Pricing Poseidon: Extreme Weather Uncertainty and Firm Return Dynamics

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Pricing Poseidon: Extreme Weather Uncertainty and Firm Return Dynamics Mathias S. Kruttli, Brigitte Roth Tran, and Sumudu W. Watugala 2019-054 Please cite this paper as: Kruttli, Mathias S., Brigitte Roth Tran, and Sumudu W. Watugala (2019). “Pricing Poseidon: Extreme Weather Uncertainty and Firm Return Dynamics,” Finance and Economics DiscussionSeries2019-054. Washington: BoardofGovernorsoftheFederalReserveSystem, https://doi.org/10.17016/FEDS.2019.054. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Pricing Poseidon: Extreme Weather Uncertainty and Firm Return Dynamics∗ Mathias S. Kruttli, Brigitte Roth Tran, and Sumudu W. Watugala† July 2019 Abstract We investigate the uncertainty dynamics surrounding extreme weather events through the lens of option and stock markets by identifying market responses to the uncertainty regarding both potential hurricane landfall and subsequent economic impact. Stock options on firms with establishmentsexposedtothelandfallregionexhibitincreasesinimpliedvolatilityof5-10percent, reflecting impact uncertainty. Using hurricane forecasts, we show that landfall uncertainty and potential impact uncertainty are reflected in prices before landfall. We find no evidence that markets incorporate better hurricane forecasts than those from NOAA. Improvements to hurricane forecasts could have economically significant effects in financial markets. JEL classification: G12, G14, Q54. Keywords: extreme weather events, uncertainty, implied volatility, stock returns, hurricanes, climate finance. ∗We thank Jawad Addoum, Rui Albuquerque, Lint Barrage, Vicki Bogan, Lauren Cohen, Kerry Emanuel, Ben Groom, Matthew Gustafson, Kristine Watson Hankins, Burton Hollifield, Andrew Karolyi, Fang Liu, David Ng, Scott Mixon, Justin Murfin, Andrew Patton, Neil Pearson, Brian Seok, Aurelio Vasquez, Scott Yonker, Youngsuk Yook, and seminar participants at the Federal Reserve Board, NOAA, Cornell University, UC San Diego, UC Santa Barbara,Caltech,theUniversityofConnecticutFinanceConference,theRiskManagementandFinancialInnovation Conference in Memory of Peter Christoffersen, the Conference on Commodities, Volatility and Risk Management, theAssociationofEnvironmentalandResourceEconomistsAnnualSummerConference,theNortheastWorkshopon Energy Policy and Environmental Economics, the CEPR-EBRD-EoT-LSE Workshop, and CEMA 2019 for helpful comments. Keely Adjorlolo, David Rubio, and Alan Yan provided outstanding research assistance. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by the Board of Governors of the Federal Reserve System or its research staff. †Kruttli: The Board of Governors of the Federal Reserve System. Email: mathias.s.kruttli@frb.gov. Roth Tran: The Board of Governors of the Federal Reserve System. Email: brigitte.rothtran@frb.gov. Watugala: Cornell University. Email: sumudu@cornell.edu. 1

1 Introduction Extreme weather events can have devastating effects and were responsible for over $300 billion in damages in the United States in 2017 alone.1 Despite significant research on extreme weather effects on real economic activity and household, firm, and financial institution decision making, little is known about uncertainty surrounding extreme weather.2 Given that uncertainty affects real economic activity and decision making (see, for example, Bernanke (1983); Bloom, Bond, and van Reenen (2007); Bloom (2009)), a comprehensive assessment of the economic effects of extreme weather events requires understanding the uncertainty dynamics surrounding them. This paper examines extreme weather uncertainty resulting from hurricanes through the lens of financial markets. Asset prices are an ideal instrument to assess the dynamics and magnitude of extreme weather uncertainty because of the frequency and scale at which financial data are available and the fact that the investor behavior underlying the asset prices is guided by financial incentives. We distinguish between two components of extreme weather uncertainty: (a) the “landfall uncertainty” regarding where, when, and whether a hurricane will make landfall, and (b) the “impact uncertainty” about a hurricane’s effect conditional on it making landfall.3 We combine firm establishment and sales data at the county level with hurricane forecast and damage data in order to identify firms that operate within regions (potentially) exposed to a particular hurricane. We use these data to test two hypotheses regarding how financial markets price a hurricane’s landfall and impact uncertainty. Our first hypothesis is that immediately after a hurricane has made landfall, implied volatilities of options of firms in the disaster region are elevated due to impact uncertainty. Implied volatility is a proxy for uncertainty (see, for example, Bloom (2009) and Kelly, Pastor, and Veronesi (2016)) as it captures the investors’ expectation of volatility. Our results support this first hypothesis. Indicative of substantial impact uncertainty, we find that immediately after hurricane landfall the 1This National Oceanic and Atmospheric Administration (NOAA) damage estimate can be found here: https://www.climate.gov/news-features/blogs/beyond-data/2017-us-billion-dollar-weather-and-climate-disasters-historic-year. 2For research on the real effects and decision making of extreme weather events, see, for example, Belasen and Polachek(2008);Imberman,Kugler,andSacerdote(2012);BarrotandSauvagnat(2016);Bernile,Bhagwat,andRau (2017); Dessaint and Matray (2017); Brown, Gustafson, and Ivanov (2017); Hong, Li, and Xu (2019). 3We focus on hurricanes because they develop and resolve over fairly short but well-defined time frames, NOAA publishes a range of forecast data, and they are likely to garner investor attention due to significant impacts and extensivemediacoverage. However,ourframeworkcanbeappliedtootherextremeweathereventslikesnowstorms and severe floods, which are also subject to landfall and impact uncertainty. 2

implied volatility of options of firms in the disaster region are between 5 and 10 percent higher than before the hurricane’s inception.4 This finding is surprisingly robust across industries, holds within industries, and for different time periods. To analyze the resolution of impact uncertainty, we examine the post-landfall stock price reactions of firms with a physical presence in a hurricane disaster region. In the short term, the abnormal returns of firms in the disaster region are not significantly different from the control firms’ abnormal returns, but the differences are substantial in the long-term. Further, the long-term differences are more pronounced for the underperforming stocks. Over the 120 trading days after hurricane landfall, the 10th percentile of the abnormal return distribution for firms in disaster regions is 12 to 14 percentage points lower than the 10th percentile of the abnormal return distribution of the control firms. Our results are consistent with a slow resolution of impact uncertainty, in line with investors only learning over time the effects of a particular hurricane and which firms were most affected by it. Our second hypothesis is that investors pay attention to hurricane forecasts before landfall and demand compensation for the landfall uncertainty and the potential impact uncertainty. This hypothesis implies that hurricane forecasts contain valuable information for investors and, if financial markets are efficient, this information should be reflected in asset prices. Other papers in the climate finance literature assessing informational efficiency have found that investors are inattentive toclimaticeventsastheyunfold(see, forexample, Hong, Li, andXu(2019)andMurfinandSpiegel (2019)). In contrast, using NOAA forecasts issued in the days or weeks leading up to a hurricane’s landfall or dissipation (in the case of a hurricane that “missed”) to measure landfall uncertainty, we find implied volatilities increase and stock returns decrease even at low landfall probabilities of 10 percent and below.5 Further, consistent with our framework, we find that the combined landfall uncertainty and expected impact uncertainty can cause implied volatility to be higher before landfall, compared to shortly after landfall when landfall uncertainty is fully resolved and only impact uncertainty remains. 4Given that our results reflect changes in implied volatilities (or stock prices) after hurricane inception, these estimates difference out potential seasonal increases in implied volatilities (or decreases in stock prices) of firms in hurricane-prone areas. As such, our estimates may understate the total increase in implied volatility (decrease in stock prices) due to hurricanes. 5Unlike at the aggregate market level, stock returns and volatility at the firm level generally exhibit positive contemporaneouscorrelationasshowninDuffee(1995);Albuquerque(2012);Grullon,Lyandres,andZhdanov(2012). As such, the negative return-volatility relationship documented for market index volatility is unlikely to drive our results. 3

We build on our main results with several key extensions. First, given anecdotal evidence that hedge funds obtain information on hurricane forecasts from sources other than the NOAA,6 we test whether financial markets can improve upon NOAA hurricane forecasts by estimating if firms that are not in the forecasted path of a hurricane but end up in the disaster region also see increases in the implied volatility of their options. We fail to reject the null hypothesis that markets do not reflect superior information to NOAA forecasts on hurricanes. In light of recent improvements in hurricane forecast accuracy7 and ongoing developments that could actually reduce the forecast accuracy going forward,8 for our second extension, we examine the potential benefits of improving NOAA forecast accuracy in this context. We estimate the additional change in implied volatility due to forecast errors for firms for which the forecasted exposure was larger (smaller) than the eventual exposure to the disaster region. We find large average effects of up to 150 basis points for thousands of firms over the sample period from 2007 to 2017. This result speaks to the outsized importance of NOAA’s hurricane forecasts for financial markets and is valuable information for legislators who make budgetary decisions. Third, although we have focused our main results on the broad universe of US public firms (excluding financial firms), we conduct a separate analysis on insurance firms. Despite limitations due to the fairly small number of public property and casualty insurance firms with liquid options and stocks, and regional exposure (fraction of written premiums) data being at the state rather than county level, we find that single stock options of property and casualty insurance firms reflect substantial impact uncertainty immediately following a hurricane landfall, increasing implied volatilities by as much as 40 percent. Our paper makes several contributions that build upon each other. We begin by presenting a novel framework to think about the uncertainty before and after extreme weather events. Second, our results show that investors are attentive to firm exposures to hurricanes even before land- 6See, for example, the discussion of the hedge fund with the name Nephila by Michael Lewis here: https://www. nytimes.com/2007/08/26/magazine/26neworleans-t.html?pagewanted=all. 7Alley, Emanuel, and Zhang (2019) show that hurricane forecasts have indeed improved dramatically in recent decades. Inparticular,theyfindthat“modern72-hourpredictionsofhurricanetracksaremoreaccuratethan24-hour forecasts were 40 years ago.” 8Spectrum auctions for the 5G network may overlap with the frequency range in which water vapor emits, thereby limiting the ability of forecasters to collect an important source of data for current forecast models (see this discussion in The Guardian: https://www.theguardian.com/world/2019/may/04/ 5g-mobile-networks-threat-to-world-weather-forecasting.) 4

fall.9 Third, our estimates imply that extreme weather uncertainty imposes significant financial costs that should be taken into account when assessing the aggregate damage of extreme weather events. Fourth, given that research has shown that other types of uncertainty can affect a firm’s decision making and that major events like elections yield similar magnitudes of uncertainty,10 the large economic magnitudes of our estimated responses together with the slow resolution of impact uncertainty suggest that extreme weather uncertainty is an important factor for real outcomes. The remainder of this paper is structured as follows. We begin with a discussion of related literature in Section 2. Then we describe our empirical design and datasets in Sections 3 and 4, respectively. We present our main results in Section 5, followed by extension in Section 6. We conclude in Section 7. 2 Related literature Our paper ties in to several diverse bodies of literature. By analyzing extreme weather uncertainty, our paper contributes to the uncertainty literature, in which several papers have focused on economic policy uncertainty and its effects on firms (see, for example, Bloom, Bond, and van Reenen (2007) and Bloom (2009)). Other researchers have focused on political uncertainty proxied by elections and how they affect firm investments and financial markets (see, for example, Julio and Yook (2012); Kelly, Pastor, and Veronesi (2016); Jens (2017)). Our paper complements this body of work by showing that extreme weather uncertainty is a different but important source of uncertainty that affects prices in financial markets. Moreover, in the case of elections, there is uncertainty about outcomes, but generally not about when and whether the elections themselves will occur because they are scheduled in advance.11 Our analysis introduces an additional layer of complexity as we separately examine the effects of the uncertainty about the landfall of a hurricane 9Investor attention to extreme weather risk is important for correctly pricing assets with extreme weather and climate change exposure and reduces the risks of sudden large price corrections that could disrupt financial stability (see, for example, Carney (2015)). 10Uncertainty has been shown to reduce firm investments by Bernanke (1983), Bloom, Bond, and van Reenen (2007), and Julio and Yook (2012). Our on average 5 to 10 percent increase in implied volatility is of a similar magnitude to the increase in implied volatility found by Kelly, Pastor, and Veronesi (2016) around major political elections. 11Empirical work on political uncertainty focuses on scheduled elections in order to isolate political uncertainty from economic uncertainty. Unscheduled elections and regime changes can be precipitated by economic conditions. In contrast, hurricanes are exogenous to economic uncertainty (economic conditions do not make hurricanes more likely), so we do not face this identification issue. 5

and the uncertainty about the impact of the event itself. Our paper differs from the research on macroeconomic uncertainty and economic growth (see, for example, Jurado, Ludvigson, and Ng (2015); Baker, Bloom, and Davis (2016); Baker, Bloom, and Terry (2018); Dew-Becker, Giglio, and Kelly (2018)) in that our analysis is at the firm level and more granular than the macroeconomy as a whole, which is important as extreme weather events are generally local phenomena. Also, the uncertainty shock in our case, the hurricane, is exactly determined. Next, by showing that extreme weather events cause substantial uncertainty in economic regions before and after landfall, our work proposes an additional factor that should be considered by the literature that examines extreme weather events’ real effects and their impact on economic agents’decisionmaking. Thisgrowingliteratureincludesworkthatexaminestheeffectsofextreme weather on labor markets and schooling (see Belasen and Polachek (2008) and Imberman, Kugler, and Sacerdote (2012)). Barrot and Sauvagnat (2016) find that shocks of extreme weather events propagate in customer-supplier firm networks. Bernile, Bhagwat, and Rau (2017) analyze the relationship between risk taking behavior and the early-life disaster experiences of CEOs. Dessaint and Matray (2017) show that managers overreact to hurricane risks after experiencing a hurricane. Brown, Gustafson, and Ivanov (2017) report that firms experience decreased cash flows after extreme snowfall events and that they respond by increasing their use of credit lines. Looking at storm-level total damages, Martinez (2018) finds that damages increase with forecast error of landfall location 12 hours before landfall. Roth Tran and Wilson (2019) find that hurricanes have a wide range of impacts on local economic activity, including on employment, population, and home prices. Finally, this paper introduces a novel topic to an emerging literature on climate finance that includes early empirical work on how Florida temperature fluctuations affect orange juice futures prices (see Roll (1984) and Boudoukh, Richardson, Shen, and Whitelaw (2007)) and how the use of a time series forecasting approach is useful for pricing weather derivatives (see Campbell and Diebold (2005)). Our research contributes to three branches of the climate finance literature. First, by examining hurricane effects, this paper builds on recent papers in the finance literature focused on extreme weather events and investor attention. Hong, Li, and Xu (2019) show that drought indices are predictive of food company stock returns, indicating that investors are inattentive to droughts’ impacts on food companies. Choi, Gao, and Jiang (2018) find evidence 6

of a positive relationship between investors’ beliefs about climate change and warmer-than-usual temperatures. Addoum, Ng, and Ortiz-Bobea (2019) examine whether firm earnings are affected by high temperatures and how analysts and investors react to temperature shocks. Second, our paper adds to climate finance papers that develop hedging strategies. While Baker, Hollifield, and Osambela (2018) and Roth Tran (2019) present theoretical models in which green or emission-oriented investors can hedge risks by investing in polluters, Andersson, Bolton, and Samama (2016) show empirically that investors can hedge against potential future prices on carbon emissions by investing in a decarbonized index. Engle, Giglio, Kelly, Lee, and Stroebel (2018) develop a climate change news index and assess strategies that can hedge an investor against such news. In contrast to these papers, we focus on market dynamics that reflect investor behavior around specific disaster events that occur at a local level. Third, by using daily hurricane forecasts from NOAA, this paper adds to recent climate finance research that analyzes how NOAA forecasts are reflected in asset prices. Drawing mixed conclusions, several papers (see Bernstein, Gustafson, and Lewis (2018); Giglio, Maggiori, Rao, Stroebel, and Weber (2018); Murfin and Spiegel (2019)) use NOAA sea level rise predictions to examine whether residential real estate prices reflect sea level rise risks. Our use of NOAA forecasts is substantiallydifferent, becauseforhurricaneswecanobservemultipleisolatedeventsfrominceptionto resolution, whereas NOAA’s forecasts for sea level rise are long-term and cannot yet be compared to realizations. We are thus able to show not only that price reactions in options and stocks are consistent with investors paying attention to NOAA’s hurricane forecasts, but also assess whether this attention is in line with the realized outcomes. 3 Empirical design 3.1 Landfall uncertainty and impact uncertainty Our framework distinguishes between two types of uncertainty that surround a hurricane: impact uncertainty and landfall uncertainty. While this paper focuses on hurricanes as an example for an extreme weather event due to the availability of high quality data, the framework of landfall and impact uncertainty can also be applied to other types of extreme weather events. The impact uncertainty is the uncertainty about a hurricane’s impact on firms with exposure to the landfall 7

area. More formally, if hurricane h is expected to make landfall at time t+1 and an all-equity firm i’s stock return at t+1 is given by r = (cid:15) +θ g , (1) i,t+1 i,t+1 h,t+1 i,h,t+1 where (cid:15) ∼ N(0,σ2) represents a random shock to the firm’s return at time t + 1 with a mean of zero and variance of σ2. The random variable g ∼ N(µ ,σ2) is independent of (cid:15) and i,h,t+1 g g captures the impact of the hurricane on the value of firm i, conditional on hurricane landfall in the firm’s geographic region. The random variable θ captures whether or not the firm is hit by the hurricane and has a Bernoulli distribution or, equivalently, a binomial distribution with one draw, θ ∼ B(1,φ), where Pr(θ = 1) = 1−Pr(θ = 0) = φ and 0 ≤ φ ≤ 1. The product of the two random variables, θ g , is the component of the return attributable to the hurricane. h,t+1 i,h,t+1 Conditional on hurricane landfall at time t+1, σ2 represents the impact uncertainty. Defining g uncertaintyasthevarianceofanunpredictabledisturbanceisinlinewithPastorandVeronesi(2012 and2013)andJurado,Ludvigson,andNg(2015). Inourframework,ahurricanelandfallintroduces uncertainty for the local economy and firms. Predicting at the time of landfall which firms will be mostaffected couldbe challengingforseveralreasons. First, thenumber ofhurricanelandfallsfora given local economy are not sufficient to predict the exact economic effect. For example, Houston, TX, had not experienced a hurricane for more than two decades before Hurricane Harvey hit in 2017. Second, a hurricane’s impact on individual firms operating within a disaster region is to a large extent unpredictable. Knowing ex-ante exactly which areas will actually flood in a particular storm, the extent of power outages, or whether a levy will break, is challenging if not impossible. Prior to (potential) landfall, there is a second source of uncertainty which we call landfall uncertainty about whether the hurricane will make landfall. More generally, in other contexts, this corresponds to the uncertainty of the incidence or occurrence of an event or the uncertainty on the extensive margin. Similarly, impact uncertainty can be thought of as uncertainty on the intensive margin. At time t, we can decompose the uncertainty generated for the firm from the hurricane into expected impact uncertainty and landfall uncertainty as follows. The expected return conditional on whether or not landfall occurs is, intuitively, E [r |θ = t i,t+1 1] = µ and E [r |θ = 0] = 0. The conditional variance of firm i’s return is, g t i,t+1 8

Var (r |θ = 0) = σ2, (2) t i,t+1 Var (r |θ = 1) = σ2+σ2. (3) t i,t+1 g Then, we can find the expected conditional variance12 and the variance of the conditional expectation,13 E[Var (r |θ)] = σ2+φσ2, (4) t i,t+1 g Var(E [r |θ]) = φ(1−φ)µ2. (5) t i,t+1 g Applying the law of total variance, we can derive Var (r ) using (4) and (5), t i,t+1 Var (r ) = E[Var (r |θ)]+Var(E [r |θ]), t i,t+1 t i,t+1 t i,t+1 = σ2+φσ2+φ(1−φ)µ2. (6) g g Landfall uncertainty is captured in the total variance by the third term in equation (6), φ(1− φ)µ2. For a given µ (cid:54)= 0, landfall uncertainty is highest when the probability of landfall, φ = 0.5. g g When µ = 0, there is no contribution from landfall uncertainty to total variance at time t. In this g case, Var (r ) varies with φ purely due to the expected impact uncertainty, φσ2. t i,t+1 g Figure1depictshowthetotalvariancepriortolandfall(Var (r ))varieswiththeprobability t i,t+1 of hurricane landfall (φ). The figure has parameters σ = 0.4 and σ = 0.05. The four dashed lines g have µ absolute values of 0.1, 0.07, 0.05, and 0, respectively. The solid line shows the level of g variance following hurricane landfall, Var (r |θ = 1) = σ2+σ2. t i,t+1 g Depending on the the parameter values of µ and σ2, as φ varies from 0 to 1, prior to landfall, g g the relative contribution to total variance from the landfall uncertainty and the expected impact uncertaintywillvary. Allelseequal,asµ increases,thecontributionoflandfalluncertaintytototal g variance increases. In Figure 1, landfall uncertainty at a given φ is the vertical distance between a curve and the dot-dash (red) straight line depicting Var (r ) when µ = 0. Var (r ) will in t i,t+1 g t i,t+1 fact be greater than Var t (r i,t+1 |θ = 1) when |µ g | > √1 σ g . In the figure, this is the case where the φ dashed lines are above the solid black line. When φ > 0 and at least one of µ or σ is non-zero, g g 12E[Var (r |θ)]=(1−φ)σ2+φ(σ2+σ2)=σ2+φσ2 t i,t+1 g g 13E[E [r |θ]]=φµ , t i,t+1 g Var(E [r |θ])=E[(E [r |θ]−φµ )2]=φ(µ −φµ )2+(1−φ)(0−φµ )2 =φ(1−φ)µ2. t i,t+1 t i,t+1 g g g g g 9

Var (r ) is greater than Var (r |θ = 0) = σ2. t i,t+1 t i,t+1 3.2 Identification strategy Changestotheexpectedvolatilityofstockreturnsduetoahurricaneeventcanbemeasuredbythe changes to the implied volatility of the stock’s options. In our analysis, we use single stock options of firms that are in the damage region or forecasted path of a hurricane to estimate treatment effects, while using single stock options of firms unaffected by a hurricane as controls. Becauseanextremeweathereventlikeahurricaneisgenerallyalocalphenomenon, ouridentification strategy is based on selecting counties where a hurricane has made (or is predicted to make) landfall. For each hurricane, we have data on which counties were damaged and which counties lay in the forecasted path of the hurricane on a given day. A firm’s exposure to these counties are then measured through the share of establishments or sales located in such counties. For firm i on day T , the landfall day of hurricane h, the exposure to counties damaged by hurricane h is given by h (cid:88) HurricaneDamageExposure = (FirmCountyExposure ×I ), (7) i,T h i,T h ,c c∈DTh c whereFirmCountyExposure istheshareoffirmi’sestablishments(sales)locatedincountyc, i,T ,c h and D is the set of counties damaged by hurricane h. Therefore, a firm’s exposure to a hurricane T h is a continuous variable that can range from 0 to 100 percent. To measure a firm’s exposure to a forecasted hurricane path, we can use the set of counties in the forecasted path of a hurricane Γ days before the landfall or dissipation of the hurricane, denoted F , in place of D : T −Γ T h h (cid:88) HurricaneForecastExposure = (FirmCountyExposure ×I ). (8) i,T h −Γ i,T h −Γ,c c∈FTh−Γ c For each hurricane, there are two groups of firms, those with and without exposure, with the degree of exposure being heterogeneous. Therefore, our analysis can be thought of as a differencesin-differences setting, where each hurricane represents a treatment, and we jointly estimate the treatment effect across all the hurricanes. Because a hurricane has an identified inception date, we can isolate and estimate a hurricane’s impact uncertainty, described in Section 3.1 by comparing the implied volatility of firms exposed to 10

thelandfallareashortlyafterlandfalltotheimpliedvolatilitybeforetheinceptionofthehurricane. The implied volatility dynamics of the options of firms with zero exposure over the same time window comprise the control set. To measure landfall uncertainty, we rely on hurricane forecasts from NOAA. NOAA releases forecasts for the path of a hurricane starting from the hurricane’s inception. For each hurricane, these forecasts provide a landfall probability for each county and each day since the inception of the hurricane. The landfall uncertainty defined in Section 3.1, can be computed based on the probabilities issued by NOAA. 4 Data and summary statistics We combine data from a range of sources. We use data both from the Federal Emergency Management Agency (FEMA) and the Spatial Hazard Events and Losses Database for the United States (SHELDUS) when determining which counties are affected by hurricanes. We identify county level pre-landfall hurricane risk levels using archived forecasts from the NOAA. We combine these data sources with National Establishment Time-Series (NETS) data on locations of firm establishments and sales to identify firm exposure to hurricanes. And finally, our stock and option outcome data come from CRSP-Compustat and OptionMetrics, respectively. We describe each of these data sources below. 4.1 Hurricane damages WeuseFEMADisasterDeclarationsSummarydataincombinationwithSHELDUSdatatoidentify counties that experienced significant hurricane damages. We create an indicator that equals 1 if a county received a FEMA disaster declaration qualifying residents for individual and household program (IHP) assistance due to a hurricane. Counties are only eligible for IHP aid if they sustain significant damage on a per capita basis. WebuildonthebasicdisasterindicatorbasedsolelyonFEMAdeclarationsbycombiningitwith SHELDUS data in order to allow for the possibility that there are areas with significant damages which do not receive IHP aid.14 SHELDUS data draw upon National Centers for Environmental 14All of our results are robust to using FEMA data alone or to using FEMA data augmented with SHELDUS data. ObservationswithoutFEMAdeclarationsbutwithSHELDUSdamagesexceedingreportedFEMAdeclaration thresholds could result from measurement error, disagreements, or FEMA choosing to diverge from reported guidelines. Onepotentialsourceofmeasurementerror,especiallyforolderdata,isthatSHELDUSshowsduplicatecounty 11

Information (formerly National Climatic Data Center) Storm Data and Unusual Weather Phenomena for hurricanes to provide county level per capita estimates of damages for named hurricane events. We set our second disaster indicator variable equal to 1 if there was a FEMA declaration for IHP aid or if the SHELDUS-reported per capita damages exceeded the published threshold for FEMA to provide IHP aid.15 Summary statistics are provided in Table 1. Figure 2 shows the numberoftimeseachcountyreceivedanIHPdeclarationfromFEMAforahurricanebetween2002 and 2017, while Figure 3 shows which counties received IHP aid or met the threshold according to SHELDUS in the case of Hurricane Sandy. 4.2 Hurricane forecasts We use NOAA’s National Hurricane Center (NHC) wind speed probability forecasts to develop our measure of a hurricane’s landfall uncertainty prior to landfall. In particular, we use text files containing probabilities that particular locations will experience winds in excess of 64 knots (KT), which is the lower bound wind speed for hurricanes. Because NOAA does not issue forecasts for the damage that counties could experience, the hurricane wind speed forecasted for a county acts as a proxy for the amount of damage, with higher wind speed implying larger damage. Thewindspeedprobabilitiesarepresentedcumulativelyfor12-hourwindowsuptofivedaysout from the time of each forecast. The NHC reports these wind speeds for cities, towns, and military bases along the coast as well as some major cities that are more inland (including Birmingham, AL, Savannah, GA, and Washington, DC.) There are three wind speed thresholds included in these reports, the lowest and highest of which are the cutoffs for tropical storm-force and hurricaneforce winds, respectively. These wind speed data are derived from the same underlying data of the hurricane forecast charts published by the NHC in real time and used by news outlets in the run-up to hurricanes. Figure 4 shows an example of the forecast chart of cumulative probability bands for hurricane force winds, as presented by the NHC, over a five day period in the case of Hurricane Sandy in 2012. level damages when the underlying data are only available to SHELDUS at a higher aggregation level. This would introduce classical measurement error and therefore attenuation bias, making our regression coefficient estimates conservative. Our results are qualitatively unchanged both for the periods 2002-2017 and 2007-2017 whether using just FEMA data or FEMA and SHELDUS data together. 15See “Per Capita Impact Indicator and Project Thresholds” by FEMA, accessed at https://www.fema.gov/ public-assistance-indicator-and-project-thresholds on January 11, 2018. 12

We have taken two steps to deal with the fact that the wind speed probabilities in these text files are only reported for particular locations, most of which are coastal. First we define a set of criteria that counties with data must meet in order to be considered at risk. For example, we might define a location as being at risk if it has at least a 20 percent probability of experiencing hurricane-force winds, that is 64 KT and above, within the next five days. Second, we define as at-risk any county within a 75-mile radius of a county with data that has been defined as being at-risk in the first step. For each day, we only use the last available forecast before close of trading, as forecasted hurricane paths can change meaningfully over the course of a day. Figure 5 illustrates a sample of processed wind speed data at different probability thresholds for Hurricane Sandy over a period four days. One benefit of using the wind speed probability data is that it provides us information not only on where the eye of the storm is expected to be but also on how intense the winds will likely be and how wide the impact will be geographically. A hurricane that technically never makes landfall because the eye of the storm never passes over land can still pass close enough to a coastline to generatesignificantdamagethroughstrongwinds, heavyrainfall, andstormsurge. Thewindspeed forecast would show strong winds in the coastal areas closest to this hurricane. We will be referring to locations with forecasted wind speeds of 64 KT or more as the locations where the hurricane is predicted to make landfall. These wind speed forecasts are available from 2007 to 2017. Table2reportssummarystatisticsonthehurricaneforecastdatausedinourempiricalanalysis. The number of storms for which we observe forecasts decreases as probability threshold or days to event resolution (hurricane landfall or, in the case of “misses”, dissipation) increases. Panel A reports the mean, median, and standard deviation of the number of county-dates observations for which we have hurricane forecasts for each storm at a given probability threshold. Panel B presents the observation count by days to resolution at a given probability threshold. 4.3 Firm data We use data on locations of firm establishments and sales in order to precisely estimate firm exposure to specific hurricanes. In particular, we use NETS data (see, for example, Neumark, Wall, and Zhang (2011) and Barnatchez, Crane, and Decker (2017)) to compute the geographic footprint of a firm. The NETS data contain establishment locations of public and private firms at 13

the county level at an annual frequency. The data also include sales data for each establishment. For each hurricane season, we use the firms’ geographic footprints from the previous year. Because the NETS data end in 2014, we use the geographic footprint from 2014 for hurricanes in 2015-2017. To obtain financial data for the firms in NETS, we map them to OptionMetrics and CRSP- Compustat, which we describe below. Our sample starts in 2002, the first year for which we have OptionMetrics data. We map firms based on the name of the firm and the address of the headquarters. We exclude all financial firms with SIC numbers from 6000 to 6799 from our analysis.16 Summary statistics for the NETS data are reported in Table 3. From 2002 to 2014, we have 4,197 (4,187) firms in our sample with establishment (sales) data. On average, a county has 69 establishments and $525 million in sales. For counties that have experienced hurricane damage the number is higher with an average of 88 establishments and $675 million in sales. Figure 6 shows counties sorted into deciles based on the number of establishments for the years 2010 and 2014. The economic activity as measured by firm establishments is high in hurricane prone areas along the Atlantic and the Gulf Coast. We obtain daily stock data from CRSP-Compustat and single-name stock options from Option- Metrics. Similar to previous studies we use the data from out-of-the-money traded options with valid pricing information,17 and we restrict the set of options to slightly out-of-the-money options. Thesearemoreliquidandhavearelativelysmallereffectfromanypotentialearly-exercisepremium for American options over European options for the time horizons that we consider. Accordingly, we include single-stock options with: (i) standard settlement, (ii) a positive open interest, (iii) a positive bid price and bid-ask spread, (iv) a valid implied volatility estimate, (v) greater than 7 days and at most 200 days to expiry, and (vi) an option delta, δ, that satisfies 0.2 ≤ |δ| ≤ 0.5. The estimate for the average implied volatility of firm i at date t is, IV = 1 (cid:80)N IV , where M is the nearest-to-maturity expiration at time t with valid opi,t N j=1 i,j,t,M tionswhichsatisfytheabovecriteriaandN isthenumberofvalidstockoptionsforfirmiwiththat expiry. Using this methodology for the period from 2002 to 2017, we obtain 10,152,776 firm-date observationsofimpliedvolatilityfromOptionMetrics. WemergethesedatawithCRSP-Compustat on firm CUSIP, which yields 9,420,182 observations covering 5,691 firms and 4,028 dates. Of these 16We provide a separate analysis on insurance firms in Section 6.3. 17See, among others, Carr and Wu (2009); Kelly, Pastor, and Veronesi (2016); Martin and Wagner (2018). 14

3,866,672observationsarefrom2,198firmsthatappearatleastonceinthefirmestablishmentdata from NETS. The summary statistics for these data are in Table 4. 5 Results In this section, we describe the regression specifications that we employ to test our hypotheses and the corresponding results. The first hypothesis is that stock options of firms in hurricane disaster regions exhibit a higher implied volatility right after a hurricane has hit, in line with investors perceiving uncertainty about the impact on firms in the disaster region of the hurricane. We posit that this impact uncertainty resolves slowly over time. The slow resolution of impact uncertainty is reflected in the abnormal stock returns of firms with exposure to the hurricane disaster region, which exhibit a large dispersion that is negatively skewed in the long-run after a hurricane has hit, consistent with investors learning over time which firms were most affected. Our second hypothesis is that prior to landfall, stock and option prices react to hurricane forecasts, with implied volatility increasingandstockpricesdecreasingforfirmslocatedintheforecastedhurricanepaths, consistent with investors paying attention to hurricane forecasts. 5.1 Impact uncertainty estimation We begin with testing the hypothesis that stock options of firms in disaster regions have higher impliedvolatilities. Theimpliedvolatilityisanestimateofexpectedfuturevolatilityandiscommonly used as a measure of uncertainty. If hurricane landfall leads to impact uncertainty for firms in the disaster region, the implied volatilities of these firms should increase. The impact uncertainty can beisolatedandestimatedbylookingattheimpliedvolatilitiesshortlyafterlandfall, wheninvestors know where the hurricane hit and the landfall uncertainty has fully resolved, but do not know what the eventual impact on the firms located in the damage region will be. To test this first hypothesis, we estimate the following panel regression model, (cid:32) (cid:33) IV i,T +τ log h = λHurricaneDamageExposure +π +ψ +(cid:15) , (9) IV i,T∗ i,T h h Ind i,h,τ h 15

where τ is the number of trading days since the hurricane made landfall on day T .18 The last h trading day before the inception of the hurricane is T∗ and IV is the implied volatility of firm i. h i HurricaneDamageExposure is a measure of firm i’s exposure to the counties with hurricane i,T h damage, as defined in equation (7).19 This measure can vary from 0 percent, for firms with no exposure to the hurricane disaster region, to 100 percent, for firms with all of their establishments (or sales) located within the disaster region. We use the NETS data to measure the share of a firm’s establishments or sales in a specific county. We include hurricane fixed effects (θ ), which h is equivalent to including time fixed effects because there is one time period per hurricane. We include industry fixed effects (ψ ) based on the firms’ two-digit SIC numbers. We exclude firms Ind that do not have implied volatility measures for each trading day from inception to τ days after landfall. As our treatment selection is geographically based, we cluster the standard errors by firm headquartercounty(see, forexample, DessaintandMatray(2017)andAbadie, Athey, Imbens, and Wooldridge (2017)).20 The regression model in equation (9) can also be seen as a differences-in-differences estimation where each hurricane acts as a treatment, such that firms with exposure to the disaster region are considered treated and those without exposure act as controls. Following the recommendation of Bertrand, Duflo, and Mullainathan (2004), for each hurricane, we collapse the time series information into a pre- and post-period, where the pre-period is T∗, the day before hurricane inception, h and the post-period is T +τ, τ days after landfall. h The coefficient estimate of λ is expected to be significant and positive if investors perceive that hurricane landfall leads to impact uncertainty surrounding the local firms. Knowing ex-ante which firms will be most affected is likely impossible because of several factors. First, the number of hurricane landfalls for a given local economy are mostly insufficient to predict the exact economic effect. Second, the hurricane’s impact on individual firms in the disaster area is to a large extent random, as described in Section 3.1. Theestimationresultsofthemodelgiveninequation(9)arereportedinTable5. PanelAshows 18If a hurricane makes multiple landfalls, the first landfall date is used as T in the analysis. h 19The inception day of a hurricane is defined as the first day on which the hurricane is predicted to make landfall with at least a 1% probability. For hurricanes before 2007, we do not have hurricane forecast data available and choose as inception day the first day that the hurricane appeared as a tropical depression. 20We use the headquarters location for clustering instead of a firm’s geographic footprint as a firm’s geographic footprint is unique in the great majority of the cases, which would leave us with clusters that are not sufficiently large to ensure conservative standard error estimates. 16

the results when the exposure of a firm to the hurricane disaster region is based on establishments. Our results are robust to both selecting counties in the disaster region based solely on FEMA damage data as well as using FEMA damage data enhanced with SHELDUS data. We show results for τ = 5 trading days after landfall, but our results are robust to choosing a different τ close to landfall. We show results both for 2002 to 2017, when we have high quality option data, andfor2007to2017,whenthehurricaneforecastdatausedinthesubsequentanalysisareavailable. Theestimateofλissignificantandpositiveforallspecifications,inlinewithourfirsthypothesis that hurricane landfall causes impact uncertainty for local firms. In particular, we find that a firm with 100 percent of its establishments located in the disaster region experiences a 5 to 10 percentincreaseinitsimpliedvolatilityrelativetobeforeinceptionofthehurricane. Thiseconomic magnitude is considerable and comparable to Kelly, Pastor, and Veronesi (2016), who show that political uncertainty leads to implied volatilities of index options increasing on average around 5 percent around major political elections compared to non-election periods. These results are robust to including industry times time fixed effects, which implies that the effects are present within industry. These results are robust to measuring the geographic footprint of a firm by sales at the county levelinsteadofestablishments, asshowninPanelB.TheestimatesofλinPanelBarealsostrongly significant for all specifications. The largest coefficient estimates for five days after the landfall are 0.06, implying that a firm with a 100 percent of its sales in the disaster region has an implied volatility that is 6 percent higher than before the inception of the hurricane. The fact that the magnitudes of the coefficients based on exposure of sales (Panel B) are smaller than those based on exposure of establishments (Panel A) could be explained by investors being more concerned about damages to production facilities or our data on locations of firm establishments being better than on the sales. Importantly, our results are not driven by small firms. The average market capitalization of a firm with exposure to disaster region counties of at least 20% is $6.0 billion and $8.2 billion when measuring the exposure by establishments and sales, respectively. The average market capitalization of a firm with less than 20% exposure to the disaster region is similar in magnitude with $6.8 billion and $6.7 billion, respectively. Firms with coastal exposure can differ from other firms based on unobserved characteristics, and it is possible that firms that would be more vulnerable to 17

hurricanes because of their particular line of business avoid being exposed to the Atlantic or Gulf Coast. However, such sorting would bias us against finding evidence of impact uncertainty. We also test if these results are driven by a particular industry but find that the impact uncertainty of hurricanes is similar across industries. Table A.1 in the appendix shows the results when an industry dummy is interacted with HurricaneDamageExposure in equation (9).21 i,T h The coefficient estimate of the interaction term is insignificant for almost all specifications, which suggests that the reported effect is not driven by one particular industry. Only the construction industry shows a consistent pattern of facing less uncertainty surrounding a hurricane than the other industries, with the caveat that the number of construction firms in our sample is small. A potential explanation for this finding is that these firms can profit from a hurricane due to a boom in reconstruction. 5.2 Impact uncertainty resolution The large impact uncertainty measured in the previous section suggests that firms in the disaster region face uncertain outcomes. The resolution of this impact uncertainty should be reflected in the firms’ stock prices in the months following a hurricane landfall. In particular, we test if the abnormalstockreturnsoffirmswithexposuretoahurricanedisasterregionshowalargedispersion inthelong-run, inlinewithinvestorslearningovertimehowdevastatingahurricanewasandwhich firms were most adversely affected. To isolate the resolution of impact uncertainty, we estimate how a hurricane affects firms’ stock returns after landfall. We first estimate daily abnormal returns relative to the Fama-French threefactor model (see Fama and French (1993)). For each firm and each hurricane in our sample, the following model is estimated: r = α +β r +β r +β r +(cid:15) , (10) i,d i 1,i m,d 2,i smb,d 3,i hml,d i,d where r is the daily market return on day d minus the risk-free rate, r and r are the m,d smb,d hml,d daily returns of the small-minus-big and high-minus-low portfolios, respectively. We estimate this model using 200 trading days before the day of hurricane landfall. We then use the coefficient 21Theindustrydummyisbasedonthetwo-digitSICnumbersofthefirms. Inthistable,weexcludetheagriculture and non-classified categories because of the small number of firms. 18

estimates from this first stage regression to compute abnormal returns for each firm and hurricane as follows: ra = r −(αˆ +βˆ r +βˆ r +βˆ r ). (11) i,d i,d i 1,i m,d 2,i smb,d 3,i hml,d We next aggregate the abnormal returns to a cumulative abnormal return, denoted rac , for i,T :T +τ h h each firm and hurricane over the time period T to T +τ, where T is again the day of the landfall h h h and τ is the number of trading days.22 The time period starts in 2002 and ends in 2017, which correspondstothetimeperiodusedinTable5. Toensurethatstockswithstalepricesareexcluded from our analysis, a stock is required to have return data for all trading days from 200 trading days before landfall to τ trading days after landfall. We compute the differences in the mean and nine percentiles between the cumulative abnormal return distributions of firms with (treated) and without (control) exposure to a hurricane damage region. The results are reported in Table 6 along with the corresponding t-stats. We estimate the standard errors using a bootstrap that clusters by county based on firm headquarters. Because we want to compare the returns in the short- and long-run after hurricane landfall, the cumulative abnormal returns are computed from the day of landfall for up to 5 and 120 trading days after landfall.23 For Panel A, we consider firms to be in the disaster region if at least 50 percent of the establishments are in the disaster region. For Panel B, the threshold is 50 percent of the sales. Table A.2 in the Appendix shows that the results are robust to lowering the threshold to 25%. Panel A shows that the cumulative abnormal returns from the landfall day to five days after yield a negative difference for all percentiles except the top one. These differences are generally between -10 and -50 basis points, and they are not significant. However, when looking at the cumulativeabnormalreturnsfromlandfalldayto120tradingdaysafterthelandfall, thedifferences in cumulative abnormal returns are strongly negatively skewed. For the 10th and 20th percentiles, the difference in cumulative abnormal returns between control and treated firms is around -14 percent and strongly statistically significant, but for the 80th and 90th percentiles, the difference is only around -5 percent and statistically insignificant. In Panel B, firm exposure to hurricane disaster regions is measured based on a firm’s sales in a county. The cumulative abnormal return 22If a hurricane makes landfall on a non-trading day, we take the next trading day as T . h 23We choose a horizon of 120 trading days as that corresponds to half a calendar year. The hurricane season lasts half a calendar year, and thus, we avoid overlaps with the following year’s hurricane season. The results are robust to different benchmarks for the horizon. 19

distributionofthetreatedfirmsisagainnegativelyskewedandcomparabletoPanelAinmagnitude andstatisticalsignificanceforthelonghorizon. Fortheperiodfromlandfallto5tradingdaysafter, the differences between percentiles of the control and treated firms’ return distributions are again insignificant. These findings are in line with a slow resolution of impact uncertainty in the aftermath of a hurricane. Investors appear to learn over time how devastating a hurricane was and which firms were most adversely affected. 5.3 Uncertainty before landfall In Table 5, we show that shortly after landfall, options price in substantial impact uncertainty in potential hurricane disaster regions. We now turn to our second hypothesis that investors pay attention to hurricane forecasts before landfall and demand compensation for the landfall uncertainty and potential impact uncertainty. While landfall uncertainty is an important component of extreme weather uncertainty as discussed in Section 3.1, this analysis also investigates if investors pay attention to climatic events as they unfold. Other papers in the climate finance literature find that investors are inattentive to climate-related forecasts such as droughts (see Hong, Li, and Xu (2019)) and sea level rise (see Murfin and Spiegel (2019)). Over the course of the days or weeks while a hurricane makes its approach toward the Atlantic or the Gulf Coast, NOAA issues hurricane forecasts that contain the probabilities of the hurricane making landfall in a particular region. Such forecasts are often highly publicized through news outlets. For example, the forecasted path of Hurricane Sandy in 2012 shown in Figure 4 likely looks familiar to people who tend to follow the news during hurricane season. Based on the efficient market hypothesis, investors should pay attention to these forecasts, and the forecasts shouldbepricedin. Ifinvestorspayattentiontohurricaneforecastsbeforelandfall,thentheimpact uncertainty will increasingly be priced into options as the likelihood of a hurricane making landfall in a specific region increases, which is represented by the term φσ2 in equation (6). In addition, investor attention to hurricane forecasts will also lead to landfall uncertainty, given by the term φ(1−φ)µ in equation (6), being reflected in option prices through higher implied volatilities. g We use the NOAA forecasts described in Section 4.2 to examine how hurricane forecasts affect implied volatilities of firms located in the path of a hurricane and estimate the following panel 20

regression model (cid:32) (cid:33) IV i,T −Γ log h = λHurricaneForecastExposure +π +ψ +(cid:15) , (12) IV i,T∗ i,T h −Γ h Ind i,h,Γ h where Γ represents the number of calendar days before the landfall or dissipation of the hurricane and we estimate the regression separately for each Γ ∈ {1,2,3,4,5}, as NOAA forecasts hurricane pathsuptofivedaysout.24 Firmi’sexposuretohurricaneh’sforecastedpath,HurricaneForecast Exposure , is as defined in equation (8). The remaining parameters are as described for re- T −Γ h gression equation (9). Only hurricanes for which the day T −Γ is a trading day are included in h the regression. The time series starts in 2007, because we have high quality hurricane forecast data from 2007 onwards, and ends in 2017. As described in detail in Section 4.2, the hurricane forecasts provide a probability that a county will experience wind speeds of at least 64 KT within five days. A storm’s wind speed has to be at least 64 KT to be classified as a hurricane by NOAA. If investors pay attention to hurricane forecasts, the estimate of λ is expected to be positive and significant. Particularly, the change in a firms’ implied volatilities should depend on the probability that a hurricane will make landfall in counties in which the firm operates. In our framework presented in Section 3.1, we show in Figure 1 that for any probability of landfall greater than zero, given by the term φ, the implied volatility will be higher than before the inception of the hurricane. Further, the total uncertainty given in equation (6) can be higher before landfall, when landfall and impact uncertainty are present,compared to after landfall when there is no uncertainty about landfall but only about the impact of a hurricane. Figure 1 shows that depending on the parametrization (depending on the expected impact (µ ), impact uncertainty (σ ), and probability g g of landfall (φ)), the total variance (uncertainty) can be higher before landfall, (when φ is less than 1) than at landfall (when φ equals 1). Whether total uncertainty is higher before landfall than right after landfall is ultimately an empirical question. We report the estimation results of the model in equation (12) in Table 7. The parameter Γ is between 1 and 5 days, and the probabilities of hurricane-level wind speeds that we require to designate a county as at-risk ranges from 1 to 50 percent. Figure 7 plots the λ coefficient estimates and confidence interval bands for Γ=1 and Γ=2. For each Γ and probability, we require a sample 24If a hurricane makes landfall on multiple days, we only consider the first landfall day. 21

with at least three hurricanes and 25 firms that have at least 20 percent of their establishments or sales in counties in the forecasted path of a hurricane. Because the days before the landfall or dissipation of a hurricane can fall on non-trading days and different hurricanes reach the specified probabilitythresholdsofmakinglandfallondifferentdays,thehurricanesincludedintheestimation can vary across the table’s columns. Theresultsareinlinewithinvestorspayingattentiontohurricaneforecastsandtheuncertainty surrounding a hurricane being reflected in the implied volatilities of firms located in the forecasted path of a hurricane. The estimates of λ are always positive, regardless of whether a firm’s exposure to a hurricane is based on establishments (Panel A) or sales (Panel B). The λ estimates are also significant with the exception of the estimates five days before landfall/dissipation. For each day, the magnitude of λ increases with higher landfall probabilities. It is clear from Figure 7 that the λ estimates are significant higher at as landfall probability increases. In fact, for high probabilities, the increase in implied volatilities is in some cases larger than the increase in implied volatilities right after landfall reported in Table 5.25 This result suggests that the landfall uncertainty is strongly reflected in option prices and can push the total uncertainty before landfall above the impact uncertainty measured at landfall, as suggested by our framework in Section 3.1. Overall, these results are consistent with hurricane forecasts containing valuable information and investors paying attention to them. An interesting observation is that the estimated magnitude of λ for the same probability is sometimes lower for days closer to the landfall or dissipation of the hurricane, although not significantly so. This could reflect that hurricanes that reach a specific probability of making landfall whentheyarestillfaroffthecoastaresimplystrongerhurricanesthatcanleadtomoredevastating effects. 25A caveat is that the sample of hurricanes in the two tables can differ. In particular, while in Table 5 we include onlythehurricanesthatmakelandfall,inTable7wealsoconsiderhurricanesthatdissipatewithoutmakinglandfall. Also, for some hurricanes included in Table 5, a specific day before landfall can be a non-trading day, and thus, the hurricane would not be included for that day in Table 7. However, the result of higher total uncertainty before the landfall than right after landfall holds even when comparing the exact same set of hurricanes both before and after landfall. 22

5.3.1 Alternative specification for forecasts and implied volatilities The estimates of the regression model shown in equation (12) support the hypothesis that investors pay attention to hurricane forecasts and the uncertainty surrounding a hurricane is reflected in option prices before landfall. To further test the robustness of this result, we use an alternative estimation where we allow for the fact that firms can reach an exposure threshold to a specific hurricane, for example, 10 percent of establishments being located in the forecasted path of a hurricane, on different days. In the regression model in equation (12), it is not possible to jointly estimate the change in implied volatilities for these firms. The model specification below allows for a joint estimation, but the hurricane exposure variable is an indicator variable instead of a continuous variable as in equation (12). We compute the measure given by (cid:32) (cid:33) (cid:32) (cid:33) IV i,t 1 (cid:88) IV j,t IVD = log i,h − log i,h , (13) i,h IV i,T∗ |J h | IV j,T∗ h J h h where t is the first trading day when the number of establishments (sales) of firm i in the path i,h of hurricane h exceed a certain threshold, and T∗ is again the last trading day before the inception h of hurricane h. The set of control firms, J , for hurricane h are the firms with zero exposure to h the forecasted path of the hurricane. We exclude from this analysis the days on which a hurricane makes landfall. We compute IVD for all hurricanes and firms and estimate the mean, IVD, for i,h the sample from 2007 to 2017. A positive and significant IVD would be consistent with the results shown in Table 7. The results for this specification are presented in Table 8. Here we use five probability thresholds ranging from 1 to 50 percent to designate which counties lie within the forecasted path of a hurricane. A 1 percent probability threshold implies that a county has at least a 1 percent chance of experiencing hurricane-force winds in the next 5 days. We consider three thresholds for a firm’s exposuretoahurricanewherein10, 25, and50percentofafirm’sestablishments(sales)arelocated in counties that we have designated as being in the hurricane’s forecasted path. For a 10 percent threshold, we compute the difference in the implied volatility of firm i on the first trading day that 10 percent of firm i establishments (sales) are located in the hurricane’s forecasted path and the implied volatility on the last trading day before the inception of the hurricane. 23

Panel A reports the estimates of IVD when the firm geographic footprints are computed based on the share of establishments in a county. The estimates of IVD are positive and significant for the great majority of the specifications. The only two exceptions are for the probability threshold of 1 percent. As in Table 7, the magnitude of the estimates is monotonically increasing with the probability of the firms being hit by a hurricane. Further, when selecting only firms with at least 25 or 50 percent of their establishments or sales in at-risk counties, the estimates are substantially larger than for the firms with an exposure of at least 10 percent. These results further support the hypotheses that investors pay attention to hurricane forecasts which leads to uncertainty landfall and expected impact being reflected in the implied volatilities of firms located in the forecasted path of a hurricane. 5.3.2 Forecasts and stock returns The results in the previous sections show that hurricane forecasts lead to price effects in option markets. Consequently, hurricane forecasts should also affect the underlying stock prices. We expect that a higher likelihood of a hurricane strike should cause negative stock returns. First, the stock returns can be depressed because investors require a premium to hold stocks during a time period of high uncertainty, as, for example, discussed by Pastor and Veronesi (2012, 2013) in the context of political uncertainty.26 Second, the possibility of a hurricane strike can decrease stock returns because of lower expected cash flows. To test if hurricane forecasts affect the stock returns, we estimate the regression model in equation (12), but with cumulative stock returns as the dependent variable: r i,T h ∗:T h −Γ = λHurricaneForecastExposure i,T h −Γ +π h +ψ Ind +(cid:15) i,h,Γ , (14) where r i,T∗:T −Γ is the cumulative return of firm i from the inception of hurricane h to Γ calendar h h days before the landfall or dissipation of the hurricane.27 An estimate of λ that is significant and negative would support the hypothesis that the uncertainty surrounding firms in a hurricane’s path 26PastorandVeronesi(2012,2013)lookatmarketeffects,whichcannotbediversified. Theuncertaintysurrounding a hurricane affects individual firms, but the presence of market frictions, for example, limited access to information as in Merton (1987), can lead to investors demanding a premium for idiosyncratic volatility. 27Weusenon-adjustedreturnsinthisanalysis,buttheresultsarequalitativelythesamewhenusingrisk-adjusted returns. 24

leads to negative stocks returns. The results are reported in Table 9, which is structured the same as Table 7. The estimates of λ are negative in all cases except for the last column that examines effects five days before landfall/dissipation with a probability of a hurricane hit of 10 percent. The estimates are strongly significant for the majority of the specifications. The estimates are also economically significant. The smallest estimate is -0.09, which implies that a firm with a 100 percent exposure to the forecasted path of a hurricane experiences a negative return of 9 percent from the inception of the hurricane to a few days before the landfall/dissipation. The estimates are similar when the geographicfootprintofacompanyisbasedonestablishments, asinPanelA,oronsales, asinPanel B. These results are consistent with our analysis on implied volatilities and support the hypothesis that the uncertainty associated with a hurricane leads to negative stock returns. 6 Extensions Having examined how markets broadly price in impact and landfall uncertainty both before and afterhurricanelandfall,wenowturnourattentiontothreekeyextensions. Weaskwhethermarkets beat NOAA forecasts, what the potential benefit of improving hurricane forecast accuracy is, and how extreme weather uncertainty affects insurance firms (which are excluded from our baseline results together with other financial firms). 6.1 Can the market forecast better than NOAA? The previous results show that market prices react to hurricane forecasts by pricing in the impact uncertainty caused by a potential hurricane strike. The hurricane forecasts in our analysis are taken from NOAA. NOAA’s hurricane forecasts are arguably the most prominent as they are widely publicized through the media. However, it is possible that large institutional investors like hedge funds, which often act as marginal investors in asset markets and move asset prices, could outperform the publicly available NOAA hurricane forecasts by trading on proprietary hurricane forecast information. In this case, markets would predict hurricane damages more precisely than NOAA forecasts. There are a few of reasons to believe that markets could predict damages more precisely. First, 25

there is anecdotal evidence that hedge funds buy information on hurricane forecasts from private companies.28 Second, the budget of NOAA’s subdivision responsible for hurricane forecasts, the National Weather Service, is minuscule compared to the value of assets managed by large institutionalinvestors.29 Therefore, a firmthat produces proprietaryhurricane forecastsand sells themto institutional investors could potentially generate sufficient revenues to rival NOAA. Third, because the NOAA forecasts are public and available in real-time, investors can use the information in the NOAA forecasts and improve upon them with proprietary information. We test this hypothesis by estimating the panel regression model in equation (12) with an additionaltermthatmeasureswhetheroptionmarketscanpredictwhichfirmsendupmoreexposed to the hurricane than predicted by the NOAA forecasts: (cid:32) (cid:33) IV i,T −Γ log h =λHurricaneForecastExposure i,T −Γ IV i,T∗ h (15) h +γUnderPrediction +π +ψ +(cid:15) . i,T h −Γ h Ind i,h,Γ Here UnderPrediction is defined as the difference between a firm’s exposure to counties that i,T −Γ h eventually experience hurricane related damages and the exposure to counties in a hurricane’s forecasted path: UnderPrediction =(HurricaneDamageExposure i,T −Γ i,T h h −HurricaneForecastExposure ) (16) i,T −Γ h ×I . (HurricaneDamageExposurei,Th −HurricaneForecastExposurei,Th−Γ)>0 Firm i will have a positive value for UnderPrediction if the share of its establishments or i,T −Γ h sales in counties that experience hurricane damages is greater than the share of its establishments or sales predicted to be affected based on NOAA forecasts made Γ days before landfall. Otherwise, UnderPrediction will assume a value of zero. If the market can forecast which counties will i,T −Γ h experiencehurricanedamagemoreaccuratelythanNOAA,theestimateofγ inequation(15)would 28See, for example, the discussion of the hedge fund Nephila by Michael Lewis here: https://www.nytimes.com/ 2007/08/26/magazine/26neworleans-t.html?pagewanted=all. 29The total budget of the National Weather Service, a subdivision of NOAA, was around $1 billion in 2017. However, this budget also includes funds appropriated for weather forecasts other than hurricane forecasts. The budget of the National Weather Service for 2017 can be found here: https://www.corporateservices.noaa.gov/nbo/. 26

be significant and positive.30 The underprediction measure can also be thought of as a measure of false negatives, and we test if markets can detect them. Welookatunderpredictedfirmsratherthanoverpredictedfirms,becauseasmallerthanaverage increase in implied volatility for overpredicted firms (firms that end up with less exposure to the damageregionthanforecasted)couldbeexplainedbymarketsbeinglessattentivetolowprobability forecasts, which include many firms in the forecasted path that end up with no or little exposure to the damage region. Therefore, analyzing underpredicted firms allows us to better isolate and estimate the forecast ability of financial markets. TheresultsareshowninTable10,whichhasthesamestructureasTable7. Whiletheestimates ofγ arepositive,inlinewiththemarketforecastingbetterthanNOAA,theyareinsignificantforall butoneweaklysignificantcoefficientestimate. Therefore, wedonotfindsupportforthehypothesis thatmarketscanforecasthurricanesbetterthanNOAA.Thisnullresultcouldbecausedbyalackof statistical power. However, given that the number of firms with a non-zero UnderPrediction i,T −Γ h measure is quite large with an average of around 1,200 across the specifications, we should have sufficient power to detect the market’s ability, if any, to beat NOAA forecasts. In contrast, the coefficient estimates on HurricaneForecastExposure are positive and strongly significant i,T −Γ h for most specifications as in Table 7.31 6.2 The economic effect of improved forecasts The previous findings show that markets price in NOAA’s hurricane forecasts and furthermore do not appear to be able to outperform NOAA forecasts. These findings lead to the question: what economic effect would improved hurricane forecasts have? In other words, how much of the price variation in options around hurricanes were caused by mispredictions, that is over- and 30Suppose, for example, that the NOAA forecast implies zero exposure for a firm four days before a hurricane’s actual landfall. If the firm ends up with significant exposure to counties affected by the hurricane four days later, UnderPrediction would equal the exposure of the firm to the actual hurricane damage region. If the markets i,Th−4 areabletopredictthefinalexposurefourdaysaheadwhentheNOAAforecastdidnot,γ wouldreflectthisbybeing significant and positive. 31We also test if our results change when we focus on options for which the underlying stocks have a large institutional ownership, because institutional investors are more likely than retail investors to have the means to obtain hurricaneforecaststhatarenotgeneratedbyNOAA.However,wedonotfindevidenceofthemarketoutperforming NOAA. 27

underpredictions.32 This price variation could potentially be reduced by more accurate forecasts.33 To answer this question, we compute by how much implied volatilities are too high or low on average due to an overprediction or underprediction. To estimate by how much implied volatilities are too high due to an overprediction, we compute the average overprediction (for the day before landfall or dissipation) for the probability thresholds shown in Table 7. We then multiply this average overprediction by the coefficient estimate on HurricaneForecastExposure of the i,T −Γ h corresponding probability threshold given in Panel A of Table 7. To estimate by how much implied volatilities are too low due to an underprediction, we multiply the average underprediction by the equation (9) coefficient on HurricaneDamageExposure estimated for the respective hurricanes i,T h in the sample using observations five days post-landfall.34 Figure8presentstheresultsofthisanalysis. PanelAshowstheaverageover-andunderreactions inimpliedvolatilitiesresultingfromover-andunderpredictions. Themagnitudesarelarge,reaching 75 basis points for the overpredictions and -150 basis points for the underpredictions. To provide moreintuitionontheeconomicsignificance,wemultiplytheover-andunderreactionsintheimplied volatilities with the average market capitalization of the respective firms and show the results in Panel B.35 The resulting product tells us by how much the expected fluctuations of the market capitalization changes due to the over- and underpredictions. The overpredictions reach values of above $60 millions, and the underpredictions lead to values of below -$80 millions. Considering that the number of firms affected by over- and underpredictions is large as reported in Panel C, these results imply that improvements to hurricane forecasts could have large economic effects on pricing of hurricane related uncertainty in option markets. 32While the measure of underprediction is described in equation 16, the overprediction measure is defined as OverPrediction = (HurricaneDamageExposure − HurricaneForecastExposure ) × i,Th−Γ i,Th i,Th−Γ I . The overprediction measure can be thought of as (HurricaneDamageExposurei,Th −HurricaneForecastExposurei,Th−Γ)<0 a measure of false positives. 33Alley, Emanuel, and Zhang (2019) show that hurricane forecasts have indeed improved dramatically in recent decades. Inparticular,theyfindthat“modern72-hourpredictionsofhurricanetracksaremoreaccuratethan24-hour forecasts were 40 years ago.” 34The damage exposure of a firm is based on the combined FEMA and SHELDUS dataset, and the number of establishments act as the geographic footprint measure. 35Forexample,fortheunderpredictionofthe50%ormoreprobabilitythreshold,wemultiplytheaverageunderreactionof-1.54%with0.44,whichistheaverageimpliedvolatilitylevelinoursample,andthenmultiplytheresulting product with the average market capitalization. 28

6.3 Insurance firms The analysis and discussion so far in this paper has been focused on the universe of firms excluding financial firms as common in the asset pricing literature. One contribution of this paper is to show that the uncertainty around extreme weather events affects a wide range of firms and not only insurance firms which are often thought of in the context of natural disasters. However, we also want to investigate if extreme weather uncertainty is reflected in the asset prices of insurance firms. Thechallengethatwefaceisthatthenumberofpubliclytradedinsurancefirmswithliquidoptions is relatively limited and we only have the exposure of an insurance firm at the state level, not at the county level. We use data on insurance statutory financials from S&P Global Market Intelligence, which provides us with the share of total premiums written by state for property and casualty insurance firmsintheUS.Weestimatetheregressioninequation(9)forthesepropertyandcasualtyinsurance firms, with HurricaneDamageExposure being replaced by a variable that measures the share i,T h of total premiums, lagged by one year, written in states that experienced damage by hurricane h. The results are reported in Table 11. Panel A (B) considers a state to have experienced hurricane damage if at least 10% (25%) of the counties experienced hurricane damage as measured by FEMA data and FEMA data enhanced with SHELDUS. Thecoefficientestimateispositiveandsignificantforallspecificationsimplyingthattheimpact uncertainty for property and casualty insurance firms is substantial in the aftermath of a hurricane. The magnitude of the coefficient estimates are economically significant, with the implied volatility being up to 40% higher for insurance firms with a 100% exposure to the damage region of the hurricane. The statistical significance is slightly weaker than for the universe of firms in Table 5 as the number of insurance firms in our sample is relatively small. We have on average 20 to 30 insurance firms per hurricane depending on the specification. 7 Conclusion This paper studies extreme weather uncertainty through prices in option and stock markets by analyzing the uncertainty surrounding hurricanes. Our framework distinguishes between landfall uncertainty (on where the hurricane will hit, if at all) and impact uncertainty (on the consequences 29

to the local firms and economy following landfall). We find that options of firms operating in regions affected by hurricanes have considerably higher implied volatilities, between 5 to 10 percent, in the immediate aftermaths of those hurricanes. The higher implied volatilities are in line with investors being concerned about substantial impact uncertainty right after a hurricane has hit. The stock returns of firms in damage regions show a strong negative skewness in the long-run consistent with a slow resolution of impact uncertainty. Using daily hurricane forecasts from NOAA, we find that landfall uncertainty combined with potential impact uncertainty are both priced before a hurricane makes landfall. Our novel analysis and framework contribute to a burgeoning climate finance literature. Further, we add to the existing uncertainty literature by showing that extreme weather uncertainty is important and reflected in the prices of options and stock markets. Future research can build on the results in this paper by linking extreme weather uncertainty to real economic activity. Extreme weather uncertainty potentially affects firm production networks, commodity and agricultural markets, and decisions by various economic agents. 30

References Abadie, Alberto, Susan Athey, Guido W. Imbens, and Jeffrey Wooldridge, 2017, When should you adjust standard errors for clustering?, Working Paper. Addoum, Jawad M., David Ng, and Ariel Ortiz-Bobea, 2019, Temperature shocks and earnings news, Working Paper. Albuquerque, Rui, 2012, Skewness in stock returns: Reconciling the evidence on firm versus aggregate returns, The Review of Financial Studies 25, 1630–1673. Alley, Richard B., Kerry A. Emanuel, and Fuqing Zhang, 2019, Advances in weather prediction, Science 363, 342–344. Andersson, Mats, Patrick Bolton, and Frederic Samama, 2016, Hedging climate risk, Financial Analysts Journal 72, 13–32. Baker, Steven D., Burton Hollifield, and Emilio Osambela, 2018, Asset prices and portfolios with externalities, Working Paper. Baker,ScottR.,NicholasBloom,andStevenJ.Davis,2016,Measuringeconomicpolicyuncertainty, Quarterly Journal of Economics 131, 1593–1636. Baker, Scott R., Nicholas Bloom, and Stephen J. Terry, 2018, Does uncertainty reduce growth? Using disasters as natural experiments, Working Paper. Barnatchez, Keith, Leland D. Crane, and Ryan A. Decker, 2017, An assessment of the national establishment time series (NETS) database, Working Paper. Barrot, Jean-Noel, and Julien Sauvagnat, 2016, Input specificity and the propagation of idiosyncratic shocks in production networks, Quarterly Journal of Economics 131, 1543–1592. Belasen, Ariel R., and Solomon W. Polachek, 2008, How hurricanes affect wages and employment in local labor markets, American Economic Review 98, 49–53. Bernanke, Ben S., 1983, Irreversibility, uncertainty, and cyclical investment, Quarterly Journal of Economics 98, 85–106. 31

Bernile, Gennaro, Vineet Bhagwat, and P Raghavendra Rau, 2017, What doesn’t kill you will only make you more risk-loving: Early-life disasters and CEO behavior, The Journal of Finance 72, 167–206. Bernstein, Asaf, Matthew Gustafson, and Ryan Lewis, 2018, Disaster on the horizon: The price effect of sea level rise, Journal of Financial Economics forthcoming. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan, 2004, How much should we trust differences-in-differences estimates?, Quarterly Journal of Economics 119, 249–275. Bloom, Nicholas, 2009, The impact of uncertainty shocks, Econometrica 77, 623–685. Bloom, Nick, Stephen Bond, and John van Reenen, 2007, Uncertainty and investment dynamics, Review of Economic Studies 74, 391–415. Boudoukh, Jacob, Matthew Richardson, YuQing Shen, and Robert F. Whitelaw, 2007, Do asset pricesreflectfundamentals? FreshlysqueezedevidencefromtheOJmarket,Journal of Financial Economics 83, 397–412. Brown, James R., Matthew T. Gustafson, and Ivan T. Ivanov, 2017, Weathering cash flows, Working Paper. Campbell, Sean D., and Francis X. Diebold, 2005, Weather forecasting for weather derivatives, Journal of the American Statistical Association 100, 6–16. Carney, Mark, 2015, Breaking the tragedy of the horizon - climate change and financial stability, Speech given at Lloyd’s of London (29 September). Carr, P., and L. Wu, 2009, Variance risk premiums, Review of Financial Studies 22, 1311–1341. Choi, Darwin, Zhenyu Gao, and Wenxi Jiang, 2018, Attention to global warming, Working Paper. Dessaint, Olivier, and Adrien Matray, 2017, Do managers overreact to salient risks? Evidence from hurricane strikes, Journal of Financial Economics 126, 97–121. Dew-Becker, Ian, Stefano Giglio, and Bryan Kelly, 2018, Hedging macroeconomic and financial uncertainty and volatility, Working Paper. 32

Duffee, Gregory R, 1995, Stock returns and volatility a firm-level analysis, Journal of Financial Economics 37, 399–420. Engle, Robert, Stefano Giglio, Bryan Kelly, Heebum Lee, and Johannes Stroebel, 2018, Hedging climate change news, Working Paper. Fama, Eugene F., and Kenneth R. French, 1993, Common risk factors in the returns on stocks and bonds, Journal of Financial Economics 33, 3–56. Giglio, Stefano, Matteo Maggiori, Krishna Rao, Johannes Stroebel, and Andreas Weber, 2018, Climate change and long-run discount rate: Evidence from real estate, Working Paper. Grullon, Gustavo, Evgeny Lyandres, and Alexei Zhdanov, 2012, Real options, volatility, and stock returns, The Journal of Finance 67, 1499–1537. Hong, Harrison, Frank W. Li, and Jiangmin Xu, 2019, Climate risks and market efficiency, Journal of Econometrics 208, 265–281. Imberman,ScottA.,AdrianaD.Kugler,andBruceI.Sacerdote,2012,Katrina’schildren: Evidence on the structure of peer effects from hurricane evacuees, American Economic Review 102, 2048– 2082. Jens, Candace E., 2017, Political uncertainty and investment: Causal evidence from U.S. gubernatorial elections, Journal of Financial Economics 124, 563–579. Julio, Brandon, and Youngsuk Yook, 2012, Political uncertainty and corporate investment cycles, Journal of Finance 67, 45–83. Jurado, Kyle, Sydney C. Ludvigson, and Serena Ng, 2015, Measuring uncertainty, American Economic Review 105, 1177–1216. Kelly, Bryan, Lubos Pastor, and Pietro Veronesi, 2016, The price of political uncertainty: Theory and evidence from the option market, Journal of Finance 71, 2417–2480. Martin, Ian, andChristianWagner, 2018, Whatistheexpectedreturnonastock?, Working Paper. Martinez, Andrew, 2018, A false sense of security: The impact of forecast uncertainty on hurricane damages, Working Paper. 33

Merton, Robert C., 1987, A simple model of capital market equilibrium with incomplete information, Journal of Finance 42, 483–510. Murfin, Justin, and Matthew Spiegel, 2019, Is the risk of sea level rise capitalized in residential real estate?, Review of Financial Studies forthcoming. Neumark,David,BrandonWall,andJunfuZhang,2011,Dosmallbusinessescreatemorejobs? New evidence for the united states from the national establishment time series, Review of Economics and Statistics 93, 16–29. Pastor, Lubos, and Pietro Veronesi, 2012, Uncertainty about government policy and stock prices, Journal of Finance 67, 1219–1264. , 2013, Political uncertainty and risk premia, Journal of Financial Economics 110, 520–545. Roll, Richard, 1984, Orange juice and weather, American Economic Review 74, 861–880. Roth Tran, Brigitte, 2019, Divest, disregard, or double down? Philanthropic endowment investments in objectionable firms, American Economic Review: Insights, forthcoming. , and Daniel J. Wilson, 2019, The local economic impact of natural disasters, Working Paper. 34

Var(r ) t t+1 0.164 0.163 0.162 0.161 ϕ 0.2 0.4 0.6 0.8 1.0 σ g =0.05,|μ g |=0.1 σ g =0.05,|μ g |=0.07 σ g =0.05,|μ g |0.05 σ g =0.05,|μ g |=0.0 Figure 1: Variance as a function of the probability of hurricane landfall Thisfigureshowsthetotalvariancepriortolandfall,Var (r )derivedinequation(6),astheprobabilityoflandfall, t i,t+1 φ,variesfrom0to1. Inthisfigure,σ=0.4andσ =0.05. Thefourdashedlineshaveabsolutevaluesof0.1,0.07,0.05, g and0forµ . Thesolidlineshowsthelevelofvarianceconditionalonhurricanelandfall,Var (r |θ=1)=σ2+σ2, g t i,t+1 g as defined in equation (3). 35

IHP Declarations Related to Hurricanes 0 1 2 3 4−5 6−7 (a) From 2002 to 2017 IHP Declarations Related to Hurricanes 0 1 2 3 (b) From 2007 to 2017 Figure 2: US counties with hurricane damage This figure highlights the counties with at least one hurricane related Individual and Household Program (IHP) declaration during the sample period from 2002 to 2017 (Panel A) and the sample period from 2007 to 2017 (Panel B). Source: FederalEmergencyManagementAdministration(FEMA)DisasterDeclarationsSummary(availableathttps: //www.fema.gov/media-library/assets/documents/28318.) 36

Figure 3: Counties with damage from Hurricane Sandy. This figure highlights the counties with Individual and Household Program (IHP) level damages from Hurricane Sandy in 2012. Source: Federal Emergency Management Administration (FEMA) Disaster Declarations Summary (available at https://www.fema.gov/media-library/assets/documents/28318); Arizona State University, Spatial Hazard Events and Losses Database for the United States (SHELDUS). 37

Figure 4: Example of a five-day forecast of a hurricane. ThisfigurefromNOAAillustratesthefive-dayforecastforHurricaneSandyonOctober27,2012. Weobtaintheraw data underpinning such hurricane forecast visualizations for our analysis. Source: NationalHurricaneCenterSANDYGraphicsArchive,availableathttps://www.nhc.noaa.gov/archive/2012/ graphics/al18/loop PROB64.shtml. 38

October 26, 2012, 4 days before landfall October 27, 2012, 3 days before landfall October 28, 2012, 2 days before landfall October 29, 2012, 1 day before landfall ≥1 percent ≥10 percent ≥20 percent ≥50 percent Figure 5: Hurricane Forecasts At Different Time Frames and Wind Speed Probability Thresholds Eachmapshowsthecountiesindicatedasbeingat-riskforHurricaneSandygiventhenumberofdaysbeforelandfall in each row and the wind speed probability threshold in each column. Source: NationalHurricaneCenterWindSpeedProbabilities,availableathttps://www.nhc.noaa.gov/archive/text/. 39

0.0 2.5 5.0 7.5 10.0 (a) Year 2010 0.0 2.5 5.0 7.5 10.0 (b) Year 2014 Figure 6: Firm establishments by county Thischartplotscountiesbasedonthenumberofestablishmentslocatedinthatcountyfortheyears2010(PanelA) and 2014 (Panel B). Only firms that could be mapped to CRSP-Compustat are included. The counties are sorted into deciles based on the number of establishments. Source: Walls & Associates, National Establishment Time-Series (NETS) Database 2014. 40

0.2 l l l 0.1 l l l l 0.0 −0.1 0 10 20 30 40 50 Minimum probability of hurricane force winds noitpecni−erp susrev egnahc VI Days Before Landfall: 1 (a) One day before hurricane landfall/dissipation 0.2 l l 0.1 l l l l 0.0 −0.1 0 10 20 30 40 50 Minimum probability of hurricane force winds noitpecni−erp susrev egnahc VI Days Before Landfall: 2 (b) Two days before hurricane landfall/dissipation Figure 7: Implied volatility difference for firms exposed to a hurricane forecast path This chart plots the implied volatility difference for firms exposed to hurricane forecast path one and two days before hurricane landfall/dissipation. This corresponds to the λ coefficient estimate in regression equation (12) by probabilitythreshold. ThecompleteresultsoftheregressionsarepresentedinTable7. Theblueshadedregionabove and below the plotted line depicts the 95 percent confidence interval of the estimates. Source: OptionMetrics, Ivy DB US, Wharton Research Data Services, http://wrds-web.wharton.upenn.edu/wrds/; Walls & Associates, National Establishment Time-Series (NETS) Database 2014; National Hurricane Center Wind Speed Probabilities, available at https://www.nhc.noaa.gov/archive/text/. 41

1.25 0.75 0.25 -0.25 -0.75 -1.25 -1.75 -2.25 1 10 20 30 40 50 )%(noitcaerrednu / noitcaerrevo VI .gvA 1 day before landfall / dissipation Overprediction (DamageExposure < ForecastExposure) Underprediction (DamageExposure > ForecastExposure) Probability (%) of hurricane hit ≥ X (a) Implied volatility overreaction / underreaction 100 80 60 40 20 0 -20 -40 -60 -80 -100 1 10 20 30 40 50 noitcaerrednu / noitcaerrevo pac tekram .gvA )snoillim( 1 day before landfall / dissipation Overprediction (DamageExposure < ForecastExposure) Underprediction (DamageExposure > ForecastExposure) Probability (%) of hurricane hit ≥X (b) Market cap overreaction / underreaction 5,000 4,500 4,000 3,500 3,000 2,500 2,000 1,500 1,000 500 0 1 10 20 30 40 50 smrif fo rebmuN 1 day before landfall / dissipation Overprediction (DamageExposure < ForecastExposure) Underprediction (DamageExposure > ForecastExposure) Probability (%) of hurricane hit ≥X (c) Number of firms overpredicted / underpredicted Figure 8: Economic effects of differences in forecast and damage exposure PanelAplotstheaverageoverreaction(underreaction)inimpliedvolatilityduetoanoverprediction(underprediction), i.e.,thefirm’sforecastexposuretoahurricaneisgreater(smaller)thanthefirm’seventualdamageexposure. PanelB, depicts the average implied volatility overreaction (underreaction) multiplied with the average market capitalization of the respective firms. Panel C plots the number of firms for which the forecast exposure to a hurricane is greater (smaller)thanthefirm’seventualdamageexposuresummedacrossallthehurricanesfrom2007to2017. Theanalysis is based on hurricane forecasts one day before landfall or dissipation. The geographic footprint of firms is based on the number of establishments in counties. The damage exposure is measured with FEMA and SHELDUS data. Source: OptionMetrics, Ivy DB US, Wharton Research Data Services, http://wrds-web.wharton.upenn.edu/wrds/; Walls & Associates, National Establishment Time-Series (NETS) Database 2014; Center for Research in Security Prices,CRSP/CompustatMergedDatabase;NationalHurricaneCenterWindSpeedProbabilities,availableathttps: //www.nhc.noaa.gov/archive/text/;FederalEmergencyManagementAdministration(FEMA)DisasterDeclarations Summary (available at https://www.fema.gov/media-library/assets/documents/28318); Arizona State University, Spatial Hazard Events and Losses Database for the United States (SHELDUS). 42

Table 1: Summary statistics of hurricane damage data Thistablereportssummarystatisticsonthedamagedataforeachhurricanefrom2002to2017basedonFEMAand SHELDUS. Counties are eligible for individual and household program (IHP) assistance if there is a FEMA disaster declaration due to significant damage from a hurricane on a per capita basis. Mean Median Std Dev Min Max Obs FEMA Counties with IHP declarations 80 72 53 16 254 20 SHELDUS Counties with property damage 175 137 126 24 466 20 SHELDUS Counties with IHP-worthy damages 100 77 88 5 345 20 SHELDUS Property damages estimate ($millions) 12,158 2,808 24,613 32 89,432 20 Source:Federal Emergency Management Administration (FEMA) Disaster Declarations Summary (available at https://www.fema.gov/media-library/assets/documents/28318); Arizona State University, Spatial Hazard Events and Losses Database for the United States (SHELDUS). Table 2: Summary statistics of hurricane forecast data ThistablereportssummarystatisticsofNOAAwindspeedforecastsfrom2007to2017forstormsthatareforecasted to make landfall within five days with wind speeds of at least 64KT with a given minimum probability. Panel A reports the mean, median, and standard deviation of the number of county-dates observations for which we have hurricane forecasts for each storm at a given probability threshold. Panel B presents the observation count by days to resolution (hurricane landfall or, in the case of “misses”, dissipation) at a given probability threshold. Panel A: Summary statistics of county date forecast observations per storm Probability ≥ 1 10 20 40 50 N Storms 49 17 14 9 9 N County Days 14,988 2,093 913 414 335 Mean 305.878 42.714 18.633 8.449 6.837 Median 124.000 0.000 0.000 0.000 0.000 Std. Dev. 402.974 91.761 43.723 20.857 18.004 Panel B: Number of county date forecast observations Days to dissipation or Probability ≥ landfall 1 10 20 40 50 1 2,251 536 371 239 199 2 3,131 678 320 149 122 3 3,198 545 159 14 14 4 2,431 187 37 12 0 5 1,929 101 21 0 0 Source: NationalHurricaneCenterWindSpeedProbabilities,availableathttps://www.nhc.noaa.gov/archive/text/. 43

Table 3: Firm establishment and sales summary statistics This table reports the summary statistics on the number of establishments and amount of sales (in USD) in the NETS dataset from 2002 to 2017 for the firms that were matched to equity data from CRSP-Compustat. Number of firms with establishment/sales data Establishments 4,197 Sales 4,187 Statistics by firm-year Avg SD 10% percentile 50% percentile 90% percentile Establishments 69.397 323.093 1.000 4.000 110.000 Sales (in millions) 524.601 2,465.746 0.310 31.642 959.216 Statistics by county-year Avg SD 10% percentile 50% percentile 90% percentile Establishments 66.659 207.904 2.000 14.000 143.000 Sales (in millions) 503.377 2,088.409 2.408 58.780 894.028 Statistics by county-year for hurricane damaged counties Avg SD 10% percentile 50% percentile 90% percentile Establishments 88.054 241.813 3.000 21.000 214.000 Sales (in millions) 675.404 2,831.049 2.923 87.663 1,260.017 Source: Walls & Associates, National Establishment Time-Series (NETS) Database 2014; Center for Research in Security Prices, CRSP/Compustat Merged Database. 44

Table 4: Summary statistics of implied volatility Thistablereportsthesummarystatisticsonthesingle-stockoptionsdatafromJanuary1,2002toDecember31,2017 from OptionMetrics including the number of observations (N), mean, median, standard deviation, 25th, 75th, 10th, and 90th percentiles. Panel A includes the options dataset once merged with CRSP-Compustat. Panel B further restricts the sample to firms appearing at least once in the NETS firm establishment data. Panel A: Firms matched to CRSP-Compustat N Mean Median Stdev 25th 75th 10th 90th IV 9,420,182 0.453 0.384 0.273 0.272 0.551 0.205 0.778 i,t (cid:16) (cid:17) log IVi,t 9,420,182 0.001 0.000 0.124 -0.042 0.045 -0.105 0.110 IVi,t−1 Days to expiry 9,420,182 38.852 29.000 35.427 17.000 40.000 11.000 93.000 i,t Total open interest 9,420,182 2,396.119 267.000 11,321.761 54.000 1,347.000 13.000 5,106.000 i,t Panel B: Firms matched to CRSP-Compustat and NETS N Mean Median Stdev 25th 75th 10th 90th IV 3,866,672 0.440 0.372 0.265 0.267 0.530 0.202 0.750 i,t (cid:16) (cid:17) log IVi,t 3,866,672 0.001 0.000 0.126 -0.043 0.045 -0.106 0.112 IVi,t−1 Days to expiry 3,866,672 39.243 29.000 35.577 17.000 40.000 11.000 93.000 i,t Total open interest 3,866,672 2,080.397 234.000 7,681.694 50.000 1,194.000 12.000 4,584.000 i,t Source: OptionMetrics, Ivy DB US, Wharton Research Data Services, http://wrds-web.wharton.upenn.edu/wrds/; Walls & Associates, National Establishment Time-Series (NETS) Database 2014; Center for Research in Security Prices, CRSP/Compustat Merged Database. 45

Table 5: Hurricane effects on implied volatility This table reports the coefficients and t-statistics (in parentheses) when estimating the panel model in equation (9). The dependent variable is the percent change in implied volatility of firm i from the day before hurricane inception day(T∗)to5tradingdaysafterlandfall(T .) DamageExposure istheshareofafirm’sgeographicfootprintthat h h i,Th is exposed to the disaster. In Panels A and B, firm geographic exposure to a hurricane is based on establishments andsalespercounty,respectively. DamagedcountiesareidentifiedwithFEMAdataandFEMAdataenhancedwith SHELDUSdata. Thedataarefrom2002to2017. Resultsarealsoshownforthesubsamplefrom2007to2017,which correspondstothetimeperiodforwhichwehavehurricaneforecastdatausedinsubsequentanalyses. Thestandard errors have been clustered by firm headquarter county. Industry and time fixed effects are used. The significance of the coefficient estimate is indicated by * for p < 0.10, ** for p < 0.05, and *** for p < 0.01. PanelA:Firms’hurricaneexposurebasedonestablishments (cid:16) (cid:17) Dependentvariable: ChangeinIV,log IV i,Th+5 /IV i,T h ∗ 2002-2017 2007-2017 FEMA FEMA+SHELDUS FEMA FEMA+SHELDUS DamageExposure 0.072∗∗∗ 0.053∗∗∗ 0.071∗∗∗ 0.057∗∗∗ 0.085∗∗∗ 0.065∗∗∗ 0.086∗∗∗ 0.071∗∗∗ i,Th (4.345) (3.222) (4.316) (3.428) (4.250) (2.933) (4.181) (3.225) AdjustedR2 (%) 13.379 14.128 13.574 14.336 14.841 15.793 14.862 15.824 Obs. total 12,156 12,156 12,621 12,621 7,204 7,204 7,204 7,204 Obs. firmexposure>0% 5,387 5,387 5,938 5,938 3,020 3,020 3,261 3,261 Obs. firmexposure≥20% 805 805 974 974 479 479 593 593 Obs. firmexposure≥50% 286 286 317 317 175 175 199 199 Hurricanes 19 19 20 20 10 10 10 10 IndustryFE Yes No Yes No Yes No Yes No TimeFE Yes No Yes No Yes No Yes No Industry×TimeFE No Yes No Yes No Yes No Yes PanelB:Firms’hurricaneexposurebasedonsales (cid:16) (cid:17) Dependentvariable: ChangeinIV,log IV i,Th+5 /IV i,T h ∗ 2002-2017 2007-2017 FEMA FEMA+SHELDUS FEMA FEMA+SHELDUS DamageExposure 0.051∗∗∗ 0.038∗∗∗ 0.050∗∗∗ 0.039∗∗∗ 0.063∗∗∗ 0.048∗∗∗ 0.063∗∗∗ 0.052∗∗∗ i,Th (3.554) (2.738) (3.726) (3.084) (3.810) (2.771) (3.964) (3.155) AdjustedR2 (%) 13.368 14.139 13.556 14.337 14.829 15.797 14.845 15.821 Obs. total 12,126 12,126 12,590 12,590 7,194 7,194 7,194 7,194 Obs. firmexposure>0% 5,365 5,365 5,916 5,916 3,006 3,006 3,249 3,249 Obs. firmexposure≥20% 808 808 953 953 488 488 595 595 Obs. firmexposure≥50% 391 391 447 447 235 235 279 279 Hurricanes 19 19 20 20 10 10 10 10 IndustryFE Yes No Yes No Yes No Yes No TimeFE Yes No Yes No Yes No Yes No Industry×TimeFE No Yes No Yes No Yes No Yes Source: OptionMetrics, Ivy DB US, Wharton Research Data Services, http://wrds-web.wharton.upenn.edu/wrds/; Walls & Associates, National Establishment Time-Series (NETS) Database 2014; Center for Research in Security Prices, CRSP/Compustat Merged Database; Federal Emergency Management Administration (FEMA) Disaster Declarations Summary (available at https://www.fema.gov/media-library/assets/documents/28318); Arizona State University, Spatial Hazard Events and Losses Database for the United States (SHELDUS). 46

Table 6: Abnormal returns post hurricane landfall Thistablereportsdifferencesincumulativeabnormalreturnspostlandfallbetweenfirmswithexposure(treated)and firmswithoutexposure(control)tothehurricanedisasterregion. Thedifferencesarereportedforthemeanandnine percentiles of the return distributions of the treated and control firms. The differences are estimated for two time periods: from landfall to 5 trading days and 120 trading days after landfall, respectively. The abnormal returns are estimated based on the Fama-French three factor model. FEMA and SHELDUS data are used to identify counties that have been hit by a hurricane. For Panel A, the hit firms are defined as firms that have 50 percent or more of theirestablishmentsinthecountiesofthedisasterregion,andforPanelB,50percentormoreofthesaleshavetobe locatedinthedisasterareacounties. Weexcludehurricanesthataffectedlessthan25firms. Thedataarefrom2002 to 2017. The standard errors are bootstrapped and clustered by county (headquarter location). The significance of the difference in abnormal returns is indicated by * for p < 0.10, ** for p < 0.05, and *** for p < 0.01. Panel A: Hit firms selected based on exposure in disaster region ≥ 50% From 0 to 5 days post landfall From 0 to 120 days post landfall Cumulative r difference T-stat Cumulative r difference T-stat Mean 0.103 0.197 -6.489 -1.562 Percentiles 10% -0.315 -0.637 -13.754∗∗∗ -3.556 20% -0.479 -0.801 -14.781∗∗∗ -3.550 30% -0.214 -0.681 -10.241∗∗ -2.310 40% -0.241 -1.003 -7.871∗∗ -2.213 50% -0.500∗ -1.858 -7.491∗∗∗ -3.402 60% -0.357 -1.215 -8.262∗∗∗ -2.734 70% -0.143 -0.327 -4.529 -1.395 80% -0.199 -0.423 -3.795 -1.061 90% 0.927 0.523 -5.708 -0.910 Hit firms 415 382 Control firms 8,376 7,027 Panel B: Hit firms selected based on sales in disaster region ≥ 50% From 0 to 5 days post landfall From 0 to 120 days post landfall Cumulative r difference T-stat Cumulative r difference T-stat Mean 0.131 0.248 -6.501∗∗ -1.968 Percentiles 10% -0.661 -1.139 -12.215∗∗∗ -2.577 20% -0.438 -0.823 -10.233∗∗ -2.502 30% -0.313 -0.967 -6.862∗∗ -2.528 40% -0.314 -0.947 -5.394∗∗ -2.195 50% -0.318 -0.848 -5.697∗∗∗ -2.693 60% -0.054 -0.144 -5.633∗∗∗ -2.923 70% 0.278 0.439 -4.838∗∗ -2.127 80% 0.251 0.382 -3.559 -1.215 90% 0.948 0.828 -6.667 -1.120 Hit firms 641 622 Control firms 10,115 9,795 Source: Walls & Associates, National Establishment Time-Series (NETS) Database 2014; Center for Research in SecurityPrices,CRSP/CompustatMergedDatabase;FederalEmergencyManagementAdministration(FEMA)Disaster Declarations Summary (available at https://www.fema.gov/media-library/assets/documents/28318); Arizona State University, Spatial Hazard Events and Losses Database for the United States (SHELDUS). 47

ytilitalov deilpmi dna htap enacirruh detsaceroF :7 elbaT eht ni )tnecrep ni( egnahc eht si elbairav tnedneped ehT .)21( noitauqe ni ledom lenap eht gnitamitse nehw scitsitats tset dna stneicffieoc eht stroper elbat sihT woh serusaem elbairav tnednepedni ehT .enacirruh eht fo , T ,noitapissid ro llafdnal erofeb syad Γ ot enacirruh eht fo noitpecni morf i mrfi fo ytilitalov deilpmi h .enacirruh eht fo noitapissid ro llafdnal eht erofeb syad Γ enacirruh a fo htap detsacerof eht ot desopxe si mrfi a fo tnirptoof cihpargoeg eht fo )tnecrep ni( hcum rofytnuocrepselasnodesabdnaAlenaProfytnuocrepstnemhsilbatsenodesabsimrfiafoenacirruhaoterusopxeehterusaemotdesutnirptoofcihpargoegehT ytilibaborpdedulcniehT .ytnuochcaerofenacirruhcfiicepsaybtihgniebfoytilibaborpasevigdnaAAONmorfsienacirruhehtfohtapdetsacerofehT.BlenaP atad ehT .htap detsacerof eht ni seitnuoc eht ot selas ro stnemhsilbatse ni %02 revo fo erusopxe na htiw smrfi 52 dna senacirruh eerht tsael ta evah sdlohserht dexfi emit dna yrtsudnI .)noitacol retrauqdaeh( ytnuoc yb deretsulc era srorre dradnats ehT .stats-t eht era sesehtnerap ni seulav ehT .7102 ot 7002 morf era .10.0 < p rof *** dna ,50.0 < p rof ** ,01.0 < p rof * yb detacidni si etamitse tneicffieoc eht fo ecnacfiingis ehT .desu era stceffe stnemhsilbatsenodesaberusopxeenacirruh’smriF:AlenaP (cid:17) (cid:16) ∗ hT,iVI/Γ−hT,iVI gol,noitapissid/llafdnalerofebsyadΓotnoitpecnienacirruhmorfVIniegnahC :elbairavtnednepeD syaD5 syaD4 syaD3 syaD2 yaD1 Γ %01 %1 %01 %1 %02 %01 %1 %04 %02 %01 %1 %05 %04 %02 %01 %1 ≥tihenacirruhfoytilibaborP 600.0 900.0 ∗∗∗721.0 ∗410.0 ∗∗∗931.0 ∗∗∗001.0 ∗∗410.0 ∗∗∗171.0 ∗∗∗760.0 ∗∗∗170.0 ∗710.0 ∗∗∗441.0 ∗∗∗731.0 ∗∗∗280.0 ∗∗∗740.0 ∗∗∗340.0 Γ−hT,ierusopxEtsaceroF )131.0( )449.0( )952.3( )327.1( )276.5( )885.3( )321.2( )844.5( )183.4( )446.4( )897.1( )770.4( )417.3( )677.4( )551.4( )461.4( 281.9 840.01 201.52 907.51 033.51 632.41 847.01 679.81 556.61 597.41 779.11 503.32 303.32 629.91 872.12 380.61 2RdetsujdA 373,2 023,01 500,4 178,11 257,4 640,7 932,41 479,3 364,5 570,7 800,61 543,2 543,2 853,5 070,6 061,12 latoT .sbO 076 055,4 901,1 920,6 382,1 759,1 546,6 829 513,1 178,1 902,6 746 407 734,1 717,1 137,5 %0>erusopxEtsaceroF .sbO 43 601,1 68 358,1 88 441 929,1 66 611 361 442,1 16 07 621 571 804 %02≥erusopxEtsaceroF .sbO 3 31 5 51 6 9 81 5 7 9 02 3 3 7 8 72 senacirruhforebmuN seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFyrtsudnI seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFemiT selasnodesaberusopxeenacirruh’smriF:BlenaP (cid:17) (cid:16) ∗ hT,iVI/Γ−hT,iVI gol,noitapissid/llafdnalerofebsyadΓotnoitpecnienacirruhmorfVIniegnahC :elbairavtnednepeD syaD5 syaD4 syaD3 syaD2 yaD1 Γ %01 %1 %01 %1 %02 %01 %1 %04 %02 %01 %1 %05 %04 %02 %01 %1 ≥tihenacirruhfoytilibaborP 020.0 600.0 ∗∗∗121.0 ∗410.0 ∗∗∗331.0 ∗∗780.0 ∗∗510.0 ∗∗∗461.0 ∗∗∗080.0 ∗∗∗070.0 ∗∗810.0 ∗∗∗051.0 ∗∗∗431.0 ∗∗∗470.0 ∗∗∗440.0 ∗∗∗630.0 Γ−hT,ierusopxEtsaceroF )256.0( )428.0( )541.3( )259.1( )955.4( )425.2( )525.2( )180.5( )772.5( )983.5( )111.2( )098.4( )719.3( )141.4( )326.4( )034.3( 451.9 230.01 102.52 777.51 664.51 082.41 677.01 501.91 537.61 838.41 699.11 485.32 145.32 730.02 543.12 431.61 2RdetsujdA 073,2 703,01 000,4 558,11 647,4 730,7 022,41 969,3 654,5 660,7 889,51 243,2 243,2 053,5 160,6 231,12 latot .sbO 366 925,4 001,1 000,6 572,1 939,1 916,6 819 403,1 558,1 771,6 146 007 624,1 607,1 066,5 %0>erusopxEtsaceroF .sbO 93 389 69 746,1 59 441 647,1 47 321 271 681,1 66 67 341 981 954 %02≥erusopxEtsaceroF .sbO 3 31 5 51 6 9 81 5 7 9 02 3 3 7 8 72 senacirruhforebmuN seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFyrtsudnI seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFemiT tnemhsilbatsE lanoitaN ,setaicossA & sllaW ;/sdrw/ude.nnepu.notrahw.bew-sdrw//:ptth ,secivreS ataD hcraeseR notrahW ,SU BD yvI ,scirteMnoitpO :ecruoS deepS dniW retneC enacirruH lanoitaN ;esabataD degreM tatsupmoC/PSRC ,secirP ytiruceS ni hcraeseR rof retneC ;4102 esabataD )STEN( seireS-emiT ./txet/evihcra/vog.aaon.chn.www//:sptth ta elbaliava ,seitilibaborP 48

ytilitalov deilpmi dna htap enacirruh detsacerof rof noitacfiiceps evitanretlA :8 elbaT deilpmi eht tcartbus ew ,i mrfi dna enacirruh hcae roF .)31( noitauqe ni debircsed DVI rof tats-t dna ,DVI detoned ,naem elpmas eht stroper elbat sihT h,i a sdeecxe htap s’enacirruh eht ni seitnuoc ot erusopxe s’mrfi eht nehw yad eht no VI eht morf enacirruh eht fo noitpecni eht erofeb yad eht no )VI( ytilitalov ot desopxe ton era taht smrfi fo VI ni ecnereffid gol naem eht yb denaemed neht si VI ni ecnereffid gol ehT .dedulcxe era syad llafdnaL .tnecrep X fo dlohserht no desab dna A lenaP rof ytnuoc rep stnemhsilbatse no desab si mrfi a fo enacirruh a ot erusopxe eht erusaem ot desu tnirptoof cihpargoeg ehT .enacirruh eht ehT .ytnuochcaerofenacirruhcfiicepsaybtihgniebfoytilibaborpasevigdnaAAONmorfsienacirruhehtfohtapdetsacerofehT.BlenaProfytnuocrepselas fo ecnacfiingis ehT .)noitacol retrauqdaeh( ytnuoc yb deretsulc era srorre dradnats ehT .stats-t eht era sesehtnerap ni seulav ehT .7102 ot 7002 morf era atad .10.0 < p rof *** dna ,50.0 < p rof ** ,01.0 < p rof * yb detacidni si etamitse tneicffieoc eht stnemhsilbatsenodesaberusopxeenacirruh’smriF:AlenaP %05 %52 %01 ≥htapenacirruhoterusopxE %05 %04 %02 %01 %1 %05 %04 %02 %01 %1 %05 %04 %02 %01 %1 ≥tihenacirruhfoytilibaborP ∗∗∗113.11 ∗∗∗404.9 ∗∗∗673.3 ∗∗871.2 ∗∗589.0 ∗∗∗594.21 ∗∗∗221.11 ∗∗∗656.4 ∗∗∗984.3 972.0 ∗∗925.8 ∗915.6 ∗∗054.2 ∗∗837.1 843.0- DVI )538.5( )962.3( )638.2( )490.2( )560.2( )606.3( )540.3( )531.4( )031.3( )827.0( )835.2( )517.1( )392.2( )312.2( )574.1-( 72 23 18 811 282,1 16 86 561 442 577,2 541 871 414 946 880,6 smrfiforebmuN 4 4 7 11 73 5 5 7 11 93 5 5 9 31 04 senacirruhforebmuN selasnodesaberusopxeenacirruh’smriF:BlenaP %05 %52 %01 ≥htapenacirruhoterusopxE %05 %04 %02 %01 %1 %05 %04 %02 %01 %1 %05 %04 %02 %01 %1 ≥tihenacirruhfoytilibaborP ∗∗∗843.11 ∗∗690.9 ∗∗∗288.3 ∗∗127.2 044.0 ∗∗∗436.11 ∗∗∗336.01 ∗∗∗387.4 ∗∗664.2 991.0 ∗∗∗712.21 ∗494.7 ∗∗778.2 ∗∗788.1 531.0- DVI )994.3( )814.2( )894.5( )545.2( )402.1( )672.4( )298.2( )998.7( )169.1( )016.0( )047.3( )728.1( )282.2( )699.1( )994.0-( 84 75 221 471 607,1 37 48 391 972 297,2 231 061 453 835 887,4 smriF 5 5 8 21 73 5 5 9 21 73 6 6 01 41 93 senacirruhforebmuN tnemhsilbatsE lanoitaN ,setaicossA & sllaW ;/sdrw/ude.nnepu.notrahw.bew-sdrw//:ptth ,secivreS ataD hcraeseR notrahW ,SU BD yvI ,scirteMnoitpO :ecruoS deepS dniW retneC enacirruH lanoitaN ;esabataD degreM tatsupmoC/PSRC ,secirP ytiruceS ni hcraeseR rof retneC ;4102 esabataD )STEN( seireS-emiT ./txet/evihcra/vog.aaon.chn.www//:sptth ta elbaliava ,seitilibaborP 49

snruter kcots dna htap enacirruh detsaceroF :9 elbaT )tnecrep ni( nruter kcots eht si elbairav tnedneped ehT .)41( noitauqe ni ledom lenap eht gnitamitse nehw scitsitats tset dna stneicffieoc eht stroper elbat sihT fo )tnecrep ni( hcum woh serusaem elbairav tnednepedni ehT .enacirruh eht fo noitapissid ro llafdnal erofeb syad Γ ot enacirruh eht fo noitpecni morf i mrfi fo cihpargoeg ehT .enacirruh eht fo ecnaraeppasid ro llafdnal eht erofeb syad Γ enacirruh a fo htap detsacerof eht ot desopxe si mrfi a fo tnirptoof cihpargoeg eht ehT.BlenaProfytnuocrepselasnodesabdnaAlenaProfytnuocrepstnemhsilbatsenodesabsimrfiafoenacirruhaoterusopxeehterusaemotdesutnirptoof sdlohserht ytilibaborp dedulcni ehT .ytnuoc hcae rof enacirruh cfiiceps a yb tih gnieb fo ytilibaborp a sevig dna AAON morf si enacirruh eht fo htap detsacerof era atad ehT .htap detsacerof eht ni seitnuoc eht ot selas ro stnemhsilbatse ni tnecrep 02 revo fo erusopxe na htiw smrfi 52 dna senacirruh eerht tsael ta evah stceffedexfiemitdnayrtsudnI .)noitacolretrauqdaeh(ytnuocybderetsulcerasrorredradnatsehT .stats-tehterasesehtnerapniseulavehT .7102ot7002morf .10.0 < p rof *** dna ,50.0 < p rof ** ,01.0 < p rof * yb detacidni si etamitse tneicffieoc eht fo ecnacfiingis ehT .desu era stnemhsilbatsenodesaberusopxeenacirruh’smriF:AlenaP Γ−hT:∗ hT,ir,noitapissid/llafdnalerofebsyadΓotnoitpecnienacirruhmorfnruterevitalumuC :elbairavtnednepeD syaD5 syaD4 syaD3 syaD2 yaD1 Γ %01 %1 %01 %1 %02 %01 %1 %04 %02 %01 %1 %05 %04 %02 %01 %1 ≥tihenacirruhfoytilibaborP 400.0 200.0- ∗∗∗880.0- 700.0- ∗∗650.0- ∗∗840.0- 500.0- ∗∗∗190.0- ∗∗∗740.0- ∗∗∗040.0- 510.0- ∗∗∗270.0- ∗∗∗660.0- ∗∗∗230.0- ∗∗∗030.0- 510.0- Γ−hT,ierusopxEtsaceroF )294.0( )005.0-( )171.3-( )526.1-( )045.2-( )165.2-( )810.1-( )864.4-( )991.6-( )608.3-( )914.1-( )367.3-( )814.3-( )517.3-( )873.3-( )301.1-( 577.91 205.4 418.2 863.2 313.21 093.5 527.0 938.02 798.8 539.8 469.0 520.1 320.1 795.1 500.2 416.3 2RdetsujdA 005,2 173,11 343,4 280,41 270,5 821,8 901,61 351,4 690,6 617,7 705,71 844,2 844,2 413,6 532,7 260,52 latot .sbO 176 231,5 831,1 098,6 423,1 880,2 543,7 449 163,1 739,1 366,6 686 747 355,1 648,1 385,6 %0>erusopxEtsaceroF .sbO 53 763,1 78 800,2 39 551 951,2 76 511 561 204,1 36 07 231 471 365 %02≥erusopxEtsaceroF .sbO 3 31 5 61 6 9 81 5 7 9 02 3 3 7 8 82 senacirruhforebmuN seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFyrtsudnI seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFemiT selasnodesaberusopxeenacirruh’smriF:BlenaP Γ−hT:∗ hT,ir,noitapissid/llafdnalerofebsyadΓotnoitpecnienacirruhmorfnruterevitalumuC :elbairavtnednepeD syaD5 syaD4 syaD3 syaD2 yaD1 Γ %01 %1 %01 %1 %02 %01 %1 %04 %02 %01 %1 %05 %04 %02 %01 %1 ≥tihenacirruhfoytilibaborP 400.0- 300.0- ∗∗∗370.0- ∗∗800.0- ∗∗∗840.0- ∗∗∗930.0- 400.0- ∗∗∗870.0- ∗∗∗140.0- ∗∗∗230.0- 110.0- ∗∗∗750.0- ∗∗∗350.0- ∗∗∗520.0- ∗∗∗420.0- 110.0- Γ−hT,ierusopxEtsaceroF )545.0-( )530.1-( )271.4-( )383.2-( )388.2-( )126.2-( )610.1-( )464.5-( )339.6-( )032.3-( )173.1-( )226.3-( )104.3-( )120.4-( )517.3-( )450.1-( 987.91 015.4 128.2 673.2 943.21 083.5 427.0 619.02 809.8 529.8 269.0 220.1 120.1 695.1 500.2 216.3 2RdetsujdA 894,2 563,11 243,4 670,41 170,5 621,8 001,61 351,4 590,6 417,7 994,71 844,2 844,2 413,6 432,7 350,52 latot .sbO 466 211,5 031,1 268,6 713,1 270,2 123,7 539 053,1 129,1 136,6 186 447 245,1 438,1 115,6 %0>erusopxEtsaceroF .sbO 14 362,1 401 568,1 111 461 700,2 97 621 771 173,1 37 78 651 102 736 %02≥erusopxEtsaceroF .sbO 3 31 5 61 6 9 81 5 7 9 02 3 3 7 8 82 senacirruhforebmuN seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFyrtsudnI seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFemiT degreM tatsupmoC/PSRC ,secirP ytiruceS ni hcraeseR rof retneC ;4102 esabataD )STEN( seireS-emiT tnemhsilbatsE lanoitaN ,setaicossA & sllaW :ecruoS ./txet/evihcra/vog.aaon.chn.www//:sptth ta elbaliava ,seitilibaborP deepS dniW retneC enacirruH lanoitaN ;esabataD 50

nosirapmoc stsacerof AAON dna tekraM :01 elbaT )tnecrep ni( egnahc eht si elbairav tnedneped ehT .)51( noitauqe ni ledom lenap eht gnitamitse nehw scitsitats tset dna stneicffieoc eht stroper elbat sihT elbairav tnednepedni ehT .enacirruh eht fo , T ,noitapissid ro llafdnal erofeb syad Γ ot enacirruh eht fo noitpecni morf i mrfi fo ytilitalov deilpmi eht ni h syadΓenacirruhafohtapdetsacerofehtotdesopxesimrfiafotnirptoofcihpargoegehtfo)tnecrepni(hcumwohserusaem erusopxEtsaceroFenacirruH Γ−hT,i eht ot erusopxe s’mrfi eht neewteb ecnereffid eht serusaem noitciderPrednU elbairav tnednepedni ehT .enacirruh eht fo noitapissid ro llafdnal eht erofeb Γ−hT,i ot desu tnirptoof cihpargoeg ehT .)61( noitauqe ni nwohs sa ,enacirruh a fo htap detsacerof eht ot erusopxe s’mrfi eht dna noiger egamad enacirruh lautneve htap detsacerof ehT .B lenaP rof ytnuoc rep selas no desab dna A lenaP rof ytnuoc rep stnemhsilbatse no desab si mrfi a fo enacirruh a ot erusopxe eht erusaem tsael ta evah sdlohserht ytilibaborp dedulcni ehT .ytnuoc hcae rof enacirruh cfiiceps a yb tih gnieb fo ytilibaborp a sevig dna AAON morf si enacirruh eht fo .7102 ot 7002 morf era atad ehT .htap detsacerof eht ni seitnuoc eht ot selas ro stnemhsilbatse ni %02 revo fo erusopxe na htiw smrfi 52 dna senacirruh eerht ehT .desu era stceffe dexfi emit dna yrtsudnI .)noitacol retrauqdaeh( ytnuoc yb deretsulc era srorre dradnats ehT .stats-t eht era sesehtnerap ni seulav ehT .10.0 < p rof *** dna ,50.0 < p rof ** ,01.0 < p rof * yb detacidni si etamitse tneicffieoc eht fo ecnacfiingis stnemhsilbatsenodesaberusopxeenacirruh’smriF:AlenaP (cid:17) (cid:16) ∗ hT,iVI/Γ−hT,iVI gol,noitapissid/llafdnalerofebsyadΓotnoitpecnienacirruhmorfVIniegnahC :elbairavtnednepeD syaD5 syaD4 syaD3 syaD2 yaD1 Γ %01 %1 %01 %1 %02 %01 %1 %04 %02 %01 %1 %05 %04 %02 %01 %1 ≥tihenacirruhfoytilibaborP 600.0 900.0 ∗∗∗621.0 ∗410.0 ∗∗∗831.0 ∗∗∗001.0 ∗∗410.0 ∗∗∗171.0 ∗∗∗860.0 ∗∗∗170.0 ∗710.0 ∗∗∗541.0 ∗∗∗831.0 ∗∗∗280.0 ∗∗∗840.0 ∗∗∗340.0 Γ−hT,ierusopxEtsaceroF )231.0( )399.0( )891.3( )237.1( )156.5( )875.3( )760.2( )473.5( )983.4( )916.4( )377.1( )540.4( )907.3( )896.4( )632.4( )271.4( 100.0 320.0 640.0 310.0 120.0 610.0 930.0- 210.0 130.0 120.0 610.0- 020.0 810.0 520.0 ∗660.0 630.0 Γ−hT,inoitciderPrednU )260.0( )809.0( )095.1( )652.0( )238.0( )276.0( )051.1-( )444.0( )479.0( )486.0( )262.0-( )885.0( )945.0( )118.0( )629.1( )100.1( 341.9 150.01 251.52 307.51 423.51 922.41 357.01 269.81 666.61 297.41 379.11 682.32 282.32 729.91 723.12 580.61 )%(2RdetsujdA 373,2 023,01 500,4 178,11 257,4 640,7 932,41 479,3 364,5 570,7 800,61 543,2 543,2 853,5 070,6 061,12 latot .sbO 076 055,4 901,1 920,6 382,1 759,1 546,6 829 513,1 178,1 902,6 746 407 734,1 717,1 137,5 %0>erusopxEtsaceroF .sbO 43 601,1 68 358,1 88 441 929,1 66 611 361 442,1 16 07 621 571 804 %02≥erusopxEtsaceroF .sbO 31 363 03 076 33 84 786 32 74 26 244 22 52 64 86 751 %05≥erusopxEtsaceroF .sbO 837 530,1 951,1 156 804,1 935,1 429 835,1 416,1 444,1 757 321,1 550,1 924,1 092,1 599 noitciderPrednU .sbO 051 151 022 67 322 802 89 492 432 702 46 171 561 371 711 87 %02≥noitciderPrednU .sbO 05 46 96 22 17 86 82 79 67 96 02 35 35 85 63 42 %05≥noitciderPrednU .sbO 3 31 5 51 6 9 81 5 7 9 02 3 3 7 8 72 senacirruhforebmuN seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFyrtsudnI seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFemiT tnemhsilbatsE lanoitaN ,setaicossA & sllaW ;/sdrw/ude.nnepu.notrahw.bew-sdrw//:ptth ,secivreS ataD hcraeseR notrahW ,SU BD yvI ,scirteMnoitpO :ecruoS deepS dniW retneC enacirruH lanoitaN ;esabataD degreM tatsupmoC/PSRC ,secirP ytiruceS ni hcraeseR rof retneC ;4102 esabataD )STEN( seireS-emiT yrammuS snoitaralceD retsasiD )AMEF( noitartsinimdA tnemeganaM ycnegremE laredeF ;/txet/evihcra/vog.aaon.chn.www//:sptth ta elbaliava ,seitilibaborP detinUehtrofesabataDsessoLdnastnevEdrazaHlaitapS,ytisrevinUetatSanozirA;)81382/stnemucod/stessa/yrarbil-aidem/vog.amef.www//:sptthtaelbaliava( .)SUDLEHS( setatS 51

)deunitnoc( nosirapmoc stsacerof AAON dna tekraM :01 elbaT selasnodesaberusopxeenacirruh’smriF:BlenaP (cid:17) (cid:16) ∗ hT,iVI/Γ−hT,iVI gol,noitapissid/llafdnalerofebsyadΓotnoitpecnienacirruhmorfVIniegnahC :elbairavtnednepeD syaD5 syaD4 syaD3 syaD2 yaD1 Γ %01 %1 %01 %1 %02 %01 %1 %04 %02 %01 %1 %05 %04 %02 %01 %1 ≥tihenacirruhfoytilibaborP 020.0 700.0 ∗∗∗121.0 ∗∗410.0 ∗∗∗331.0 ∗∗780.0 ∗∗410.0 ∗∗∗461.0 ∗∗∗080.0 ∗∗∗070.0 ∗∗710.0 ∗∗∗051.0 ∗∗∗431.0 ∗∗∗470.0 ∗∗∗540.0 ∗∗∗730.0 Γ−hT,ierusopxEtsaceroF )156.0( )988.0( )421.3( )169.1( )745.4( )225.2( )164.2( )501.5( )462.5( )233.5( )770.2( )429.4( )959.3( )361.4( )455.4( )134.3( 100.0 220.0 030.0 300.0 010.0 210.0 140.0- 300.0- 500.0 100.0 740.0- 200.0 000.0 600.0 920.0 210.0 Γ−hT,inoitciderPrednU )920.0( )612.1( )303.1( )270.0( )524.0( )685.0( )052.1-( )031.0-( )281.0( )050.0( )199.0-( )580.0( )610.0( )282.0( )092.1( )224.0( 190.0 001.0 252.0 851.0 551.0 341.0 801.0 191.0 761.0 841.0 021.0 632.0 532.0 002.0 412.0 161.0 )%(2RdetsujdA 073,2 703,01 000,4 558,11 647,4 730,7 022,41 969,3 654,5 660,7 889,51 243,2 243,2 053,5 160,6 231,12 latot .sbO 366 925,4 001,1 000,6 572,1 939,1 916,6 819 403,1 558,1 771,6 146 007 624,1 607,1 066,5 %0>erusopxEtsaceroF .sbO 93 389 69 746,1 59 441 647,1 47 321 271 681,1 66 67 341 981 954 %02≥erusopxEtsaceroF .sbO 12 884 05 988 65 37 509 04 07 98 206 53 14 47 101 732 %05≥erusopxEtsaceroF .sbO 637 680,1 751,1 817 304,1 355,1 159 135,1 606,1 254,1 218 221,1 550,1 624,1 192,1 320,1 noitciderPrednU .sbO 631 591 122 79 332 122 731 113 462 432 79 602 302 502 851 311 %02≥noitciderPrednU .sbO 36 88 69 04 79 69 35 441 411 601 14 09 78 29 16 54 %05≥noitciderPrednU .sbO 3 31 5 51 6 9 81 5 7 9 02 3 3 7 8 72 senacirruhforebmuN seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFyrtsudnI seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY seY EFemiT 52

Table 11: Hurricane effects on implied volatility of insurance firms This table reports the coefficients and test statistics when estimating the panel model in equation (9) for insurance firms. The dependent variable is the change (in percent) in the implied volatility of firm i from the day before the inception day of the hurricane T∗ until 5 trading days after the landfall T . The independent variable measures the h h shareoftotalpremiumswrittenbyaninsurancefirminstatesthatexperienceddamagebyahurricane. ForPanelA, a state is considered to have experienced hurricane damage if at least 10% of the counties experienced damage, and for Panel B, the threshold is 25% of the counties. To identify counties that have been damaged by a hurricane we use FEMA data and FEMA data enhanced with SHELDUS data. The data are from 2002 to 2017. Results are also shown for the subsample from 2007 to 2017 as in Table 5. The values in parentheses are the t-stats. The standard errors are clustered by insurance firm. Time fixed effects are used. The significance of the coefficient estimate is indicated by * for p < 0.10, ** for p < 0.05, and *** for p < 0.01. Panel A: State considered hit if 10% or more of the counties were damaged (cid:16) (cid:17) Dependent variable: Change in IV, log IV i,Th+5 /IV i,T h ∗ 2002-2017 2007-2017 FEMA FEMA+SHELDUS FEMA FEMA+SHELDUS StateDamageExposure 0.348∗∗ 0.328∗ 0.357∗ 0.360∗∗ i,Th 1.985 1.963 1.850 1.995 Adjusted R2 (%) 36.894 34.790 33.399 33.597 Obs. total 400 418 238 238 Obs. insurance firm exposure > 0% 356 374 207 207 Obs. insurance firm exposure ≥ 20% 49 88 30 53 Obs. insurance firm exposure ≥ 50% 11 11 11 11 Hurricanes 18 19 9 9 Time FE Yes Yes Yes Yes Panel B: State considered hit if 25% or more of the counties were damaged (cid:16) (cid:17) Dependent variable: Change in IV, log IV i,Th+5 /IV i,T h ∗ 2002-2017 2007-2017 FEMA FEMA+SHELDUS FEMA FEMA+SHELDUS StateDamageExposure 0.426∗ 0.400∗∗ 0.424∗ 0.413∗∗ i,Th 1.922 2.329 1.733 2.251 Adjusted R2 (%) 38.482 36.595 37.328 38.254 Obs. total 367 385 205 205 Obs. insurance firm exposure > 0% 326 345 177 179 Obs. insurance firm exposure ≥ 20% 22 43 14 27 Obs. insurance firm exposure ≥ 50% 7 11 7 11 Hurricanes 17 18 8 8 Time FE Yes Yes Yes Yes Source: OptionMetrics, Ivy DB US, Wharton Research Data Services, http://wrds-web.wharton.upenn.edu/wrds/; S&PMarketIntelligence;CenterforResearchinSecurityPrices,CRSP/CompustatMergedDatabase;FederalEmergency Management Administration (FEMA) Disaster Declarations Summary (available at https://www.fema.gov/ media-library/assets/documents/28318); Arizona State University, Spatial Hazard Events and Losses Database for the United States (SHELDUS). 53

Appendix A Additional tables Table A.1: Hurricane effects on implied volatility with industry interactions Thistablereportsthecoefficientsandteststatisticswhenestimatingthepanelmodelinequation(9)withanindustry interaction term. The dependent variable is the percent change in implied volatility of firm i from the day before hurricane inception to 5 trading days after landfall. The first independent variable, DamageExposure , is the i,Th share of a firm’s geographic footprint that is exposed to the disaster. The second independent variable interacts the exposure to the disaster area with an indicator for industry g. The industry classification is based on two-digit SIC numbers. In Panels A and B, firm geographic exposure to a hurricane is based on establishments and sales per county,respectively. TheanalysisisbasedondamagedatafromFEMAandFEMAenhancedwithSHELDUSdata. The data are from 2002 to 2017. The values in parentheses are the t-stats. The standard errors are clustered by headquartercounty. Industryandtimefixedeffectsareused. Thesignificanceofthecoefficientestimateisindicated by * for p < 0.10, ** for p < 0.05, and *** for p < 0.01. PanelA:Firms’hurricaneexposurebasedonestablishments (cid:16) (cid:17) Dependentvariable: ChangeinIV,log IV i,Th+5 /IV i,T h ∗ DamagebasedonFEMA Interactionindustry Construct. Manufact. Mining Retail Services Transport. Wholesale DamageExposure 0.072∗∗∗ 0.086∗∗∗ 0.072∗∗∗ 0.073∗∗∗ 0.073∗∗∗ 0.060∗∗∗ 0.067∗∗∗ i,Th (4.359) (4.643) (3.403) (4.492) (4.078) (3.100) (3.840) DamageExposure ×I -0.125 -0.043 -0.002 -0.039 -0.009 0.068 0.078 i,Th i∈Industryg (-1.528) (-1.385) (-0.048) (-0.367) (-0.196) (1.294) (1.242) AdjustedR2 (%) 13.375 13.384 13.372 13.374 13.373 13.392 13.381 Obs. total 12,156 12,156 12,156 12,156 12,156 12,156 12,156 Obs. exposure>0% 5,387 5,387 5,387 5,387 5,387 5,387 5,387 Obs. exposure≥20% 805 805 805 805 805 805 805 Obs. exposure≥50% 286 286 286 286 286 286 286 Obs. i∈Industry 212 5,356 1,146 1,218 2,221 1,512 383 g Numberofhurricanes 19 19 19 19 19 19 19 DamagebasedonFEMA+SHELDUS Interactionindustry Construct. Manufact. Mining Retail Services Transport. Wholesale DamageExposure 0.072∗∗∗ 0.080∗∗∗ 0.071∗∗∗ 0.075∗∗∗ 0.074∗∗∗ 0.059∗∗∗ 0.067∗∗∗ i,Th (4.359) (4.226) (3.597) (4.484) (4.223) (3.022) (3.915) DamageExposure ×I -0.216∗∗∗ -0.027 0.001 -0.084 -0.018 0.075 0.086 i,Th i∈Industryg (-3.137) (-0.902) (0.032) (-0.953) (-0.383) (1.433) (1.417) AdjustedR2 (%) 13.576 13.572 13.567 13.577 13.568 13.591 13.578 Obs. total 12,621 12,621 12,621 12,621 12,621 12,621 12,621 Obs. exposure>0% 5,938 5,938 5,938 5,938 5,938 5,938 5,938 Obs. exposuregeq 20% 974 974 974 974 974 974 974 Obs. exposuregeq 50% 317 317 317 317 317 317 317 Obs. i∈Industry 221 5,562 1,174 1,273 2,317 1,569 394 g Numberofhurricanes 20 20 20 20 20 20 20 IndustryFE Yes Yes Yes Yes Yes Yes Yes TimeFE Yes Yes Yes Yes Yes Yes Yes Source: OptionMetrics, Ivy DB US, Wharton Research Data Services, http://wrds-web.wharton.upenn.edu/wrds/; Walls & Associates, National Establishment Time-Series (NETS) Database 2014; Center for Research in Security Prices, CRSP/Compustat Merged Database; Federal Emergency Management Administration (FEMA) Disaster Declarations Summary (available at https://www.fema.gov/media-library/assets/documents/28318); Arizona State University, Spatial Hazard Events and Losses Database for the United States (SHELDUS). 54

Table A.1: Hurricane effects on implied volatility with industry interactions (continued) PanelB:Firms’hurricaneexposurebasedonsales (cid:16) (cid:17) Dependentvariable: ChangeinIV,log IV i,Th+5 /IV i,T h ∗ DamagebasedonFEMA Interactionindustry Construct. Manufact. Mining Retail Services Transport. Wholesale DamageExposure 0.052∗∗∗ 0.063∗∗∗ 0.047∗∗∗ 0.053∗∗∗ 0.058∗∗∗ 0.041∗∗∗ 0.047∗∗∗ i,Th (3.621) (3.699) (2.874) (3.944) (3.633) (2.588) (3.156) DamageExposure ×I -0.129∗∗ -0.034 0.019 -0.037 -0.037 0.057 0.068 i,Th i∈Industryg (-2.028) (-1.428) (0.561) (-0.438) (-1.039) (1.560) (1.517) AdjustedR2 (%) 13.367 13.372 13.364 13.363 13.369 13.381 13.372 Obs. total 12,126 12,126 12,126 12,126 12,126 12,126 12,126 Obs. exposure>0% 5,365 5,365 5,365 5,365 5,365 5,365 5,365 Obs. exposure≥20% 808 808 808 808 808 808 808 Obs. exposure≥50% 391 391 391 391 391 391 391 Obs. i∈Industry 212 5,333 1,146 1,211 2,221 1,512 383 g Numberofhurricanes 19 19 19 19 19 19 19 DamagebasedonFEMA+SHELDUS Interactionindustry Construct. Manufact. Mining Retail Services Transport. Wholesale DamageExposure 0.051∗∗∗ 0.056∗∗∗ 0.047∗∗∗ 0.052∗∗∗ 0.058∗∗∗ 0.038∗∗ 0.047∗∗∗ i,Th (3.777) (3.378) (3.168) (4.053) (3.996) (2.500) (3.473) DamageExposure ×I -0.119∗∗ -0.019 0.016 -0.050 -0.049 0.070∗ 0.041 i,Th i∈Industryg (-2.327) (-0.808) (0.516) (-0.665) (-1.457) (1.823) (0.908) AdjustedR$ˆ2$(%) 0.136 0.136 0.136 0.136 0.136 0.136 0.136 Obs. total 12,590 12,590 12,590 12,590 12,590 12,590 12,590 Obs. exposure$¿$0% 5,916 5,916 5,916 5,916 5,916 5,916 5,916 Obs. exposure≥20% 953 953 953 953 953 953 953 Obs. exposuregeq 50% 447 447 447 447 447 447 447 Obs. i∈Industry 221 5,538 1,174 1,266 2,317 1,569 394 g Numberofhurricanes 20 20 20 20 20 20 20 IndustryFE Yes Yes Yes Yes Yes Yes Yes TimeFE Yes Yes Yes Yes Yes Yes Yes 55

Table A.2: Abnormal returns post hurricane landfall (alternative threshold) Thistablereportsdifferencesincumulativeabnormalreturnspostlandfallbetweenfirmswithexposure(treated)and firmswithoutexposure(control)tothehurricanedisasterregion. ComparedtoTable6,theexposurethresholdfora firm to be characterized as treated is reduced from 50% to 25%. The differences are reported for the mean and nine percentiles of the return distributions of the treated and control firms. The differences are estimated for two time periods: from landfall to 5 days and 120 days after landfall, respectively. The abnormal returns are estimated based ontheFama-Frenchthreefactormodel. FEMAandSHELDUSdataareusedtoidentifycountiesthathavebeenhit by a hurricane. For Panel A, the hit firms are defined as firms that have 25 percent or more of their establishments in the counties of the disaster region, and for Panel B, 25 percent or more of the sales have to be located in the disaster area counties. We exclude hurricanes that affected less than 25 firms. The data are from 2002 to 2017. The standard errors are bootstrapped and clustered by county (headquarter location). The significance of the difference in abnormal returns is indicated by * for p < 0.10, ** for p < 0.05, and *** for p < 0.01. Panel A: Hit firms selected based on establishments in disaster region ≥ 25% From 0 to 5 days post landfall From 0 to 120 days post landfall Cumulative r difference T-stat Cumulative r difference T-stat Mean 0.049 0.120 -4.997∗∗ -1.980 Percentiles 10% -0.476 -1.038 -12.693∗∗∗ -3.103 20% -0.657 -1.924 -7.216∗∗ -2.283 30% -0.251 -0.980 -6.168∗∗∗ -2.661 40% -0.268 -1.195 -4.392∗∗ -2.211 50% -0.224 -0.980 -4.672∗∗∗ -2.944 60% -0.276 -1.210 -4.427∗∗ -2.243 70% -0.079 -0.170 -3.727∗∗ -2.395 80% -0.066 -0.154 -3.553 -1.556 90% 1.223 1.264 -2.197 -0.483 Hit firms 1,131 1,096 Control firms 9,977 9,665 Panel B: Hit firms selected based on sales in disaster region ≥ 25% Mean 0.037 0.086 -4.113 -1.640 Percentiles 10% -0.085 -0.221 -8.691∗∗ -2.129 20% -0.452 -1.221 -4.930∗ -1.658 30% -0.201 -0.752 -4.622∗ -1.763 40% -0.193 -0.712 -3.738∗∗ -1.984 50% -0.237 -0.931 -3.501∗ -1.935 60% -0.132 -0.469 -3.366∗ -1.855 70% 0.020 0.046 -2.478∗ -1.653 80% -0.022 -0.051 -2.897 -1.353 90% 0.268 0.289 -4.740 -1.194 Hit firms 1,164 1,131 Control firms 9,898 9,583 Source: Walls & Associates, National Establishment Time-Series (NETS) Database 2014; Center for Research in SecurityPrices,CRSP/CompustatMergedDatabase;FederalEmergencyManagementAdministration(FEMA)Disaster Declarations Summary (available at https://www.fema.gov/media-library/assets/documents/28318); Arizona State University, Spatial Hazard Events and Losses Database for the United States (SHELDUS). 56