Reasons Behind Words: OPEC Narratives and the Oil Market

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Reasons Behind Words: OPEC Narratives and the Oil Market Celso Brunetti, Marc Jo¨ets, Val´erie Mignon 2024-003 Please cite this paper as: Brunetti, Celso, Marc Jo¨ets, and Val´erie Mignon (2024). “Reasons Behind Words: OPEC Narratives and the Oil Market,” Finance and Economics Discussion Series 2024-003. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2024.003. 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.

Reasons Behind Words: OPEC Narratives and the Oil Market Celso Brunetti∗ Marc Joëts† Valérie Mignon‡ January 10, 2024 Abstract We analyze the content of the Organization of the Petroleum Exporting Countries (OPEC) communications and whether it provides information to the crude oil market. To this end, we derive an empirical strategy which allows us to measure OPEC’s public signal and test whether market participants find it credible. Using Structural Topic Models, we analyze OPEC narratives and identify several topics related to fundamental factors, such as demand, supply, and speculative activity in the crude oil market. Importantly,wefindthatOPECcommunicationreducesoilpricevolatilityandpromptsmarket participants to rebalance their positions. Our analysis indicates that market participants assess OPEC communications as providing an important signal to the crude oil market. JEL Classification: G10, Q35, Q40, C45, C50 Keywords: OPEC Announcements, Structural Topic Models, Volatility, Traders’ Positions Acknowledgements: We would like to thank Christiane Baumeister, Anna Creti, Thomas Grjebine, Lutz Kilian, Julia Silbert, and participants at the CEMA annual meeting, and the 16th CFE conferences for very constructive comments and suggestions. ∗Division of Research and Statistics, Board of Governors of the Federal Reserve System, United States of America. E-mail: celso.brunetti@frb.gov †IESEGSchoolofManagement;Univ. Lille,CNRSUMR9221-LEMLilleEconomieManagementF-59000 Lille, France. E-mail: m.joets@ieseg.fr ‡EconomiX-CNRS, University of Paris Nanterre and CEPII, Paris, France. E-mail: valerie.mignon@parisnanterre.fr 1

1 Introduction Communication is essential in policy institutions, governmental and intergovermental organizations as well as firms. As the leading entity in oil markets, the Organization of the Petroleum Exporting Countries (OPEC) regularly shares information by releasing communications. OPEC’s objective is “[...] to coordinate and unify the petroleum policies of its Member Countries and ensure the stabilization of oil markets [...]”.1 A well-functioning crude oil market may have positive implications for the economy and inflation.2 Falling into this context, this paper aims to extract the information content in OPEC communication using textual analysis. In particular, we are interested in identifying the topics of OPEC’s announcements, the fundamental factors driving these topics, and how these topics are connected to form OPEC’s overall narrative. We are then interested in understanding whether OPEC’s narrative percolates into the crude oil market. Traditional finance theory suggests that information flows into markets through volatility and volume (Epps & Epps (1976), and Gallant et al. (1992)). Hence, our objective in the second part of the paper is to understand whether OPEC communication is credible and, if so, how it relates to both oil price volatility and the trading behavior of commercial and non-commercial market participants. We begin by analyzing all OPEC communications, then use Structural Topic Models (STM)3 to extract the information content (signal) in OPEC narratives. In line with Morris & Shin (2002), weassumethatforOPEC’spublicsignaltobecredible, itshouldreflectmarketfundamentals. Therefore,toestimatethestructuraldynamicsofOPECcommunication,weconsider several exogenous factors, such as oil demand, oil supply, and speculative activity in the oil market.4 We are able to identify numerous topics embedded in OPEC communication, which characterize the Organization’s public signal. Besides the obvious topics, such as “prices,” “oil shortage,” and “economic growth,” we detect topics related to “climate change” and “energy policies.” This is not surprising, since climate change and climate-related risks have both direct (i.e., new policies to reduce fossil fuels emissions) and indirect (i.e., new technologies) effects on oil-producing countries. The richness of our textual analysis results allows us to study the rationale behind OPEC communication. We do so in two ways. First, we map the network of topics in OPEC communication in order to investigate the interconnectedness of OPEC announcements and illustrate the complexity of OPEC’s narrative. Second, we identify factors that impact OPEC topics. The results of this analysis allow us to better understand the structure of OPEC communi- 1See https://www.opec.org 2OPEC’sproductionagreementsarenotbinding,andthereisnoenforcementmechanism. Infact,compliance of OPEC’s countries to production agreements has fluctuated over time. 3See Roberts et al. (2013). 4Asitiswell-establishedintheliterature,thesefactorsareexogenouswithrespecttoOPECcommunications and crude oil price—see, e.g., Kilian (2023) and the references therein. 2

cation, which is a first step in our quest to test whether OPEC communication contains a credible signal. Our textual analysis shows that OPEC topics are linked to important factors and are clustered in a meaningful manner, suggesting that OPEC narratives are based on crude oil fundamentals. To test how OPEC narratives impact the oil market, we adopt Lasso penalized regressions. BasedontheMorris&Shin(2002)andAmatoetal.(2002)theoreticalframework, wedevelop an empirical strategy to crucially test the hypotheses that OPEC’s public signal (i) matters to the oil market and (ii) changes as a function of the precision of the private signal. We are aware that endogeneity concerns may affect our analysis. However, our identification strategy is motivated by two observations. First, OPEC is the dominant player in the crude oil market, and its announcements are closely monitored by market participants. Second, our analysis concentrates only on the days and weeks when there are announcements. Therefore, and similar to Känzig (2021), we isolate the impact of OPEC communication and consider only the window around OPEC’s announcements. OurresultsunequivocallyshowthatOPECsignalishighlyrelatedtocrudeoilmarketvolatility and traders’ positions. We find that OPEC communication is associated with lower volatility levels.5 These results are stronger for longer-dated futures contracts, indicating that OPEC narratives are linked to the entire futures curve. OPEC communication that reassures the market on production capacity and supply contributes to oil market stabilization. Turning to traders’ positions, our findings indicate that market participants’ trading activity is deeply linked to OPEC topics.6 In particular, different topics are associated with different traders. Traders engaged in the physical business (i.e., producers, merchants, processors and users) change their positions with OPEC topics related to economic growth, while non-commercial traders (i.e., traders with no business in the underlying physical market), change their positions with topics related to oil supply and “energy policy.” All market participants rebalance their positions with the topic “cooperation.” Since OPEC does not have an enforcing mechanism, cooperation among members is essential for a credible signal.7 We also find substantial evidence that OPEC’s public signal is stronger when the private signal is noisier, in line with the predictions in Morris & Shin (2002). We runseveral robustness checks whichconfirm ourresults. Importantly, wealsoimplement a placebo test, consisting of selecting days and weeks randomly with no OPEC announcements and constructing different control groups. We show that, in the absence of OPEC announcements, there is no effect on volatility and traders’ positions, providing further support to our 5We measure volatility with the daily range, i.e., the difference between the daily log-high and log-low prices. 6WeusedatafromtheDisaggregatedCommitmentofTraders(DCOT)reportsfromtheCommodityFutures Trading Commission (CFTC) to measure trading activity of crude oil market participants. 7The absence of an enforcing mechanism could be particularly important during crisis periods. In the last part of the paper, we analyze the Global Financial Crisis (GFC) and the COVID-19 pandemic. Our results are robust to the two crises. 3

identification strategy and confirming the existence of a causal relationship. The paper provides several contributions. First, we are the first to apply Structural Topic Models to analyze OPEC announcements. This technique allows us to identify relevant topics and study how those topics are connected in a network. Furthermore, we analyze the drivers (supply, demand, and speculative factors) of the estimated topics. Pescatori & Nazer (2022) study OPEC communication and use cosine similarity and term-frequency-inverse document frequency techniques to assess the distance, in terms of content, between different OPEC announcements. Repetitive communications are considered to be uninformative. They find that OPEC statements do not vary much except when oil prices fluctuate dramatically, such as in 2008 when they reached about $150 per barrel. Our approach is different, since we are able to precisely estimate the topics of OPEC narratives and, thus, the information content of OPEC announcements. Second, to the best of our knowledge, we are the first to apply the Morris & Shin (2002) and Amato et al. (2002) theoretical framework to OPEC communication and to the oil market. A large literature applies this approach to central bank communication,8 but not to OPEC announcements. Starting from the seminal work of Morris & Shin (2002), we derive testable hypotheses and build an empirical strategy to test them. Third, we test whether topics in OPEC communication are linked to volatility and traders’ positions. We are particularly interested in these two variables because there is substantial evidenceshowinghowtradingvolumeandvolatilityarerelatedthroughtheinformationflow.9 We consider crude oil price volatility over the entire maturity futures curve and disaggregated traders’positionsfromtheCFTCpublicdata. Thereisalargeliteraturestudyingthecredibility of OPEC communication with mixed results. Wirl & Kujundzic (2004) find weak evidence of the impact of OPEC communication on the world oil market, while Guidi et al. (2006) show that the effectiveness of OPEC decisions varies over time. Similarly, Demirer & Kutan (2010), using an event study approach, find that only OPEC production cut announcements have an impact on oil prices, and this impact vanishes for longer maturity futures contracts. In a similar camp are the results of Fattouh & Mahadeva (2013), which show that OPEC pricing power varies over time. Brunetti et al. (2013) provide empirical evidence that OPEC fair price pronouncements have limited effects on the actual price of crude oil. Looking at more recent data, Quint & Venditti (2020) also find a limited effect of OPEC+ on crude oil markets. Some studies, however, provide substantive evidence that OPEC announcements have a significant influence on crude oil markets. Lin & Tamvakis (2010) and Loutia et al. (2016), using an event study approach, find significant crude oil market responses to OPEC production decisions.10 8See, among others, Hermann & Fratzscher (2007a), Hermann & Fratzscher (2007b), and Evans et al. (2012). 9See, Epps & Epps (1976), Tauchen & Pitts (1983), and Gallant et al. (1992). 10Loutia et al. (2016) account for the volatility structure of crude oil prices using an Exponential GARCH model. 4

Our approach is different, as we do not use an event study methodology. Rather, we precisely measure the topics in OPEC communication. Our results suggest that OPEC communication is based on fundamental factors and generates a credible public signal. In particular, we find that OPEC topics reduce volatility levels, in line with OPEC’s mandate of market stabilization, and induce market participants to rebalance their positions. This is particularly true when the private signal is noisy, in accordance with the predictions in Morris & Shin (2002). The remainder of the paper is organized as follows. Section 2 is devoted to the information content of OPEC narratives. Section 3 presents the theoretical framework and the related empirical strategy to test the effectiveness of OPEC communication. Section 4 studies the linkagesbetweenOPECtopicsandvolatilityandtraders’positions. Section5providesseveral robustness checks, including a sensitivity analysis to recent crisis periods. Section 6 concludes the paper. 2 The information content of OPEC communication This section begins by introducing the database of OPEC press releases, as well as identifying thefundamentaldriversinfluencingOPECcommunications. Next,weoutlinethemethodology employed to quantify the information content in OPEC statements through textual analysis. Finally,weexaminethetemporalevolutionofOPECnarratives,exploretheirinterconnections, and identify the driving forces behind each topic. 2.1 OPEC press releases and fundamental drivers The information content of OPEC communication is estimated using textual analysis, as described in Section 2.2. Our base corpus consists of OPEC press releases extracted from OPEC’s website, starting from March 2002 and updated each time a representative member gives an official talk to the press.11 Declarations from the Organization to the press usually come around major events, such as ordinary and extraordinary OPEC conferences, but also during episodes of oil market turbulence. In this respect, our corpus is quite imbalanced, with years of intense communications (more than one per month) followed by periods of limited communications. Considering every press release from March 2002 to March 2021, the sample includes 343 announcements.12 Figure 1 reports in grey the number of announcements per year over the sample period and shows how OPEC communication is cyclical. As expected, periods of intense communication coincide with episodes of prolonged low oil prices, such as the development of the US shale oil in 2016-2017 and the COVID-19 pandemic in 2020. From the total speeches, we drop talks with no topical content (such as data exercises or administrative 11See www.opec.org/opec_web/press_room/28.htm. 12The detailed database including dates, speakers, locations, and titles of each press release, is available upon request from the authors. 5

issues), leaving 262 press releases over our sample period (in red in Figure 1). Figure 1: OPEC press releases over time (2002-2021) Note: ThisfiguredepictsthenumberofannualOPECpressreleasesfromMarch2002toMarch2021(ingrey), totaling 343 speeches (source: OPEC’s website), with our selected sample of 262 speeches highlighted in red. As highlighted in the existing literature, OPEC continuously assesses the oil market to establish targeted prices and appropriate supply levels (Kilian & Murphy (2014), Baumeister & Kilian (2016)). Consequently, we posit that OPEC’s signaling is endogenous to market conditions, with the Organization observing fundamental factors in the crude oil market prior to issuing a statement. In order to estimate the endogenous structural dynamics influencing OPEC’s communications, we incorporate several exogenous variables that could impact the Organization’s messaging to the market. These variables enable us to consider multiple market components—including demand, supply, and speculative factors—and serve to identify the topics contained in OPEC’s public statements.13 These variables were constructed for the period from March 2002 to March 2021 to align with the frequency of OPEC press releases and are elaborated upon below. While some of the considered exogenous factors are quantitative in nature (i.e., evolve over time), we transform all variables to be qualitative, categorizing them into groups or classes. The transformation is performed for two reasons: (i) analytically, we are interested in OPEC signal with respect to some reference point, for instance when the US business conditions improve (or worsen), and (ii) technically, computation and interpretation are made easier when working with qualitative variables. 13See,e.g.,Kilian&Murphy(2014),Baumeister&Kilian(2016),andBrunettietal.(2016)foradiscussion. 6

We consider a total of eight exogenous qualitative variables organized into three categories: supply, demand, and speculative. Regarding supply, we rely on spare crude oil capacity. We construct a two-class variable, which we label “low” when capacity is low or close to zero (for instance between 2004-2008 and between 2012-2018) and “high” otherwise. We build five demand-related variables: • US business conditions: we use the ADS index developed by Aruoba et al. (2009) to proxy US business conditions. By construction, the average value of the index is zero, and progressively larger positive (negative) values indicate progressively better-thanaverage (worse-than-average) conditions. We construct a two-class variable capturing “worse” conditions if negative, and “better” if positive. • Economic uncertainty for the US and Europe: we use the uncertainty index of the state of the economy developed by Scotti (2016). We construct a two-class variable which we label “high” when the index is higher than 1.65 standard deviation, and “low” otherwise. • Economic surprise for the US and Europe: we use the surprise index developed by Scotti(2016)summarizingrecenteconomicdatasurprisesandmeasuresofoptimismand pessimism about the state of the economy. We construct a two-class variable recording “optimism” when negative, and “pessimism” when positive. Finally, we construct two variables that capture speculative activity, namely US & OECD crude oil inventories. We use US crude oil stocks from Energy Information Administration and OECD crude oil stocks from the International Energy Agency. We then construct a twoclass variable for each inventory considering “high” when stocks are outside the 5-year average band (for instance during the shale oil period of 2015-2017), and “normal” otherwise. The use of inventories to capture speculative activity has been considered in several studies, such as Kilian & Murphy (2014). 2.2 Structural Topic Model For topics estimation, we rely on unsupervised probabilistic topic models applied to OPEC press releases. Among these models, mixed-membership approaches allowing each document to be composed of multiple topics, have become a common tool for mining large corpora in various fields.14 The intuition, popularized by Blei et al. (2003)’s Latent Dirichlet Allocation (LDA), is that a document is a collection of multiple topics, which are themselves a collection of words. A topic is then defined as a mixture of words where each word has a probability of belonging to a topic. A document is a mixture of topics, and so a single document can be composed of multiple topics. The main goal of these models is to estimate the following three components: • Topic proportions (i.e., document-topic probability distributions) for each document d ∈ {1,2...,D} (also called topic prevalence) described by the parameter θ . d 14See Blei & Lafferty (2009) and Blei (2012) for a review. 7

• Wordproportions(i.e.,topic-wordprobabilitydistributions)foreachtopick ∈ {1,2...,K} (also called topic content) described by the parameter β . k • Core language combining the two previous components to produce the actual words in eachdocument. Inotherterms, foreachwordn ∈ {1,2...,N}indocumentd, aper-word topic assignment z conditional on the document-topic probability distribution is drawn from a multinomial distribution (z |θ ). Given the topic, words are randomly chosen d,n d from a multinomial distribution (w |z ,β ). d,n d,n k In the LDA-type framework, both topic and word proportions (θ and β ) are randomly chod k sen from a Dirichlet distribution with priors (the hyperparameters α and η). While such a standard topic model has proven to be quite efficient in discovering latent topics in economics and finance,15 it has some limitations. First, the Dirichlet distribution does not allow topics within documents to be correlated and vary over time.16 Second, the model does not permit topic prevalence and topic content to be influenced by exogenous factors or covariates. In other words, it does not allow us to discover topics and estimate their relationships with factors that may affect their dynamics. To overcome these limitations, we estimate OPEC topicsandanalyzetheirrelationshipswithcovariatesusingtheStructuralTopicModel(STM) developed by Roberts et al. (2013). Both LDA and STM share the same spirit by estimating the quantities described previously. However, in STM, the estimation of the parameters depends on exogenous factors, X and Y (X and Y can be the same set of covariates). Technically, topic prevalence θ is assumed to d be a random variable drawn from a Logistic-Normal distribution conditional on covariates, as: θ | X ,Σ ∼ Logistic−Normal(µ = X ,Σ) d dγ dγ where X is a vector of covariates, γ ∼ N(0,σ2) is a matrix of coefficients with σ2 ∼ d k k Gamma(sγ,rγ), and Σ is the covariance matrix. The topic content β is replaced with a multinomial logit such that a word’s distribution is k the combination of three effects (topic κk, covariates κy, and topic-covariate interaction κy,k) over v individual words in the relevant vocabulary of possible words, such as: β exp(m+κk +κy +κy,k) d,k v v v (cid:29) 15See Hansen & McMahon (2016), Hansen et al. (2018), Larsen & Thorsrud (2019). 16See Blei & Lafferty (2006) and Blei & Lafferty (2007) for some extensions. 8

where m is the baseline word frequency, and (κk+κy+κy,k) is a collection of coefficients with v v v κy,k ∼ Laplace(0,ry,k) and ry,k ∼ Gamma(sκ,rκ). v v v Thisframeworkallowsustoevaluatehowoursupply, demand, andspeculativevariablesaffect OPEC communication defined as topics and word proportions. Measuring OPEC communication with mixed-membership topic models is difficult because of the latent structure of the parameters, as well as the intractable and non-convex posterior. Two approximate inference algorithms are popular for the estimation: Gibbs sampling (Griffiths & Steyvers (2004)) and variational inference (Blei et al. (2003)). As suggested by Roberts et al. (2016), we estimate the model using a semi-collapsed variational EM algorithm. We further induce sparsity on the collection of parameters by regularizing prior distributions for κ (with Laplace prior), and γ (with L1-penalty) to improve interpretability, prevent overfitting, and increase computational efficiency.17 A more challenging exercise in estimating topic models is the dimensionality of the latent space, namely the number of topics K. The procedure always involves a trade-off between statistical goodness-of-fit (i.e., higher K) and output interpretability (i.e., lower K).18 We use different values of K, ranging from K = 20 to 60, and compute several statistical criteria (see Appendix A for technical details). We select K = 40 based on both statistical power and interpretability. 2.3 OPEC narratives and endogenous factors 2.3.1 Selecting topics from OPEC communication As is common in text-mining, our OPEC press releases database is high-dimensional and sparse (a 262×12586 document-term matrix with 90% of scarcity). Consequently, we need to reduce the dimensionality of the corpus before estimation. In other words, we have to remove wordscontaininglittletopicalcontent(seeAppendixAfortechnicaldetails). Inanutshell,the process resides in removing stopwords19 (i.e., ’the’, ’are’, ’but’, ...), given names, surnames, numbers and punctuation, as well as converting remaining terms into their linguistic roots (i.e., stemming). Once the dimensionality problem is reduced, STM can be estimated on the new document-term matrix. We estimate the 40-topics STM on our OPEC press release corpus from March 2002 to March 2021. ThetwomainoutputsaretopicsandwordproportionscoveringdifferentfacetsofOPEC communication. The model does not give any label, but provides the probability of each word within topics. While the label in itself plays no role in the analysis, it provides a convenient way to discuss OPEC communication. We propose to label topics based on both the top 10 17Additional results are available upon request from the authors. 18See Chang et al. (2009) for a discussion. 19The stopword list we used is from http://snowball.tartarus.org/algorithms/english/stop.txt, and is available upon request from the authors. 9

FREX (FRequency and EXclusivity) terms and most-probable bigrams (topic labeling and technical details are available in Appendix B). We select some topics from the estimation to highlight different aspects of OPEC communication. They are represented as clouds of keywords in Figure 2 together with their labels. As shown, OPEC communications are very diverse, with topics related to crude oil prices (Topic 2, panel a) and shortages (Topic 3, panel b), production adjustments during turbulent times (Topic 20, panel e), economic growth (Topic 11, panel c), climate change (Topic 19, panel d), and energy policy (Topic 36, panel f).20 OPEC narratives over the whole period also reveal that some topics are related with others. In other words, specific topics in OPEC communication tend to co-occur during particular circumstances. Figure 3 provides a static picture of these correlations over the period as a communities network map using an infomap algorithm (see Rosvall & Bergstrom (2007) for more details).21 Several observations emerge from the map. First, OPEC has a large and well developed spectrum of communication with connected narratives. Second, the weight of each spectrum is not homogeneous with different levels of topics’ importance: Topic 34 (OPEC/Non-OPEC Production Participation), Topic 25 (OPEC Production Adjustment), Topic 13 (Spare Oil Production Capacity), and Topic 4 (Oil uncertainty/Volatility) are examples of extensive topics. Third, while OPEC communication is very diverse, the overall narrative structure can be grouped into eight main communities in which topics are densely connected (each community is identified by a different color). These communities provide a better idea on the types of signals OPEC sends to markets. For instance, considering topics’ labels, we can identify that theorangeclusterisabout“supply/productionadjustment” signalduringtimesofuncertainty, when cooperation within OPEC producers and between OPEC and non-OPEC countries is needed. While both dark blue and yellow groups seem to be related to long-term production, the light blue cluster appears to be linked to supply shortages linked to natural disasters. Similarly, the green cluster is primarily related to price volatility and crude oil market stability,andbothpurpleandbrowncommunitiesareassociatedtoOPECinternationalrelations. Topics’ evolution over time for each community is reported in Figures 13 to 18 in Appendix C. These figures provide insights on the link between market conditions and OPEC signals, and alsohelpusunderstandthenatureofthesignalsperse. Forinstance,Figure13clearlyconfirms the “supply adjustment” signal during periods of high uncertainty, such as the Afghanistan Warin2002, theUSshaleoildevelopmentin2015-2018, andtheCOVID-19pandemic. Figure 15 depicts “price volatility and market stability” signals which peak during periods of strong price fluctuations, for instance, during the GFC. 20For the full list of topics see Tables 6, 7, and 8 in Appendix B. 21As a robustness check (not reported), we also compute communities using walktrap (Pons & Latapy (2006)),louvain(Blondeletal.(2008)),andpropagatinglabels(Raghavanetal.(2007))algorithms,forwhich we get very similar clusters in terms of information variation (Meilă (2003)). 10

Figure 2: Selected topics from OPEC communication Note: ThesefiguresreportestimatedtopicdistributionsfromSTMaswordclouds. Thesizeofwordsinclouds corresponds to the probability of occurrence in the corresponding topic. The larger the word, the higher the probability to occur. Note that we report the stemmed tokens. The label is from the methodology discussed in Appendix B. 11

Figure 3: Communities topics network in OPEC communication Note: Thisfigurerepresentstopiccorrelationoverthewholeperiodasanetworkstructure. Nodesizeindicates topic’s weight in the whole system (i.e., the bigger the node, the more important is the topic with respect to others). Edgesize(thickness)indicatesthestrengthoftheconnectionbetweentwotopics(i.e.,thethickerthe edge, the stronger the connection). Colors characterize nodes’ communities based on an infomap algorithm. For simplicity, isolated nodes without huge contributions to the system have been removed from the network. 2.3.2 Drivers of OPEC communication We are interested in identifying the variables that most impact OPEC topics. We extract from our STM model the estimated coefficients (γ,κk,κy,κy,k) and run linear regressions conv v v sidering (i) each topic as endogenous, and (ii) supply, demand, and speculative side factors as exogenous. Measurement uncertainty, potential serial correlation, and heteroskedasticity problems are treated “locally” by stepping through each document, updating the parameters, then saving the local covariance matrix. 12

As shown in Figure 4, eight exogenous factors influence OPEC communication. Regarding the supply-side factors (in green), the impact of spare crude oil capacity on OPEC communication is particularly significant, especially when capacity is high. This result is expected since OPEC defines its production levels as a function of its reserves and demand conditions. Communicating on factors related to supply (Topics 21 and 37 in blue) when reserves are high is a way for OPEC to reassure markets about oil reserves in the event of strong demand, as well as about possible shortages (Topic 3 in light blue). In the event of high reserves, OPEC also communicates to stabilize the market (Topic 14 in dark green), for instance through productioncutstolimitapotentialpricedecrease. Whenreservesarelow,communicationmainly concerns cooperation between OPEC members as well as OPEC and non-OPEC countries. Turning to demand-side factors (in dark gray) high economic uncertainty in the US and, to a lesser extent, in Europe is highly significant in explaining OPEC communication. When economic conditions are highly uncertain, OPEC tends to intervene on topics related to production adjustments (Topics 20 and 25 in orange) and unscheduled extraordinary meetings (Topic 1 in dark green). These meetings are typically conveyed when unanticipated events occur and are usually associated with falling prices. OPEC communication aims at reducing uncertainty and reaffirming its market power to stop prices from falling further. As an illustration, the COVID-19 pandemic pushed OPEC to communicate its commitment to limit the negative consequences of the crisis on the oil market caused by the slowdown of economic activity due to the lockdown. Similarly,asurpriseeffectregardingforecasters’predictionsontheUSandEuropeaneconomies, as well as the observed US economic conditions (in dark gray), affect OPEC communication. In the case of pessimistic surprises, OPEC mainly intervenes on demand-related concerns (Topic 12 in black), while it deals with topics related to its strategy (Topics 1 in dark green, and 5 in orange) when surprises are optimistic. In case of bad US economic conditions, OPEC communicates on supply-side factors to limit both volatility in the oil market and the subsequent fall in crude oil prices. Speculative-related factors (in brown) affect several topics. OECD crude oil inventories significantly impact OPEC communication, mainly when stocks are high—i.e., outside the 5-year average band. In this case, OPEC communication focuses on several topics linked to its strategy (Topics 5 in orange, and 8 in purple), the oil market (Topics 14 in dark green, and 33 in purple), and the world economic situation (Topic 35 in black). OECD inventories have been relatively stable for many years, but an upward trend pushes OPEC communication to reaffirm its market power. A similar pattern is observed when US crude oil inventories are high. This is to be expected, since OPEC has a strong incentive to intervene when prices are relatively low. In the case of low/normal inventories (in brown), OPEC communication mainly focuses on 13

Figure 4: Causes of OPEC communication Note: This figure reports the effects of exogenous factors (left panel) on topics’ distributions (right panel). Factors colors indicate supply (in green), demand (in gray), and speculative (in brown ) covariates. Topics colors indicate communities as in Figure 3. Only statistically significant relationships (at the 5% level) are reported. 14

cooperation and agreements (Topics 1 in dark green, 9 and 34 in orange). The aim is to provide a reassuring speech, limiting market uncertainty and volatility (Topics 4 in dark green, and 39 in yellow) that could be caused by fears of insufficient stocks or even shortages. The underlying idea is that the credibility of OPEC signals’ is higher when its members behave cooperatively. Indeed, when its members act in a non-cooperative way, geopolitical tensions are reignited, creating uncertainty and undermining OPEC’s credibility. 3 Signalling game: causes and consequences This section describes our methodology for testing the effectiveness of OPEC communication on both price volatility and trading positions. First, we briefly present the theoretical framework borrowed from Morris & Shin (2002) and Amato et al. (2002), taking the perspective of OPEC market power. Second, we delve into our empirical strategy and elaborate on the variables we have selected for analysis. 3.1 Theoretical framework To illustrate the influence of OPEC communication on market volatility and trading behaviors, we introduce a simplified signal coordination game model, drawing inspiration from the central bank communication literature.22 Like central banks, OPEC faces a dual challenge in communication: balancing market expectations with behavior coordination—reminiscent of Keynes’ “beauty contest” analogy. OPEC grapples with trade-offs between transparency and signalcredibility, especiallyininfluencingexpectationsandpricingpowerincrudeoilmarkets. The evolution of OPEC’s role has been influenced by a shift towards futures markets, which encompass a heterogeneous pool of actors like producers, swap dealers, refiners, and money managers (Fattouh (2007)). Consequently, OPEC employs a multifaceted strategy combining production decisions and public communications to steer market expectations (Fattouh & Mahadeva (2013)). Despite these efforts, the efficacy of such tools remains under debate.23 Numerous studies contend that OPEC’s signals often lack credibility because they are costfree and thus easily dismissed (Farrell & Rabin (1996), Fattouh (2007)). This skepticism leads markets to withhold judgment until OPEC’s announced decisions are implemented. Contrary to this view, our paper argues that, when considering the full spectrum of OPEC communications, their public signals matter. 22For seminal works on this topic, see Morris & Shin (2002, 2005), Morris et al. (2006) as well as related discussions in Svensson (2006) and Ehrmann & Fratzscher (2007). 23For comprehensive discussions on production decisions and their effectiveness, see Fattouh (2007). For insights into the impact of OPEC communication on oil prices, consult Wirl & Kujundzic (2004), Demirer & Kutan (2010), and Brunetti et al. (2013). 15

Paralleltocentralbanks, OPECservesadualroleas(i)anobserver, gatheringcluestoinform future actions, and (ii) a market influencer, shaping expectations. InlinewithMorris&Shin(2002)andEhrmann&Fratzscher(2007), weconsideracontinuum of private agents i ϵ[0,1] who make decision p guided by a utility function:24 i (cid:104) (cid:105) U (p ,θ) = − (1−r)(p −θ)2+r (cid:0) L −L¯(cid:1) (1) i i i i where θ denotes market fundamentals (i.e., in our case, the determinants of the crude oil market), r∈[0,1] is a constant, L = (cid:82)1 (p −p )2dj, and L¯ = (cid:82)1 L dj. The first term on the i 0 j i 0 j right-hand side of this equation emphasizes that agents make decisions aligned with fundamentals, while the second term captures a coordination game between agents’ actions, akin to Keynes’s beauty contest. The optimal decision rule of agent i is: p = (1−r)E (θ)+rE (p¯) (2) i i i where p¯ is the average decision across agents. The social planner aims to maximize welfare, which in our model is given by W (p,θ) = − (cid:82)1 (p −θ)2di, focusing solely on the component 0 i related to market fundamentals. Agents receive two types of signals about θ: an individual private signal specific to agent i, x = θ+ε and a public signal y = θ+η. Upon receiving these signals, each agent’s optimal i i decision rule simplifies to25 αy+(1−r)βx p = i . (3) i α+(1−r)β leading to an expected social welfare function α+β(1−r)2 V (α) = − . (4) [α+β(1−r]2 Thecruxofourinvestigationliesinassessingwhetheramoreprecisepublicsignal(orincreased transparency) positively or negatively affects welfare. According to Morris & Shin (2002), social welfare decreases with transparency if α < (2r−1)(1−r). (5) β This counter-intuitive result serves as a cautionary note for authorities about the amount and type of information they disclose. Specifically in the OPEC context, this leads to an analysis of how OPEC could strategically decide on the depth of forward-looking communication when settingnewproductionlevels. Thepivotalissueistodeterminethethresholdlevelofprecision, denoted by α¯, that makes the release beneficial to societal welfare α¯ = β(2r−1). (6) 24We borrow the notations from Ehrmann & Fratzscher (2007). 25The expected value of the fundamentals is E (θ)= αy+βxi. i α+β 16

This suggests that public signals with a precision level of α are more likely to be advantageous when the private signals are of lower quality—specifically, when α > β. The crux is whether the precision of the public information is sufficient to warrant its disclosure. Based on this theoretical groundwork and Equation (6), we set the stage for two important questions: 1. Does OPEC signal matter on average? 2. How does OPEC signal change with respect to the quality of the private signal? These questions lead to empirically testable hypotheses, which we explore in the next subsection. Inparticular, weintroduceanempiricalframeworkspecificallydesignedtoscrutinizethe effectiveness of OPEC’ signalling – as gauged by OPEC communications – in shaping market expectations. Our empirical framework also addresses the dilemma OPEC faces over the level of information it should disclose about its market perceptions. 3.2 Empirical strategy As discussed in the previous subsection and in Ehrmann & Fratzscher (2007), the Morris-Shin model postulates that both the public and private signals have an impact on social welfare. To measure the effectiveness of the OPEC signal on the crude oil market, we consider a two-step procedure that consists of (i) estimating the OPEC signal using the unsupervised learning model described in Section 2.2, and (ii) measuring and testing the effect of signals on the crude oil market using penalized regression. Let us present the econometric framework. To test the effectiveness of OPEC signal on the crude oil market, we rely on two empirical models. The first specification is given by: y = κ+λz +µz +ϑX +υ (7) t α,t β,t t t where y stands for the endogenous variable, κ is the constant term, X denotes a set of cont t trol variables, and λ and µ capture the effect of public and private signals on the crude oil market, respectively. Statisticallysignificantestimatesofλandµimplythatbothprivateand public signals convey important information. The public signal, however, should have a larger impact since it is related to underlying fundamentals as in Morris et al. (2006) and speaks to the effectiveness of OPEC communication. 17

WeareawarethatOPECcommunicationsareendogenoustomarketfundamentals.26 Inother words, based on market fundamentals, OPEC decides to communicate the level of intensity of the public signal to send. OPEC can then act either as a catalyst or a buffer of market fundamentals to influence prices and expectations. The intuition of Equation (7) is illustrated in Figure 5. The circle frame indicates the machine learning part of our approach devoted to estimating OPEC’s public signal (see Section 2.2). The dashed frame is the econometric part of the model, which we use to test the implications of the Morris et al. (2006)’s model. Figure 5: Empirical model 1 Supply Demand Speculative Public signal Private signal Crude oil Note: ThisfigurereportsourempiricalstrategyforEquation(7). Theshadedvariableisestimatedoutofthe econometric model. Unshaded variables are observed. The circle frame is the machine learning part devoted to the estimation of the OPEC signal as described in Section 2.2. The dashed frame is the econometric part Supply Demand Speculative of the model. For simplicity, this figure omits control variables. Equation (7) and Figure 5 imply that market participants directly observe both private and public signals (dashed frame). Yet, OPEC’s public signal reflects the Organization’s perception of crude oil market conditions, such as supply, demand, and speculative components High noise (circle frame).27 Public signal Private signal The second empirical model investigates the public Lsoiwg nnoisae l effect on the crude oil market depending on the precision of the private signal. In other words, we evaluate whether private 26Aswewilldescribefurther,todealwithsuchendogeneityissuesandtopreciselyidentifytheeffectofthe public signal, our experiment design is strictly limited to the days (weeks) when OPEC communicates. By considering the specific window of the announcements, we arCeraubdlee otiol isolate the impact of OPEC narratives on the market variables we are analyzing. 27See Kilian & Murphy (2014), Baumeister & Kilian (2016), and Brunetti et al. (2016) for a discussion on oil market determinants. 18

signalnoiseaffectstheinformationflowofthepublicone(asdiscussedbyMorrisetal.(2006)). To this end, webuild adummy variableDβ = 1ifthe noisiness ofprivateinformationisabove its mean over the whole sample period (high noise) and 0 otherwise (low noise). We then estimate the following equation which includes an interaction term between the public signal and Supply Demand Speculative noisiness in private signal (cid:16) (cid:17) (cid:16) (cid:16) (cid:17)(cid:17) y = κ+λ z Dβ +λ z 1−Dβ +µz +ϑX +υ (8) t 1 α,t t 2 α,t t β,t t t Public signal Private signal where λ 1 (λ 2 ) denotes the effect of public signal in periods of high noise (low noise) in private information. κ, µ, and ϑ still stand for the constant, the effect of the private signal, and the control variables. The intuition of Equation (8) is illustrated by Figure 6, which is similar to Figure 5 with the addition of the noisiness of the private signal. Crude oil Figure 6: Empirical model 2 Supply Demand Speculative High noise Public signal Private signal Low noise Crude oil Note: This figure reports our empirical strategy for Equation (8). The shaded variable is estimated out of the model. Unshaded variables are observed. The circle frame is the machine learning part devoted to the estimation of the OPEC signal as described in Section 2.2. The dashed frame is the econometric part of the model. For simplicity, this figure omits control variables. Equation (8) relates the effectiveness of public information to the quality of the private one. More generally, Equations (7) and (8) act as the empirical counterparts of the theoretical expressions (4) and (6) in Morris-Shin’s framework. In particular, λ and µ in the empirical setting (Equation (7)) stand for α and β in the theoretical one (Equation (4)). Additionally, 19

λ in Equation (8) captures the α > β condition implied by Equation (6). 1 3.3 Variables selection 3.3.1 Dependent variables Social welfare is difficult to capture in our framework. Hence, we assess the effectiveness of OPEC signal on oil market dynamics using two categories of dependent variables: (i) crude oil futures price volatility, and (ii) trading positions of crude oil futures market participants. Our analysis spans two time periods: March 2002-March 2021 and June 2006-March 2021, based on data availability.28 Traditional finance models argue that both volatility and trading volume serve as channels for new information to enter markets.29 Moreover, the stated objective of OPEC is to coordinate the policies of member countries to stabilize the oil market. Therefore, it is natural to consider oil price volatility when assessing the efficacy of OPEC’s public signal. Similarly, looking at market participants’ positions guides our understanding of the functioning of the crude oil market and provides evidence of the reaction of market forces to OPEC communication. We employ daily volatility estimates from WTI futures contracts traded on the Chicago Mercantile Exchange over the 2002-2021 period. Volatility is measured by the daily range—the difference between log-high and log-low prices, a widely-accepted estimator of price volatility.30 Our study incorporates the entire term structure of futures prices, ranging from 1- to 12-month maturities. This breadth allows us to understand how OPEC communications may differentially affect price expectations at various time horizons. Figure 7 reports the 3- and 12-monthfuturespricevolatility. Asexpected, volatilityistime-varying, andshorter-maturity contracts exhibit higher volatility than longer-maturity contracts. Two notable periods can be distinguishedwithimportantspikes: theGlobalFinancialCrisisandtheCOVID-19pandemic. 28To investigate the sensitivity of our results to crisis episodes, such as the COVID-19 pandemic, we also assess OPEC communication effectiveness during those specific periods (see Section 5.1.1). 29See, Epps & Epps (1976), Tauchen & Pitts (1983), and Gallant et al. (1992). 30See, for instance, Brunetti & Lildholdt (2007). For robustness, we also consider squared returns as an alternative measure of volatility. 20

Figure 7: WTI crude oil futures volatility (2002-2021) Note: This figure reports the daily WTI crude oil futures price volatility for the 3-month and 12-month maturity contracts. Volatility is proxied by log(high) - log(low) price. Data on WTI futures contracts are taken from the Chicago Mercantile Exchange. We consider weekly trading positions as measured by the DCOT from the CFTC over the 2006-2021 period.31 All the DCOT reports provide a breakdown of each Tuesday’s open interest for markets in which at least 20 traders hold positions equal to or above the reporting levels established by the CFTC. We concentrate on four categories of traders as classified by the CFTC: traditional hedgers, swap dealers, money managers, and other reportable traders. Traditional hedgers are producers, merchants and dealers (i.e., wholesalers, exporters, importers, shippers, etc.), as well as processors and users (i.e., fabricators, refiners, etc.)— hereafter P/M/D/P/U. Their main line of business concerns the physical commodity, as they are primarily engaged in the production, processing, packing, or handling a physical commodity, and they use futures markets to manage or hedge risks associated with their main activities. We also consider the positions of financial participants, such as swap dealers (SD) and money managers (MM). The former use futures markets to manage or hedge the risk associatedwithswaptransactions. Theswapdealers’counterpartiesmaybespeculativedealers, likehedgefunds, ortraditionalcommercialclientsthatmanageriskarisingfromtheirdealings in the physical commodity. Money managers are engaged in managing and conducting futures trading on behalf of clients (i.e., registered commodity trading advisory, registered commodity pool operators, or unregistered funds identified by the CFTC). Finally, the last residual category consists of positions of all other reportable (OR) traders not included in the previous 31DCOT reports are weekly publications showing holdings of different participants in futures markets. The reports are published by the CFTC and contain details on long and short open interest positions of selected categories of market participants in each futures market. See, among others, Sanders & Irwin (2013). 21

three categories.32 For each trader category, we compute the weekly net positions as the difference between long and short positions as reported in Figure 8. P/M/D/P/U are on average short, which is in line with their role of hedgers—e.g., a producer sells its production in advance in the futures market (short position) to reduce price uncertainty. Money managers are usually long and may act as counterparties to hedgers. Swap dealers’ net positions change dramatically over time depending on their swap business, but they are usually short. Figure 8: Net disaggregated trading positions (2006-2021) 4,00E+05 2,00E+05 0,00E+00 13/06/2006 13/06/2007 13/06/2008 13/06/2009 13/06/2010 13/06/2011 13/06/2012 13/06/2013 13/06/2014 13/06/2015 13/06/2016 13/06/2017 13/06/2018 13/06/2019 13/06/2020 13/06/2021 -2,00E+05 -4,00E+05 -6,00E+05 -8,00E+05 Producer/Merchant/Processor/User Swap Dealers Money Manager Other Reportables Note: Thisfiguredepictstheweeklynettradingpositions(source: DCOTreportsfromtheCFTC)asmeasured by the difference between long and short positions for each considered category of traders. 3.3.2 Independent variables We employ two types of signals: public and private. The public signal consists of OPEC communications, quantified through textual analysis, as detailed in Section 2. The private signal is approximated by the Consensus Forecast Inc., which conducts monthly surveys on oil price forecasts of about 30 market participants for 3-month and 12-month horizons. While these forecasts are certainly influenced by crude oil market fundamentals, differences across agents should reflect private information. Hence, the cross-sectional standard deviation across experts serves as our measure of the private signal—see Figure 9.33 The private signal is very high during the Global Financial Crisis (GFC), in 2011-12 when oil prices went above $100 per barrel and there was political instability in some producing 32See https://www.cftc.gov/MarketReports/CommitmentsofTraders/index.htm for more details. 33See, Ehrmann & Fratzscher (2007) for further discussion. 22

Figure 9: Standard deviation of WTI oil price consensus forecasts (2002-2021) Note: This figure reports the standard deviation of the consensus forecasts on the oil prices for 3-month and 12-month horizons. countries, in 2015-16 when oil prices plunged driven by a growing supply glut, and at the beginning of the COVID-19 pandemic. Overall, the private signal captures significant market developments.34 To capture significant market events, we control for various external factors connected to OPEC’s activities.35 Dummy variables for the March 2002-March 2021 period are introduced as follows: • Production decision variables: (i) Dp1 = 1 else 0 when OPEC decides to increase prot duction, and (ii) Dp2 = 1 else 0, when OPEC decides to decrease production. The t baseline case is “neutral” when either the level of production is kept unchanged or when there is no mention of any other decision. • Meeting type variable (anticipated vs. unanticipated announcements): Dm = 1 else 0, t when the OPEC meeting is not scheduled in advance (e.g., ordinary vs. extraordinary meetings). • OPEC behavior (acting as a cartel or not): Db = 1 else 0, when OPEC members t cooperate during the period.36 34See Ehrmann & Fratzscher (2007). 35For robustness, we also consider the global macroeconomic environment; see Section 5. 36A cartel is defined as production coordination with respect to quotas (see Brémond et al. (2012) for an empiricalanalysis). Todistinguishperiodsofcooperativeandnon-cooperativeOPECbehaviors,wefollowthe 23

4 Cheap talk or credible signal? Ouridentificationstrategyissimilar,inspirit,toKänzig(2021)andisbasedontwoconsiderations. ThefirstrelatestothedominantroleofOPECinthecrudeoilmarket. OPECproduces about 40 percent of the world’s crude oil and accounts for an estimated four-fifths of total crudeoilreserves. Hence, itisamajorplayerinthecrudeoilmarket. Marketparticipantspay attention to OPEC announcements, which can be considered the dominant event on the days when they are pronounced. The second is related to our experiment design, which is limited precisely to the days when OPEC announcements occur. Considering the specific window of the announcements’ days allows us to isolate the impact of OPEC narratives on the market variables we analyze. Applying this methodology to estimate the effects of OPEC communication on the volatility process is straightforward since we have access to daily volatility data. It is more complex when considering traders’ positions which are measured at a weekly frequency (they are reported every Tuesday). To overcome the problem of mismatch frequencies between OPEC announcements and trading positions, we align data points with respect to the OPEC signal by considering either the corresponding day of the announcement (when it coincides) or the closest, but preceding, available day. Figure 10 depicts two examples. First (in green), both the signal and the positions are reported on the same day (Tuesday), and no alignment is needed. Second (in red), OPEC signal occurs on Friday (Week 1 - Friday) between two reported trading positions (Week 1 - Tuesday and Week 2 - Tuesday). Because traders do not knowinadvancethecontentofOPECcommunication,wealignthesignal(Week1-Friday)to the next available data point (Week 2 - Tuesday). It is important to note that our alignment strategy may introduce a downward bias to our estimates since trader positions may quickly react (within the same day) to OPEC communications. We also acknowledge that other factors may drive both OPEC communication and positions of market participants in a given week.37 To address this issue, we consider various fundamental determinants of the public signal, allowing us to account for such factors that may affect OPEC announcements.38 Overall, we believe our approach is able to identify the effects of OPEC narratives on volatility and traders’ positions. In order to estimate Equations (7) and (8) and test the credibility of OPEC signal, we rely on Lasso penalized regressions. The reason is twofold. First, technically our framework faces a dimensionality problem as the sample size is not large enough compared to the parameters’ methodology discussed in Almoguera et al. (2011) and compare production quotas assigned by OPEC to the actual production levels. If actual production in period t is at least 5 percent over the quota established for that period, it indicates non-cooperation otherwise cooperation. 37Ideally, we would like to have access to daily traders’ positions. 38ThosevariablesaredescribedinSection2.1. Notethatourresultsremainrobusttovariousspecifications (see Section 5 and Appendix E). 24

space. Equations (7) and (8) count 46 and 86 coefficients, respectively, while our sample consists only of 262 observations. Second, analytically the purpose is to investigate which set of signals is credible for price volatility and trading positions. Lasso regressions allow us to overcome both problems by selecting variables that are statistically relevant, and forcing to zero the coefficient of less important variables.39 Figure 10: Frequency mismatch alignment As common for penalized regressions, the constraints on the size of coefficients depend on the magnitude of each variable. Therefore, as recommended by Tibshirani (1997), we standardize all our variables.40 Table 1 displays the estimated coefficients of the Lasso regression for Equation (7). We report results for price volatility at 1-, 6- and 12-month maturities, and trading positions for the considered categories.41 From a general viewpoint, Table 1 shows that OPEC communication affects traders’ positions morethanpricevolatility. Foreachcategoryoffactors,severaltopicshaveasignificantimpact on traders’ positions. This is particularly true for swap dealers (SD), money managers (MM), and other reportable (OR) traders. It is worth noting the dichotomy between traders engaged in the physical business (Producers/Merchants/Processors/Users) and financial traders (SD, MM, OR). Topics which are statistically important for the former are not relevant for the latter categories of traders. The only exception is Topic 9, cooperation, which implies that market participants pay attention to the credibility of OPEC communications as captured by cooperation. The effect of OPEC communication on traders’ positions is thus dependent on whether they are involved in physical or financial activities, and is particularly significant for swap dealers. This result can be explained by the fact that SD rebalance their portfolio frequently over time. They are particularly sensitive to OPEC signals aiming at stabilizing the market, as they are mainly interested in managing and hedging the risk associated with swap transactions. 39For recent applications of penalized Lasso regressions in finance, see Chinco et al. (2019), Calomiris & Mamaysky (2019), Kozak et al. (2020), Freyberger et al. (2020), and Gu et al. (2020). 40Wealsostandardizedummyvariables. Interpretationofstandardizeddummyvariablesinpenalizedregressions is often difficult. So, as a robustness check, we also perform regressions with non-standardized dummy variables (see Section 5). Results are similar to those reported in the paper and are available upon request from the authors. 41The results for other maturities are available upon request from the authors. 25

Regarding oil price volatility, OPEC tends to reduce it significantly when it intervenes on topics related to uncertainty and volatility (Topic 4), global production capacity (Topic 40), and energy policy (Topic 36). In those cases, the public signal is credible in that OPEC’s reassuring communication contributes to stabilizing the oil market. It is worth mentioning that credibility increases with maturity as some topics—such as those related to long-term production and the petroleum industry—become significant at a 12-month horizon. This result is highly interesting since it shows that OPEC communication effectively reduces oil price volatility and favors market stability, especially for long-term contracts, even in the presence of significant private signals. These results contrast those in Demirer & Kutan (2010). However, the technical approaches and the relevant variables are different.42 Overall, Equation(7)allowsustoexplainasubstantialpartofthefluctuationsofpricevolatility and trading positions over the considered period. Based on the adjusted-R2, our results show that, on average, 34 percent of crude oil price volatility variation and 57 percent of net trading positions variance are explained by our model.43 For P/M/U/P, SD, and OR we get more than 50 percent of explanatory power. Tables 2 and 3 report the effect of OPEC signals on both price volatility and trading positions dependingonthelevelofthenoiseinprivatesignals(at3months)—Equation(8).44 TheintuitionisthatOPECcommunicationmaygaincredibilitywhenprivatesignalsbecomeuncertain. As shown in Table 2, the effect of OPEC communication on price volatility appears to be more credible when private signal noise is high (credibility is measured as the proportion of significant topics in low and high regimes). OPEC credibility is, however, not constant and varies with maturity (four topics are significant at 1-month against twelve and eleven for 6- and 12-month, respectively). Based on the adjusted-R2, OPEC communication is an important element of price volatility, especially in longer maturity contracts (on average, the regressions explain 48 percent of the variance of the endogenous variables). Combining the results in Tables 1 and 2, we find that the maturity of the crude oil futures contract matters. In particular, our results indicate that OPEC signals have a stronger impact on longer maturity contracts. In fact, while only supply-related topics (Topics 20 and 29) are important at 1-month horizon, multiple types of signals on supply, price, and shortage are significant at 6and 12-month maturities—the most important are global supply (Topic 20), market stability (Topic14),economicgrowthandoildemand(Topics9and12). Multiplesignalcommunication isthenanefficientstrategyinthelongruniftheOrganizationintendstoaffectpricevolatility. 42Demirer & Kutan (2010) rely on the event study methodology to assess the effects of OPEC conferences on the crude oil market activity in the US. 43We also use the deviance ratio as a measure of the explanatory power of our model. Results are robust and available upon request from the authors. 44Resultsfor12-monthnoisearesimilartothosereportedandareavailableuponrequestfromtheauthors. 26

Table 1: Effectiveness of public and private signals Note: Thistable reports theestimated coefficients ofthe penalized bootstrap Lassoregressionsfor both price volatility(inpercentagepoints)andtradingpositionsorganizedasclusters(seecommunitiesinFigure3). We only report statistically significant topics. ∗ indicates significance at the 5% level. "X" indicates zero value coefficients. To save space, coefficients of trading positions are divided by 1000. TheeffectofOPECcommunicationontradingpositions(Table3)isgenerallymoresignificant than on price volatility. Our models explain on average 61 percent of net positions’ variability across all traders. The adjusted-R2 amounts to more than 50 percent for P/M/D/P/U (67 percent), SD (59 percent), and OR (67 percent). While OPEC signal matters for each trading category, it is even more important when private noise is high and mainly for investors 27

thatarepredominantlyengagedinphysicalcommodities(namelyP/M/D/P/U).Surprisingly, for traders involved in financial activities with no physical exposures (MM and SD), OPEC credibility is qualitatively unaffected by the amount of private noise. Similar to the results in Table 1, the dichotomy between financial and physical traders also matters in terms of the signal (topics of OPEC communication). Physical traders’ (P/M/D/P/U) net positions are mainlypositivelyaffectedbysupply-(Topics9and34)andshortage-related(Topic11)topics. Financial (MM and SD) and OR net positions, on the other hand, are impacted by many topics. In particular, MM and OR mainly respond positively to supply-related communications and negatively to long-term investment signals. SD move in the opposite direction, reacting negatively to supply signals and positively to price and investment signals, perhaps because SD act as a counterpart to MM and OR. Interestingly, in line with current debates, Topic 19 on climate change only impacts positions from traders engaged in physical commodities (P/M/D/P/U) but has no role on financial ones. It is also important to point out the strong effect of OPEC cooperation on trading positions. Overall, our results indicate that OPEC communication is relevant and effective. First and importantly, it achieves the objective of stabilizing the crude oil market by reducing volatility levels of crude oil futures prices (Table 1). These results are stronger for longer maturity contracts,indicatingthatOPECcommunicationaffectsthetermstructureofoilfutures. Sincethe main mandate of the Organization is to stabilize the oil market, our results provide evidence that OPEC fulfills its mandate. Second, market participants react to OPEC communication by readjusting their net positions. This represents additional evidence that OPEC communication matters. Topics covered in OPEC communications explain a large part (measured by adjusted-R2) of the variation in trader positions. Topics related to OPEC credibility seem to be particularly relevant. Finally, our results represent an empirical test of the Morris and Shin (2002) theory. OPEC communication seems to provide a credible signal, and this signal is stronger the higher the noise in the private signal. 5 Robustness and placebo tests 5.1 Robustness checks 5.1.1 OPEC communication in recent crises As emphasized above, OPEC communications are credible in providing signals about market fundamentals influencing crude oil price volatility and net trading positions. Credibility increases when public announcements interact with noisy private signals, making communications an important tool to shape expectations. During financial and economic crises, however, OPEC lacks the ability to implement efficient coordinated production decisions. This is mainly due to the structure of the Organization, whichlacksaformalenforcementmechanismconstrainingmemberstocomplywiththeagreed 28

production quotas.45 As for central banks, public communications then become even more important to anchor market expectations.46 To investigate how efficient OPEC communications are during unconventional times, we consider two recent turmoil periods included in our sample: the Global Financial Crisis (3/5/2007 - 12/17/2008) and the COVID-19 pandemic (3/5/2020 - 3/4/2021).47 We gather these two episodes together to avoid small sample issues, and run our estimations of Equations (7) and (8). Table 4 reports the effectiveness of OPEC communication during crisis times. Interestingly, compared to our previous results (see Table 1), the explanatory power (based on the adjusted- R2)increasessubstantiallyforallofourdependentvariables,butP/M/D/P/U(yet,itremains high at 50 percent). However, only a few topics are statistically significant. Not surprisingly, “OPECcooperation,” “Productionadjustment/COVID,” and“Oilmarketstability” areimportant topics for the volatility process. The coefficients are positive, which implies that OPEC communication relative to those topics is associated with higher volatility, perhaps linked to OPEC’s fragile structure in terms of enforcing production decisions. Positions of both commercial and non-commercial traders are mainly characterized by negative coefficients, implying that OPEC communications are associated to a reduction in net traders’ positions. The opposite is true for other positions. Interestingly, Topic 15 refers to the OPEC-Russia relationship and is important for traders’ positions, suggesting that in crisis periods, OPEC+ alliance may play a critical role. Tables 9 and 10 in Appendix D report the results from Equation (8) with high private noise over the selected crisis periods. Compared to Tables 2 and 3, the explanatory power of our models is globally more important during crisis periods for both crude oil price volatility and trading positions. In line with the Morris-Shin predictions and with our findings in Section 4, OPEC’s public signal is more significant when private noise is high. 45See Fattouh & Mahadeva (2013). 46See Eggertsson & Woodford (2003), Coenen et al. (2017), and Hubert & Labondance (2018) for some discussions on the role of central banks’ communications during unconventional times. 47Selected time ranges are based on NBER business cycles dating. 29

Table 2: Effectiveness of public and private signals on price volatility in high and low private noise (3 months) Note: This table reports the estimated coefficients of the penalized bootstrap Lasso regressions for price volatility (in percentage points) during high and low private noise (3 months) organized as clusters (see communities in Figure 3). We only report statistically significant topics. ∗ indicates significance at the 5% level. "X" indicates zero value coefficients. 30

Table3: Effectivenessofpublicandprivatesignalsontradingpositionsinhighandlowprivate noise (3 months) Note: This table reports the estimated coefficients of the penalized bootstrap Lasso regressions for trading positions during high and low private noise (3 months) organized as clusters (see communities in Figure 3). Weonlyreportstatisticallysignificanttopics. ∗indicatessignificanceatthe5%level. "X"indicateszerovalue coefficients. To save space, coefficients of trading positions are divided by 1000. 31

5.1.2 Robustness to different specifications To assess the robustness of our findings, we estimate various alternative specifications: (i) including additional macroeconomic variables,48 (ii) excluding the 2007-2008 boom-bust period, (iii) distinguishing between high and low OPEC spare capacity, and (iv) without normalizing dummy variables. We also investigate the effectiveness of public and private signals in high andlow12-monthprivatenoise. Finally, insteadofconsideringthevolatilityontheannouncement day, we computed the change in volatility with respect to the previous day. The results from these alternative specifications are similar to those reported in the paper, illustrating the robustness of our findings. As an example, Table 11 in Appendix E reports the estimation results corresponding to the case where the ADS index has been replaced by Kilian (2009)’s index of global real economic activity (denoted as “dry cargo” in the table). As shown, they are identical to those we previously obtained.49 5.1.3 Selective inference problem A challenging task when estimating high-dimensional statistics is to make inference accounting for uncertainty and hypothesis testing. This statistical problem is known as “selective inference” and raises concerns about the effects of variables’ selection on inference.50 Several methodsexisttocorrecttheproblem.51 Inthecoreofthepaper,wereporttheresultsfromthe residual bootstrap Lasso regression proposed by Chatterjee & Lahiri (2013). The bootstrap procedure allows us to measure parameters uncertainty in terms of confidence intervals. To check the robustness of our results, we use numerous frequentist and Bayesian methods that proved to be efficient in high-dimensional setting. Tables 12 to 14 in Appendix E report the results from the following methods for Equation (7): (i) bootstrap Lasso + Partial Ridge (Liu et al. (2020), Table 12), (ii) bootstrap de-sparsed Lasso (Zhang & Zhang (2014), Table 13), and (iii) Bayesian Lasso (Park & Casella (2008), Table 14).52 Our main conclusions remain valid regardless of the estimation method adopted. 48Specifically, we estimate three specifications: (i) a specification including the eight variables mentioned in Section 2.1, (ii) the same model in which the ADS index is replaced by the index of global real economic activity proposed by Kilian (2009), and (iii) a specification in which only the index of global real economic activity is included. 49To save space, we do not report all the results related to our robustness checks to different specifications, but they are available upon request from the authors. 50See, e.g., Taylor & Tibshirani (2015) for a discussion. 51See Dezeure et al. (2015) for a review of the most common existing methods. 52WealsoperformedbootstrapLasso+OLS(Liu&Yu(2013)),andmultisample-splittingmethods. Results are robust and available upon request from the authors. Results for Equation (8) also corroborate our main conclusions and are available upon request. 32

5.2 Placebo test As with all regression settings, an important identification assumption we make is that the responses of price volatility and trading positions we observe are the consequence of OPEC communications rather than the result of intrinsic dynamics in the oil market. To test the relevanceofouridentificationstructure, weconductaplacebotestduringdayswithnoOPEC announcements. We construct placebo samples by suppressing OPEC announcements days. From the placebo samples, we build control groups by sampling out with replacement 262 observations from futures price volatility over each maturity, and 208 observations from trading positions for each category of traders.53 For each control group, we repeat the estimation procedure of Equations (7) and (8).54 Our results unequivocally indicate that all coefficients capturing OPEC announcements are zero (this is due to the use of LASSO penalized regressions). This finding shows that OPEC announcements are relevant when they occur and provide support to our identification strategy. 6 Conclusions InthispaperweareinterestedinanalyzingthecontentofOPECcommunicationsandwhether it provides valuable information to the crude oil market. Starting from the Morris & Shin (2002) framework, we derive an empirical strategy which assumes that fundamental factors related to supply, demand, and speculative activity drive OPEC’s public signal. Both public and private signals affect the crude oil market. Our results suggest that OPEC narratives cover a wide range of topics that are indeed linked to the fundamental factors we consider. We also find that OPEC narratives are relevant in the sense that they reduce crude oil price volatility and prompt market participants to rebalance their positions. Our results stimulate further research. It would be interesting to know which market participants have the largest and fastest reactions to OPEC announcements. However, we recognize that data limitations pose an obstacle. In fact, to perform this analysis we would need access to confidential, detailed market participant positions. It would also be important to understand how our results in the crude oil market are transmitted to other markets. Crude oil is extremely important for both the real economy and financial markets. Understanding possible contagion mechanisms will help identify interconnectedness effects. Finally, our findings show that climate change is an important topic in OPEC communication. Studying how climate-related risks influence OPEC narratives and, in turn, the crude oil market, is critical to policymakers, market participants, and the general public. 53Recall,thereare262and208OPECannouncementdayswhenconsideringvolatility(sampleperiod2002- 2021) and traders’ positions (sample period 2006-2021), respectively. 54The random placebo procedure is used to avoid any subjectivity in the choice of the pre-OPEC announcementdays. Asarobustnesscheck,wealsoperformthesameanalysisbymanuallyselectingpre-OPEC announcement days. Results are robust and available upon request. 33

Table 4: Effectiveness of communication during crisis periods Note: Thistable reports theestimated coefficients ofthe penalized bootstrap Lassoregressionsfor both price volatility (in percentage points) and trading positions organized as clusters (see communities in Figure 3) during crisis periods. We only report statistically significant topics. ∗ indicates significance at the 5% level. "X" indicates zero value coefficients. To save space, coefficients of trading positions are divided by 1000. References Almoguera, P., Douglas, C. & Herrera, A. (2011), ‘Testing for the cartel in OPEC: non-cooperative collusion or just non-cooperative?’, Oxford Review of Economic Policy 34

27(1), 144–168. Amato, J. D., Morris, S. & Shin, H. S. (2002), ‘Communication and monetary policy’, Oxford Review of Economic Policy 18(4), 495–503. Aruoba, S., Diebold, F. & Scotti, C. (2009), ‘Real-time measurement of business conditions’, Journal of Business & Economic Statistics 27(4), 417–427. Baumeister, C. & Kilian, L. (2016), ‘Forty years of oil price fluctuations: Why the price of oil may still surprise us’, Journal of Economic Perspectives 30(1), 139–60. Bischof, J. & Airoldi, E. M. (2012), Summarizing topical content with word frequency and exclusivity, in‘Proceedingsofthe29thInternationalConferenceonMachineLearning(ICML- 12)’, pp. 201–208. Blei, D. & Lafferty, J. (2009), ‘Topic models’, Text mining: classification, clustering, and applications 10(71), 34. Blei, D. M. (2012), ‘Probabilistic topic models’, Communications of the ACM 55(4), 77–84. Blei, D. M. & Lafferty, J. D. (2006), Dynamic topic models, in ‘Proceedings of the 23rd international conference on Machine Learning’, pp. 113–120. Blei, D. M. & Lafferty, J. D. (2007), ‘A correlated topic model of science’, The Annals of Applied Statistics 1(1), 17–35. Blei, D., Ng, A. & Jordan, M. (2003), ‘Latent dirichlet allocation’, Journal of Machine Learning Research 3(Jan), 993–1022. Blondel, V. D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. (2008), ‘Fast unfolding of communities in large networks’, Journal of Statistical Mechanics: Theory and Experiment 2008(10), P10008. Brémond, V., Hache, E.&Mignon, V.(2012), ‘DoesOPECstillexistasacartel? anempirical investigation’, Energy Economics 34(1), 125–131. Brunetti, C., Büyükşahin, B. & Harris, J. H. (2016), ‘Speculators, prices, and market volatility’, Journal of Financial and Quantitative Analysis 51(5), 1545–1574. Brunetti, C., Buyuksahin, B., Robe, M. & Soneson, K. (2013), ‘OPEC "fair" price pronouncements and the market price of crude oil’, The Energy Journal 34, 79–108. Brunetti, C. & Lildholdt, P. M. (2007), ‘Time series modeling of daily log-price ranges for CHF/USD and USD/GBP’, The Journal of Derivatives 15(2), 39–59. Calomiris, C. W. & Mamaysky, H. (2019), ‘How news and its context drive risk and returns around the world’, Journal of Financial Economics 133(2), 299–336. 35

Chang, J., Gerrish, S., Wang, C., Boyd-Graber, J. & Blei, D. (2009), ‘Reading tea leaves: How humans interpret topic models’, Advances in neural information processing systems 22, 288–296. Chatterjee, A.&Lahiri, S.N.(2013), ‘RatesofconvergenceoftheadaptiveLASSOestimators to the oracle distribution and higher order refinements by the bootstrap’, The Annals of Statistics 41(3), 1232–1259. Chinco, A., Clark-Joseph, A. D. & Ye, M. (2019), ‘Sparse signals in the cross-section of returns’, The Journal of Finance 74(1), 449–492. Coenen, G., Ehrmann, M., Gaballo, G., Hoffmann, P., Nakov, A., Nardelli, S., Persson, E. & Strasser, G. (2017), Communication of monetary policy in unconventional times, Working Paper Series 2080, European Central Bank. URL: https://ideas.repec.org/p/ecb/ecbwps/20172080.html Demirer, R. & Kutan, A. M. (2010), ‘The behavior of crude oil spot and futures prices around OPEC and SPR announcements: An event study perspective’, Energy Economics 32(6), 1467–1476. Dezeure, R., Bühlmann, P., Meier, L. & Meinshausen, N. (2015), ‘High-dimensional inference: confidence intervals, p-values and r-software hdi’, Statistical Science pp. 533–558. Eggertsson, G. B. & Woodford, M. (2003), ‘Optimal monetary policy in a liquidity trap’. Ehrmann, M. & Fratzscher, M. (2007), Social value of public information: testing the limits to transparency, Working Paper Series 821, European Central Bank. URL: https://ideas.repec.org/p/ecb/ecbwps/2007821.html Epps, T. & Epps, M. (1976), ‘The stochastic dependence of security price changes and transaction volumes: Implications for the mixture-of-distributions hypothesis’, Econometrica 44(2), 305–321. Evans, C., Fisher, J., Justiniano, A. & Campbell, J. (2012), ‘Macroeconomic effects of FOMC forward guidance’, Brookings Papers on Economic Activity . Farrell,J.&Rabin,M.(1996),‘Cheaptalk’,Journal of Economic Perspectives10(3),103–118. Fattouh, B. (2007), OPEC pricing power: The need for a new perspective, Oxford Institute for Energy Studies. Fattouh,B.&Mahadeva,L.(2013),‘OPEC:whatdifferencehasitmade?’,Annu.Rev.Resour. Econ. 5(1), 427–443. Freyberger, J., Neuhierl, A. & Weber, M. (2020), ‘Dissecting characteristics nonparametrically’, The Review of Financial Studies 33(5), 2326–2377. Gallant, R., Rossi, P. & Tauchen, G. (1992), ‘Stock prices and volume’, The Review of Financial Studies 5(2), 199–242. 36

Griffiths, T. & Steyvers, M. (2004), ‘Finding scientific topics’, Proceedings of the National Academy of Sciences 101(suppl 1), 5228–5235. Gu, S., Kelly, B. & Xiu, D. (2020), ‘Empirical asset pricing via machine learning’, The Review of Financial Studies 33(5), 2223–2273. Guidi, M. G., Russell, A. & Tarbert, H. (2006), ‘The effect of OPEC policy decisions on oil and stock prices’, OPEC Energy Review 30(1), 1–18. Hansen, S. & McMahon, M. (2016), ‘Shocking language: Understanding the macroeconomic effects of central bank communication’, Journal of International Economics 99, S114–S133. Hansen, S., McMahon, M. & Prat, A. (2018), ‘Transparency and deliberation within the FOMC: a computational linguistics approach’, The Quarterly Journal of Economics 133(2), 801–870. Hermann,M.&Fratzscher,M.(2007a),‘Communicationanddecision-makingbycentralbank committees: Differentstrategies, sameeffectiveness’, Journal of Money, Credit and Banking 39(2-3), 509–541. Hermann, M. & Fratzscher, M. (2007b), ‘Transparency, disclosure and the Federal Reserve’, International Journal of Central Banking 3(1), 179–225. Hubert, P. & Labondance, F. (2018), ‘The effect of ECB forward guidance on the term structure of interest rates’, International Journal of Central Banking 56. Känzig, D. (2021), ‘The macroeconomic effects of oil supply news: Evidence from OPEC announcements’, American Economic Review 111(4), 1092–1125. Kilian, L. (2009), ‘Not all oil price shocks are alike: Disentangling demand and supply shocks in the crude oil market’, American Economic Review 99(3), 1053–69. Kilian, L. (2023), How to Construct Monthly VAR Proxies Based on Daily Futures Market Surprises, Working Papers 2310, Federal Reserve Bank of Dallas. Kilian, L. & Murphy, D. P. (2014), ‘The role of inventories and speculative trading in the global market for crude oil’, Journal of Applied Econometrics 29(3), 454–478. Kozak, S., Nagel, S. & Santosh, S. (2020), ‘Shrinking the cross-section’, Journal of Financial Economics 135(2), 271–292. Larsen, V. H. & Thorsrud, L. A. (2019), ‘The value of news for economic developments’, Journal of Econometrics 210(1), 203–218. Lin, S. & Tamvakis, M. (2010), ‘OPEC announcements and their effects on crude oil prices’, Energy Policy 38(2), 1010–1016. Liu, H., Xu, X. & Li, J. J. (2020), ‘A bootstrap lasso+ partial ridge method to construct confidence intervals for parameters in high-dimensional sparse linear models’, Statistica Sinica 30(3), 1333–1355. 37

Liu, H. & Yu, B. (2013), ‘Asymptotic properties of lasso+ mLS and lasso+ ridge in sparse high-dimensional linear regression’, Electronic Journal of Statistics 7, 3124–3169. Loutia, A., Mellios, C. & Andriosopoulos, K. (2016), ‘Do OPEC announcements influence oil prices?’, Energy Policy 90, 262–272. Meilă, M. (2003), Comparing clusterings by the variation of information, in ‘Learning theory and kernel machines’, Springer, pp. 173–187. Mimno,D.,Wallach,H.,Talley,E.,Leenders,M.&McCallum,A.(2011),Optimizingsemantic coherence in topic models, in ‘Proceedings of the 2011 conference on empirical methods in natural language processing’, pp. 262–272. Morris, S. & Shin, H. S. (2002), ‘Social value of public information’, American Economic Review 92(5), 1521–1534. Morris, S. & Shin, H. S. (2005), ‘Central bank transparency and the signal value of prices’, Brookings Papers on Economic Activity 2005(2), 1–66. Morris, S., Shin, H. S. & Tong, H. (2006), ‘Social value of public information: Morris and shin (2002) is actually pro-transparency, not con: Reply’, American Economic Review 96(1), 453–455. Park, T. & Casella, G. (2008), ‘The bayesian lasso’, Journal of the American Statistical Association 103(482), 681–686. Pescatori, M. A. & Nazer, Y. F. (2022), OPEC and the Oil Market, IMF Working Papers 2022/183, International Monetary Fund. Pons, P. & Latapy, M. (2006), Computing communities in large networks using random walks, in ‘J. Graph Algorithms Appl’, Citeseer. Quint, D. & Venditti, F. (2020), ‘The influence of OPEC+ on oil prices: a quantitative assessment’, ECB Working Paper 2467 . Raghavan, U. N., Albert, R. & Kumara, S. (2007), ‘Near linear time algorithm to detect community structures in large-scale networks’, Physical Review E 76(3), 036106. Roberts, M. E., Stewart, B. M. &Tingley, D. (2016), ‘Navigatingthe local modesof bigdata’, Computational Social Science 51. Roberts, M. E., Stewart, B. M. & Tingley, D. (2019), ‘Stm: An R package for structural topic models’, Journal of Statistical Software 91(1), 1–40. Roberts, M. E., Stewart, B. M., Tingley, D., Airoldi, E. M. et al. (2013), The structural topic model and applied social science, in ‘Advances in neural information processing systems workshop on topic models: computation, application, and evaluation’, Vol. 4, Harrahs and Harveys, Lake Tahoe, pp. 1–20. 38

Roberts,M.E.,Stewart,B.M.,Tingley,D.,Lucas,C.,Leder-Luis,J.,Gadarian,S.K.,Albertson, B. & Rand, D. G. (2014), ‘Structural topic models for open-ended survey responses’, American Journal of Political Science 58(4), 1064–1082. Rosvall, M. & Bergstrom, C. T. (2007), ‘Maps of information flow reveal community structure in complex networks’, arXiv preprint physics.soc-ph/0707.0609 . Sanders, D. & Irwin, S. (2013), ‘Measuring index investment in commodity futures markets’, The Energy Journal 34(3), 105–127. Scotti, C. (2016), ‘Surprise and uncertainty indexes: Real-time aggregation of real-activity macro-surprises’, Journal of Monetary Economics 82, 1–19. Svensson, L. E. (2006), ‘Social value of public information: Comment: Morris and shin (2002) is actually pro-transparency, not con’, American Economic Review 96(1), 448–452. Taddy, M. (2012), On estimation and selection for topic models, in ‘Artificial Intelligence and Statistics’, PMLR, pp. 1184–1193. Tauchen, G. & Pitts, M. (1983), ‘The price variability-volume relationship on speculative markets’, Econometrica 51(2), 485–505. Taylor, J. & Tibshirani, R. J. (2015), ‘Statistical learning and selective inference’, Proceedings of the National Academy of Sciences 112(25), 7629–7634. Tibshirani, R. (1997), ‘The lasso method for variable selection in the Cox model’, Statistics in Medicine 16(4), 385–395. Wallach, H. M., Murray, I., Salakhutdinov, R. & Mimno, D. (2009), Evaluation methods for topic models, in ‘Proceedings of the 26th annual international conference on machine learning’, pp. 1105–1112. Wirl, F. & Kujundzic, A. (2004), ‘The impact of OPEC conference outcomes on world oil prices 1984-2001’, The Energy Journal 25(1). Zhang, C.-H. & Zhang, S. S. (2014), ‘Confidence intervals for low dimensional parameters in high dimensional linear models’, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 76(1), 217–242. Appendix A Pre-processing, and model selection As discussed in Section 2.3, dimensionality reduction is a key first step when using computational linguistic models since text datasets are often both large and scarce. Table 5 reports 39

thepre-processingstepsusedinthepaperwiththeevolutionoftotalwords. Ourcorpusstarts with 12586 words and ends with 2573 meaningful words after preprocessing. Table 5: Data dimensionality reduction of each preprocessing step Raw Remove Remove Remove words < 3 length + Stopwords Numbers Stemming Text & Given Names & Punctuations Algorithm Total words 12586 11883 7400 5027 Notes: The table reports the evolution of total words through pre-processing. The stemming algorithm is the Porter stemmerimplementedinRusing‘tm’package. Another important element in estimating topic models is the number K of topics. We use several methods to help choosing the number of topics. Figure 11 reports our considered evaluationmeasuresforTopics20to60,suchastheheld-outlikelihood(Wallachetal.(2009)), the residual checks (Taddy (2012)), and the lower bound.55 One needs to find the right trade-off for all measures, namely the number of topics for which each considered criteria is reasonably good. Based on our diagnostic, we select K = 40. Another selection criteria used in the literature is the semantic coherence developed by Mimno et al. (2011).56 As noted by Roberts et al. (2014), semantic coherence alone is relatively easy to achieve by having only a couple of topics which all are dominated by the most common words. We therefore follow Roberts et al. (2014) and report in Figure 12 a combination of semantic coherence and exclusivity of words to topics.57 The coherence-exclusivity trade-off confirms our choice of K = 40. 55For a discussion on each measure, see Roberts et al. (2019). 56Semanticcoherenceisrelatedtopointwisemutualinformationandismaximizedwhenthemostprobable words in a given topic frequently co-occur together. 57Inourcase,exclusivityismeasuredbyFREXmetric(seeBischof&Airoldi(2012)). SectionBinAppendix discusses in more details the FREX measure. 40

Figure 11: Diagnostic values by number of topics Note: This figure reports different measures of topic selection for several topics values (from 20 to 60). Both held-out likelihood and lower bound have to be maximized, while residual diagnostic need to be minimized. Figure 12: Semantic exclusivity vs coherence 41

B Topic labeling This section briefly presents the two approaches we use for topic labeling. Recall that labels play no role in the analysis but provide a convenient way to discuss our results. For each of the 40 topics, we first use the FREX metric defined as the weighted harmonic mean of the word’s rank in terms of exclusivity and frequency:  −1 ω 1−ω FREX k,ν =  (cid:16) (cid:17) +  ECDF β / (cid:80)K β ECDF (β k,ν ) k,ν j=1 j,ν whereECDF isthefrequencyscoregiventheempiricalCDF ofthewordinitstopicdistribution. ω is the weight sets to 0.7 (to favor exclusivity). Exclusivity is calculated by normalizing theβ matrix(i.e.,theconditionalprobabilityoftopicsgiventheword). Wordswithhighvalue are those where most of the mass for these words is assigned to the given topic. Together with FREX, we also use the most-probable bigrams.58 Both metrics are reported for each topic in Tables 6 to 8. 58A bigram is an association of two words. 42

Table 6: Estimated topics and labeling (Topic 1 to 15) Topics Label Top 10 terms meet, market, opec, organ, petroleum, Topic 1 Extraordinary meetings current, countri, extraordinari, republ, suppli basket, refer, wti, cut, crude, Topic 2 Basket price barrel, quarter, averag, russia, month howev, quarter, extraordinari, wish, like, Topic 3 Oil shortage ceil, level, can, purpos, deepest volatil, specul, fundament, crude, geopolit, Topic 4 Oil uncertainty/volatility oilpric, price, increas, day, comfort algier, accord, agreement, committe, algeria, Topic 5 Rebalancing market forward, rebalanc, overhang, reactiv, high-level strategi, long-term, object, futur, Topic 6 Long-term strategy consist, identifi, adopt, multilater, role, technolog Physical/Financial physic, workshop, financi, ief, interact, Topic 7 interaction evolv, iea, regul, three, event iraq, visit, iraqi, said, algier Topic 8 Iraq-Saudi relations aramco, achiev, prime, venezuela, extens committe, declar, nopec, technic, monitor, Topic 9 Cooperation adjust, voluntari, joint, particip, return join, sovereign, declar, peopl, join, Topic 10 OPEC producers right, organ, cooper, withdraw, nation growth, barrel, economic growth, averag, project, Topic 11 Economic growth year, forecast, like, oecd, balanc like, locat, recoveri, sign, posit, Topic 12 Oil demand move, citi, district, general, libyan Spare oil capac, increas, rise, spare, downstream, Topic 13 prod. capacity avail, around, product, addit, crude oil countri, market, oilmarket, opec, oil, Topic 14 Oil market stability meet, stabil, member, world, global india, high-level, parti, russianf, long-term, Topic 15 OPEC-Russia relations opec-russia, dialogu, meet, senior, technic Note: This table reports labels for Topics 1 to 15 based on both most probable bigrams (column “Label”) and top 10 FREXterms(column“Top10terms”). 43

Table 7: Estimated topics and labeling (Topic 16 to 30) Topics Label Top 10 terms market, price, oil, opec, suppli, Topic 16 Energy investments invest, produc, consum, energi, oilpric composit, eleven, heavier, distil, orb, Topic 17 Heavy crude iran, trial, weight, temporarili, index data, media, tool, avail, big, Topic 18 Oil and gas market exercis, phase, project, uae, statist climat, chang, pari, convent, framework Topic 19 Climate change negoti, agreement, diversif, implement, sustain adjust, reaffirm, epidem, COVID, declar, Topic 20 Prod. adjustment / COVID compens, outbreak, particip, product, agre adequ, suppli, situat, level, price, Topic 21 Adequate supply band, close, light, qatar, consum gecf, area, data, experi, mutual, Topic 22 Intergovernmental Relations gas, exchang, two, common, sign kuwait, declar, implement, prime, met, Topic 23 Kuwait cooperation earlier, visit, cooper, role, congratul hurrican, condol, peopl, katrina, devast, Topic 24 Natural disaster caus, unit, sad, capac, govern OPEC production committe, compens, conform, month, particip, Topic 25 adjustments adjust, schedul, full, overal, rebalanc believ, prove, humankind, back, abl, Topic 26 Energy poverty statement, challenge, reflect, histori, togeth china, workshop, iea, region, opec-china, Topic 27 OPEC-MENA/China relations uncertainti, prospect, dialogu, demand, mena nigeria, univers, oil, nation, serv, Topic 28 Nigeria crude oil nigerian, former, petroleum, institut, gas iran, qatar, algeria, current, consult, Topic 29 Minister energy negotiations accord, restor, attend, negat, oilmarket ief, compar, iea, energi, outlook, Topic 30 Energy Outlook scenario, transit, iea, transpar, agenc Note: ThistablereportslabelsforTopics16to30basedonbothmostprobablebigrams(column“Label”)andtop10 FREXterms(column“Top10terms”). 44

Table 8: Estimated topics and labeling (Topic 31 to 40) Topics Label Top 10 terms india, dialogu, iea, visit, pandem, Topic 31 OPEC/Asia dialogue research, energi, center, exchang, cooper coal, unit, gas, visit, center, Topic 32 Gas & coal markets commod, state, oil, outlook, imf compani, ceo, india, sector, invest, Topic 33 Oil industry offici, total, industri, spoke, eni OPEC/Non-OPEC meet, declar, adjust, particip, cooper, Topic 34 production participation joint, voluntari, opec-nopec, month, produc special, ceremoni, recess, exhibit, activ, Topic 35 World Economy govern, packag, golden, stamp, perform energy, opec, technolog, polici, brussel, Topic 36 Energy policy european, progress, dialogu, demand, fuelenergi, south, africa, osaka, shall, japan, Topic 37 Energy security cop, poverti, secur, kyoto, protocol ministri, egypt, meet, let, oil-produc, Topic 38 Petroleum industry shall, observ, come, scientif, ministri factor, reason, price, geopolit, pressur, Topic 39 Production ceiling stabil, specul, ceil, measur, tension nigeria, gas, reserv, doha, therefor, Topic 40 Global prod. capacity domin, price, noc, proud, polit Note: ThistablereportslabelsforTopics31to40basedonbothmostprobablebigrams(column“Label”)andtop10 FREXterms(column“Top10terms”). 45

C Time evolution of OPEC signals Figure 13: Supply adjustment signal evolution (orange community) Note: Thisfigurerepresentsthetopicsprobabilityovertimeintheorangecommunityusingakernelsmoothing transformation(Daniellmethod). Thewindowsizeis6pointswhichroughlycorrespondsto6-monthsperiod. 46

Figure 14: OPEC international relations signal (brown and purple communities) Note: This figure represents the topics probability over time in the brown (panel (a)) and purple (panel (b)) communities using a kernel smoothing transformation (Daniell method). The window size is 6 points which roughly corresponds to 6-months period. 47

Figure 15: Price volatility and market stability signal (green community) Note: Thisfigurerepresentsthetopicsprobabilityovertimeinthegreencommunityusingakernelsmoothing transformation(Daniellmethod). Thewindowsizeis6pointswhichroughlycorrespondsto6-monthsperiod. 48

Figure 16: Long-term investment signal (yellow and darkblue communities) Note: Thisfigurerepresentsthetopicsprobabilityovertimeintheyellow(panel(a))anddarkblue(panel(b)) communities using a kernel smoothing transformation (Daniell method). The window size is 6 points which roughly corresponds to 6-months period. 49

Figure 17: Oil shortage signal (light blue community) Note: This figure represents the topics probability over time in the light blue community using a kernel smoothing transformation (Daniell method). The window size is 6 points which roughly corresponds to 6months period. 50

Figure 18: Energy policy signal (grey community) Note: This figure represents the topics probability over time in the grey community using a kernel smoothing transformation(Daniellmethod). Thewindowsizeis6pointswhichroughlycorrespondsto6-monthsperiod. 51

D OPEC communication in recent crises Table9: Effectivenessofcommunicationonpricevolatilityinhighandlowprivatenoiseduring crisis periods (3 months) Note: This table reports the estimated coefficients of the penalized bootstrap Lasso regressions for price volatility (in percentage points) during high and low private noise (3 months) organized as clusters (see communities in Figure 3) during crisis period. We only report statistically significant topics. ∗ indicates significance at the 5% level. "X" indicates zero value coefficients. 52

Table 10: Effectiveness of public and private signals on trading positions in high and low private noise during crisis periods (3 months) Note: This table reports the estimated coefficients of the penalized bootstrap Lasso regressions for trading positions during high and low private noise (3 months) organized as clusters (see communities in Figure 3). Weonlyreportstatisticallysignificanttopics. ∗indicatessignificanceatthe5%level. "X"indicateszerovalue coefficients. To save space, coefficients of trading positions are divided by 1000. 53

E Robustness checks E.1 Macro specification Table 11: Effectiveness of public and private signals (macro specification) Note: ThistablereportstheestimatedcoefficientsofthepenalizedbootstrapLassoforbothpricevolatility(in percentage points) and trading positions organized as clusters (see communities in Figure 3). We only report statistically significant topics. ∗ indicates significance at the 5% level. "X" indicates zero value coefficients. To save space, coefficients of trading positions are divided by 1000. “Dry Cargo” is the index of global real economic activity developed by Kilian (2009). 54

E.2 Selective inference Table 12: Bootstrap Lasso + Partial Ridge estimation Note: ThistablereportstheestimatedcoefficientsofthepenalizedbootstrapLasso+partialridgeregressions forbothpricevolatility(inpercentagepoints)andtradingpositionsorganizedasclusters(seecommunitiesin Figure3). Weonlyreportstatisticallysignificanttopics. ∗indicatessignificanceatthe5%level. "X"indicates zero value coefficients. To save space, coefficients of trading positions are divided by 1000. 55

Table 13: Bootstrap de-sparsed Lasso estimation Note: This table reports the estimated coefficients of the penalized bootstrap de-parsed Lasso regressions for both price volatility (in percentage points) and trading positions organized as clusters (see communities in Figure 3). We only report statistically significant topics. Estimations and confidence intervals are computed over B = 5000 bootstrap replications. A 5-fold cross validation procedure has been performed to select the Lasso penalty λ. ∗ indicates significance the at 5% level. "X" indicates zero value coefficients. To save space, coefficients of trading positions are divided by 1000. 56

Table 14: Bayesian Lasso estimation Note: This table reports the Bayesian lasso estimation for both price volatility (in percentage points) and trading positions organized as clusters (see communities in Figure 3). "X" indicates zero value coefficients. ∗ indicates significance in a bayesian sense. To save space, coefficients of trading positions are divided by 1000. 57