spb/phd_thesis Public

1\chapter*{Introduction}         % ne pas numéroter2\label{chap-introduction}       % étiquette pour renvois3\phantomsection\addcontentsline{toc}{chapter}{\nameref{chap-introduction}} % inclure dans TdM45% BROUILLON À RÉVISER PAR L'AUTEUR — introduction générale de la thèse6% (rédigée à partir des trois articles ; les renvois bibliographiques utilisent7% la bibliographie consolidée de la thèse).89Futures markets are where much of the world's price discovery takes place.10Crude oil, natural gas, gold, equity indexes, Treasury notes: for each of11these assets, the futures contract is typically the most liquid instrument,12the first to react to news, and the reference point from which spot prices,13inventories, and investment decisions are set. Understanding how information14is impounded into futures prices --- and how that process shapes volatility15--- is therefore central to asset pricing, to market design, and to a series16of policy debates that have accompanied the transformation of these markets17over the past two decades: the growth of speculative and index-based trading18in commodities \citep{tang2012index, cheng2014financialization}, the rise of19exchange-traded funds (ETFs) as the retail gateway to commodity exposure, and20the increasing weight of central bank communication as a market-moving event21in its own right.2223A common empirical obstacle runs through these debates. Information arrives24in minutes, but much of the literature has measured its effects in days.25Daily data conflate the announcement of interest with everything else that26happens the same day, attenuate estimates through microstructure noise and27intraday reversals, and are silent about the dynamics --- amplification,28absorption, resolution of uncertainty --- that unfold in the first minutes29and hours after an information event \citep{andersen2003micro,30kothari2007econometrics}. The high-frequency literature pioneered by31\citet{andersen1998deutsche} and extended by realized-measure econometrics32\citep{andersen2001distribution, barndorff2004power, corsi2009simple} showed33that intraday data are not merely more precise: they change what can be34identified. Event windows measured in minutes isolate the causal effect of a35single release; realized variance turns latent volatility into an observable;36and the separation of continuous and jump components distinguishes gradual37information diffusion from discrete repricing.3839This dissertation applies that identification strategy to three questions40about commodities and financial futures markets. Each essay constructs a new41high-frequency dataset, each exploits a well-defined information event or42transmission channel, and each documents effects whose magnitude, direction,43or even sign would be invisible at the daily frequency. The common thread is44the same throughout: \emph{who} transmits information to prices, through45\emph{which channel}, and at \emph{what horizon}.4647\section*{Essay 1: Speculative trading and macroeconomic surprises in energy48  markets}4950The first essay addresses one of the most persistent controversies in51commodity market policy: whether speculative trading destabilizes energy52markets. The commodity price run-up of 2004--2008 spawned an influential53narrative --- the ``Masters hypothesis'' --- according to which the growing54presence of financial traders amplifies price movements and volatility, with55harmful consequences for the real economy. The theoretical literature shows56that such distortions are possible \citep{basak2016model,57goldstein2022commodity}, but the empirical evidence, largely based on daily58data and Granger-causality designs, is inconclusive \citep{fattouh2013role,59irwin2012testing}.6061The essay takes a different route. Macroeconomic announcement releases62provide sharply identified, exogenous information shocks whose surprise63component can be standardized following \citet{balduzzi2001economic}. Using645-minute futures data for two energy commodities (crude oil and natural gas)65and four metals (gold, silver, copper, and palladium) from April 2007 to66February 2024, together with a time-varying, commodity-specific measure of67speculative intensity built from the CFTC's disaggregated Commitments of68Traders data, the essay asks whether the reaction of returns, volatility, and69bid-ask spreads to macroeconomic surprises is amplified or dampened when70speculative trading is more intense. The answer is unambiguous: the71interaction between surprises and speculative intensity is consistently72stabilizing. Prices and conditional volatility react \emph{less} to surprises73when speculative positions are larger, and bid-ask spreads narrow rather than74widen. Disaggregating trader categories shows that these beneficial effects75are driven by money managers --- informed traders in the sense of76\citet{cheng2015convective} --- rather than swap dealers. For energy markets,77where volatility feeds directly into household energy costs and investment78under uncertainty, the finding that speculation dampens rather than amplifies79news-driven volatility speaks directly to the design of position limits and80to the broader regulatory debate.8182\section*{Essay 2: The iNAV and volatility transmission between commodity83  ETFs and their underlying assets}8485The second essay turns from the futures market itself to the fastest-growing86channel through which investors access commodities: exchange-traded funds. An87ETF is an equity-like wrapper around an underlying basket --- physical88bullion for gold and silver funds, futures positions for oil and natural gas89funds --- and the arbitrage mechanism that keeps the wrapper aligned with its90contents is also a conduit for volatility \citep{ben2018etfs}. Measuring that91conduit precisely requires observing the value of the basket at the same92frequency as the ETF price. The essay's central contribution is the93construction of a novel minute-level dataset of indicative net asset values94(iNAV) for four commodity ETFs (GLD, SLV, USO, UNG), validated against95official NAVs and exchange-disseminated values, over more than a decade.9697With ETF and basket volatilities observable minute by minute, the essay98measures transmission using HAR-type cascade models \citep{corsi2009simple}99extended with cross-asset terms, decomposes realized variance into continuous100and jump components, and complements the analysis with Bayesian vector101autoregressions. Three results stand out. First, the direction of102transmission mirrors the arbitrage technology: for physically backed precious103metal funds, volatility flows one way, from the basket to the ETF; for104futures-based energy funds, transmission is bidirectional and asymmetric.105Second, transmission operates primarily through jumps rather than through the106continuous component --- a channel that standard connectedness measures,107built on daily total volatility, largely obscure. Third, sampling frequency108matters quantitatively: 1-minute data reveal transmission up to twice as109large as 30-minute estimates. Beyond the substantive findings, the essay110demonstrates that the informational content of the iNAV --- a series111disseminated in real time but rarely archived or studied --- provides a112sharper image of the ETF--underlying relationship than price-based proxies.113114\section*{Essay 3: Policy tone, informational novelty, and the115  high-frequency reaction to FOMC announcements}116117The third essay studies the information events that arguably matter most for118financial futures: announcements of the Federal Open Market Committee. A119large literature measures monetary policy surprises through interest-rate120futures \citep{kuttner2001, gurkaynak2005} or through the joint reaction of121rates and equities \citep{nakamura2018}. Yet FOMC statements are texts, and122their market impact depends not only on what they imply for the policy rate123but on how they say it. The essay decomposes each statement into two semantic124dimensions: \emph{policy tone} --- the hawkish-versus-dovish orientation of125the language --- and \emph{informational novelty} --- the distance between a126statement and its predecessor. The measurement apparatus is a dual-model127ensemble of transformer language models (MiniLM and BERT) fine-tuned on128Federal Reserve communications, with reference statements selected by a129data-driven, PCA-based procedure that removes researcher discretion.130131Combining these measures with 1-minute data for seven futures contracts ---132equities, VIX, Treasuries, the dollar, crude oil, and gold --- across 148133FOMC announcements between 2008 and 2025, the essay documents a clean134division of labour between the two dimensions. Tone predicts directional135returns: a one-standard-deviation dovish shift is associated with equity136gains that accumulate to roughly twelve basis points over two hours. Novelty137predicts volatility, with a sign opposite to the naive prediction: statements138that depart more from their predecessor \emph{reduce} uncertainty, consistent139with communication acting as clarification rather than noise. The140tone--novelty interaction on VIX futures is the most robust effect in the141paper, and pre-announcement placebo tests together with five independent142inference methods support a causal reading. The two channels also operate on143different clocks --- volatility effects resolve within roughly twenty144minutes, return effects build for two hours --- a pattern inconsistent with145any single-channel account of central bank communication.146147\section*{Contributions and structure of the dissertation}148149Taken together, the three essays make three kinds of contributions. First,150they build and document new high-frequency datasets --- standardized151macroeconomic surprises matched to 5-minute commodity futures, a minute-level152iNAV panel for commodity ETFs, and a semantically decomposed corpus of FOMC153communications aligned with 1-minute futures data --- each of which has value154beyond the questions asked here. Second, they deliver identified answers to155long-standing questions: speculation stabilizes rather than destabilizes156energy futures around news; ETF--underlying volatility transmission is157jump-driven and direction-dependent; and central bank communication moves158returns through tone but volatility through novelty. Third, they carry policy159implications --- for the calibration of speculative position limits, for the160design and monitoring of commodity index products, and for the communication161strategy of central banks.162163The remainder of the dissertation is organized as follows.164Chapter~\ref{chap-chapitre1} presents the first essay, on speculative trading165and macroeconomic surprises in energy futures markets.166Chapter~\ref{chap-chapitre2} presents the second essay, on iNAV-based167volatility transmission between commodity ETFs and their underlying baskets.168Chapter~\ref{chap-chapitre3} presents the third essay, on the high-frequency169effects of the tone and novelty of FOMC statements. A general conclusion170summarizes the findings, discusses their limitations, and outlines avenues171for future research. Appendices~\ref{ch3:sec:proofs}172and~\ref{ch3:sec:app_additional} collect the mathematical proofs and173additional results of Chapter~\ref{chap-chapitre3}.174