spb/phd_thesis Public

1\chapter*{Conclusion}           % ne pas numéroter2\label{chap-conclusion}         % étiquette pour renvois3\phantomsection\addcontentsline{toc}{chapter}{\nameref{chap-conclusion}} % inclure dans TdM45% BROUILLON À RÉVISER PAR L'AUTEUR — conclusion générale de la thèse.67This dissertation set out to understand how information is transmitted to8prices and volatility in commodities and financial futures markets, using the9identification power of high-frequency data. Three essays examined three10distinct transmission mechanisms: the interaction between macroeconomic news11and speculative positioning in energy futures; the arbitrage-driven12propagation of volatility between commodity ETFs and their underlying13baskets; and the semantic channels through which central bank communication14moves asset prices. This concluding chapter summarizes the findings, draws15out what they imply jointly, acknowledges limitations, and sketches avenues16for future research.1718\section*{Summary of contributions}1920The first essay showed that speculative trading in energy and commodity21futures markets acts as a stabilizing force around macroeconomic22announcements. Using 5-minute data over 2007--2024 and a time-varying measure23of speculative intensity built from disaggregated CFTC positions, it found24that higher speculative activity dampens the reaction of returns and25conditional volatility to standardized macroeconomic surprises and narrows26bid-ask spreads, with the effects concentrated among money managers rather27than swap dealers, and stronger for procyclical energy commodities than for28safe-haven gold. The essay thus provides sharply identified evidence against29the view that speculation amplifies news-driven volatility, and speaks30directly to the design of position limits.3132The second essay built the first long-sample, minute-level dataset of33indicative net asset values for commodity ETFs and used it to measure34volatility transmission between funds and their underlying assets. It found35that the direction of transmission reflects the arbitrage technology ---36one-way, from basket to fund, for physically backed precious metals;37two-way and asymmetric for futures-based energy funds --- that transmission38operates predominantly through jumps rather than the continuous component of39volatility, and that 1-minute data reveal transmission up to twice as large40as 30-minute estimates. The iNAV emerges as a measurement instrument of41independent value for the microstructure of index-linked products.4243The third essay decomposed FOMC statements into policy tone and informational44novelty using an ensemble of transformer language models with data-driven45reference selection, and traced their effects through 1-minute futures data46across 148 announcements. Tone predicts directional returns, building over47two hours; novelty predicts volatility, resolving within roughly twenty48minutes and with a clarifying rather than noise-inducing effect; and their49interaction on VIX futures is the single most robust result. The essay shows50that the informational content of central bank communication is51multidimensional, and that its dimensions travel through different economic52channels at different speeds.5354Across the three essays, a consistent methodological lesson emerges: the55minute-level resolution is not a refinement but a prerequisite. The damping56effect of speculation, the jump channel of ETF volatility transmission, and57the divergent dynamics of tone and novelty are all features of the first58minutes and hours after an information event; at the daily frequency they are59attenuated, conflated, or invisible.6061\section*{Limitations}6263Several limitations qualify these results and delimit their scope. First,64the measures of trader positioning in the first essay derive from weekly CFTC65reports, so that speculative intensity is observed at a coarser frequency66than the market reactions it conditions; the identification rests on the67persistence of positioning rather than on its intraday variation. Second, the68iNAV series of the second essay are constructed from disseminated and69reconstructed data whose quality, while validated against official NAVs,70cannot be audited tick by tick over the whole sample; and the analysis covers71four large single-commodity funds, leaving open how the results extend to72broad-basket or leveraged products. Third, the semantic measures of the third73essay are estimated from a finite corpus of Federal Reserve communications;74although reference selection is data-driven and robustness was assessed with75multiple inference methods, language-model-based measures inevitably embed76modelling choices, and the analysis concerns statements rather than the full77communication apparatus (minutes, press conferences, speeches). Finally, all78three essays are about the United States markets; the external validity of79the findings for other trading venues and regulatory environments remains to80be established.8182\section*{Avenues for future research}8384The results open several paths. On the policy side, the trader-level85heterogeneity documented in the first essay --- stabilizing money managers,86amplifying swap dealers --- invites a finer analysis of position-limit design87and of the role of speculative capital in financing the energy transition,88where volatility dampening lowers the real-option value of delaying89investment. On the market-structure side, the jump-based transmission channel90identified in the second essay suggests extending the iNAV apparatus to91broad-basket, leveraged, and fixed-income ETFs, and studying how arbitrage92frictions --- creation/redemption costs, settlement technology, market-maker93inventory --- shape the propagation of discontinuous risk. On the94communication side, the tone/novelty decomposition of the third essay can be95carried to press conferences and minutes, to other central banks, and to the96cross-section of individual assets, where the divergent horizons of the97return and volatility channels may help separate expectations formation from98uncertainty resolution. More broadly, the combination of language-model99measurement with high-frequency identification --- used here for monetary100policy --- applies to any recurring, text-borne information event: earnings101calls, regulatory releases, geopolitical announcements.102103High-frequency data have turned questions that were once matters of narrative104--- does speculation destabilize? do ETFs transmit shocks? do words move105markets? --- into questions of measurement. The three essays of this106dissertation are a step in that direction for commodities and financial107futures markets.108