\chapter*{Conclusion}           % ne pas numéroter
\label{chap-conclusion}         % étiquette pour renvois
\phantomsection\addcontentsline{toc}{chapter}{\nameref{chap-conclusion}} % inclure dans TdM

% BROUILLON À RÉVISER PAR L'AUTEUR — conclusion générale de la thèse.

This dissertation set out to understand how information is transmitted to
prices and volatility in commodities and financial futures markets, using the
identification power of high-frequency data. Three essays examined three
distinct transmission mechanisms: the interaction between macroeconomic news
and speculative positioning in energy futures; the arbitrage-driven
propagation of volatility between commodity ETFs and their underlying
baskets; and the semantic channels through which central bank communication
moves asset prices. This concluding chapter summarizes the findings, draws
out what they imply jointly, acknowledges limitations, and sketches avenues
for future research.

\section*{Summary of contributions}

The first essay showed that speculative trading in energy and commodity
futures markets acts as a stabilizing force around macroeconomic
announcements. Using 5-minute data over 2007--2024 and a time-varying measure
of speculative intensity built from disaggregated CFTC positions, it found
that higher speculative activity dampens the reaction of returns and
conditional volatility to standardized macroeconomic surprises and narrows
bid-ask spreads, with the effects concentrated among money managers rather
than swap dealers, and stronger for procyclical energy commodities than for
safe-haven gold. The essay thus provides sharply identified evidence against
the view that speculation amplifies news-driven volatility, and speaks
directly to the design of position limits.

The second essay built the first long-sample, minute-level dataset of
indicative net asset values for commodity ETFs and used it to measure
volatility transmission between funds and their underlying assets. It found
that the direction of transmission reflects the arbitrage technology ---
one-way, from basket to fund, for physically backed precious metals;
two-way and asymmetric for futures-based energy funds --- that transmission
operates predominantly through jumps rather than the continuous component of
volatility, and that 1-minute data reveal transmission up to twice as large
as 30-minute estimates. The iNAV emerges as a measurement instrument of
independent value for the microstructure of index-linked products.

The third essay decomposed FOMC statements into policy tone and informational
novelty using an ensemble of transformer language models with data-driven
reference selection, and traced their effects through 1-minute futures data
across 148 announcements. Tone predicts directional returns, building over
two hours; novelty predicts volatility, resolving within roughly twenty
minutes and with a clarifying rather than noise-inducing effect; and their
interaction on VIX futures is the single most robust result. The essay shows
that the informational content of central bank communication is
multidimensional, and that its dimensions travel through different economic
channels at different speeds.

Across the three essays, a consistent methodological lesson emerges: the
minute-level resolution is not a refinement but a prerequisite. The damping
effect of speculation, the jump channel of ETF volatility transmission, and
the divergent dynamics of tone and novelty are all features of the first
minutes and hours after an information event; at the daily frequency they are
attenuated, conflated, or invisible.

\section*{Limitations}

Several limitations qualify these results and delimit their scope. First,
the measures of trader positioning in the first essay derive from weekly CFTC
reports, so that speculative intensity is observed at a coarser frequency
than the market reactions it conditions; the identification rests on the
persistence of positioning rather than on its intraday variation. Second, the
iNAV series of the second essay are constructed from disseminated and
reconstructed data whose quality, while validated against official NAVs,
cannot be audited tick by tick over the whole sample; and the analysis covers
four large single-commodity funds, leaving open how the results extend to
broad-basket or leveraged products. Third, the semantic measures of the third
essay are estimated from a finite corpus of Federal Reserve communications;
although reference selection is data-driven and robustness was assessed with
multiple inference methods, language-model-based measures inevitably embed
modelling choices, and the analysis concerns statements rather than the full
communication apparatus (minutes, press conferences, speeches). Finally, all
three essays are about the United States markets; the external validity of
the findings for other trading venues and regulatory environments remains to
be established.

\section*{Avenues for future research}

The results open several paths. On the policy side, the trader-level
heterogeneity documented in the first essay --- stabilizing money managers,
amplifying swap dealers --- invites a finer analysis of position-limit design
and of the role of speculative capital in financing the energy transition,
where volatility dampening lowers the real-option value of delaying
investment. On the market-structure side, the jump-based transmission channel
identified in the second essay suggests extending the iNAV apparatus to
broad-basket, leveraged, and fixed-income ETFs, and studying how arbitrage
frictions --- creation/redemption costs, settlement technology, market-maker
inventory --- shape the propagation of discontinuous risk. On the
communication side, the tone/novelty decomposition of the third essay can be
carried to press conferences and minutes, to other central banks, and to the
cross-section of individual assets, where the divergent horizons of the
return and volatility channels may help separate expectations formation from
uncertainty resolution. More broadly, the combination of language-model
measurement with high-frequency identification --- used here for monetary
policy --- applies to any recurring, text-borne information event: earnings
calls, regulatory releases, geopolitical announcements.

High-frequency data have turned questions that were once matters of narrative
--- does speculation destabilize? do ETFs transmit shocks? do words move
markets? --- into questions of measurement. The three essays of this
dissertation are a step in that direction for commodities and financial
futures markets.