Sentiment as a Brake, Not a Signal: A Profit-at-Risk and Placebo Attribution of Affective Features in Equity Trading Systems
Trading systems increasingly use the affective content of financial text to size positions, assuming that sentiment carries information a price-based model lacks. We ask whether that assumption survives a falsification test. Using more than three million social-media and news messages scored for sentiment, sarcasm and emotion, we compare three strategies differing only in their affective input: technical indicators alone, technical indicators plus sentiment, and a content-free placebo whose feature matches sentiment's statistical profile but carries no meaning. Each is evaluated on return relative to tail risk across a bull and a bear year. Sentiment does not improve how the system trades; it changes how much. It acts almost entirely as a brake on exposure, pulling the strategy out of a rising market and cutting its participation in a falling one. The placebo is decisive: compared on common trading days, which holds the exposure difference aside, a meaningless feature performs as well as real sentiment, with no significant daily difference in either regime. What looks like skill is reduced exposure, not better timing. Any textual feature proposed for a trading system should therefore be tested against a distribution-matched placebo, compared at equal exposure, and judged on profit-at-risk rather than raw return.
Authors
- Younès Lahrichi (ORCID: https://orcid.org/0009-0007-9417-1669)
- Saïd Achchab (ORCID: https://orcid.org/0000-0001-5752-8699)
- Abdelkhaleq El Haddad (ORCID: https://orcid.org/0009-0009-8239-1044)
Institutions
- Mohammed V University (MA)
- Institut Supérieur de Commerce et d'Administration des Entreprises (MA)
- ISCA Technologies (United States) (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-09-24
- DOI
- https://doi.org/10.5281/zenodo.22936901
- Primary Topic
- Financial Markets and Investment Strategies
- Type
- preprint