{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,19]],"date-time":"2026-10-19T13:40:26Z","timestamp":1792417226692,"version":"4.4.0"},"posted":{"date-parts":[[2026]]},"group-title":"SSRN","reference-count":0,"publisher":"Elsevier BV","license":[{"start":{"date-parts":[[2026,1,1]],"date-time":"2026-01-01T00:00:00Z","timestamp":1767225600000},"content-version":"unspecified","delay-in-days":0,"URL":"https:\/\/www.uspto.gov\/ip-policy\/copyright-policy\/copyright-basics"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"abstract":"<jats:p>Simple price-based equity strategies often fail honest validation, and the ones that pass can depend on a single stock. This paper asks whether machine learning does better under strict rules. I add cross-asset indicators (the S&amp;amp;P 500 and its breadth, gold, oil, Treasury yields, credit, the dollar and the VIX term structure) and 240,000 news headlines scored with FinBERT to the usual price features, and I test five hypotheses that were written down before any model was trained. Machine learning was expected to improve volatility forecasts (H1) and drawdown prediction (H2), to filter a simple trend-following rule (H3), to rank stocks by future return (H4), and to extract information from news (H5). Models were built on 2020-2024 with a purged walk-forward, checked with negative controls and a Deflated Sharpe Ratio over 63 logged configurations, and then run once on new data from January 2025 to September 2026. None of the five hypotheses survived the holdout. Two of them had passed their rules in the development period and failed afterwards. The trend rule itself also fell apart: a Sharpe ratio of 0.16 against 0.83 for the index once NVIDIA's run ended. One result held up. During development, an ablation suggested that a gradient-boosted model using only the S&amp;amp;P 500's own price history forecast its volatility better than the standard HAR model. I committed to testing that single idea before opening the holdout, and it held: QLIKE loss of 0.80 against 1.05, Diebold-Mariano p = 0.04. Trading on the better forecast, however, did not produce a gain I can distinguish from chance, and adding cross-asset or news inputs never helped. In this setting, the most I can say for machine learning is that it forecast risk from the asset's own past a little better than a linear model.<\/jats:p>","DOI":"10.2139\/ssrn.7623819","type":"posted-content","created":{"date-parts":[[2026,10,19]],"date-time":"2026-10-19T12:45:56Z","timestamp":1792413956000},"source":"Crossref","is-referenced-by-count":0,"title":["Where Machine Learning Helps: Risk Timing, Trend Filtering, and an Untouched 2025-26 Holdout"],"prefix":"10.2139","author":[{"ORCID":"https:\/\/orcid.org\/0009-0005-5608-2943","authenticated-orcid":true,"given":"Prashant","family":"Malan","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,10,19]],"date-time":"2026-10-19T12:45:56Z","timestamp":1792413956000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7623819"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7623819","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}