A central unresolved problem in systematic equity signal generation is regimenon-stationarity: machine learning models trained on historical price data routinelyexhibit strong in-sample performance but near-zero out-of-sample precision whenthe market environment shifts. This paper presents Olympia, a framework thataddresses this failure mode directly by conditioning individual security probabilityscores on a five-feature broad market regime layer derived from the S&P 500 proxyindex. Adding these regime features increased top-50 out-of-sample precision fromapproximately 12% to 62%—a five-fold improvement in a single model iteration—while models trained without regime features continued to produce near-zero top-10precision in the 2025 evaluation period.The full framework combines gradient-boosted decision trees with isotonic probability calibration and physics-inspired monotone feature constraints. Evaluatedusing a rigorous walk-forward backtesting methodology across 11,845 US equitiesand approximately 3.46 million (ticker, date) observations, Olympia achieved approximately 90% top-10 out-of-sample precision across the evaluated test eras underfavorable market regime conditions (2023–2025). The system operates as a fullyautomated weekly scanning pipeline capable of scoring the entire US equity universein approximately 22 minutes on commodity hardware.
Michael Rupert (Sun,) studied this question.