Machine Learning in Markets: What Survives Contact with Financial Data

Standard ML assumes the world that generated your training data will keep existing. Markets violate that assumption on purpose: they adapt, and profitable patterns attract the capital that erases them. This lesson is about which ML ideas survive that environment — the judgment layer quant-research interviews probe hardest.

What makes financial data hostile

• Tiny signal-to-noise. A very good daily equity signal has an information coefficient (correlation with next-period returns) around 0.02–0.05. Compare handwriting recognition, where the "signal" is nearly the whole variance. Consequence: variance reduction dominates — simple models, heavy regularization, massive cross-sectional breadth to average the noise away.
• Non-stationarity. The data-generating process drifts (regimes, structure changes) and reacts to being learned — alpha decays as others find it. Models need retraining schedules and, more importantly, humility about their shelf life.
• Effectively small samples. A century of monthly returns is ~1,200 points — and not independent ones. Cross-sectional strategies survive by trading thousands of assets, converting a time-series problem (few observations) into a panel problem (many).

The professional's summary: in finance, the binding constraint is statistical, not algorithmic. Fancier architectures don't manufacture signal that isn't there.

The model menu, market-adjusted

• Linear/regularized regression — still the workhorse. Interpretable, stable, hedgeable (a linear signal has clean factor exposures). The default until proven insufficient.
• Decision trees → random forests / gradient boosting. Trees capture interactions and nonlinearities without feature engineering and are indifferent to monotone transformations. Single trees are high-variance; ensembles fix that by averaging — bagging (many trees on resampled data, averaged: variance ↓) or boosting (trees fitted sequentially to residuals: bias ↓, overfits if unrestrained). Gradient boosting (XGBoost-style) wins most tabular-data problems, including many alpha problems — knowing why (tabular data, interactions, limited data) beats name-dropping it.
• Neural networks. Dominant where features must be learned from raw structure — text (news, filings), order-book sequences, images. For a 50-feature tabular alpha problem they mostly offer extra variance and worse interpretability. "Where would you use a neural net in trading?" wants this differentiated answer, not enthusiasm.

Features beat models

The consistent empirical finding: a mediocre model on well-constructed features beats a sophisticated model on raw data. Market-standard feature discipline:

• Normalize cross-sectionally (rank or z-score each day) so the model learns relative attractiveness, not market direction.
• Winsorize outliers; financial tails will otherwise dominate the fit.
• Respect information timing ruthlessly: a feature must use only data available at prediction time. Point-in-time databases exist because restated earnings and index rebalances quietly leak the future — the most common fatal backtest bug.

Judgment questions, with answers

• "Your deep model beats linear by 2% accuracy in backtest. Ship it?" — Ask what a 2% edge costs in interpretability, stability across regimes, and turnover/costs first; verify the comparison was walk-forward and the deep model wasn't tuned harder (more hyperparameter trials = more selection bias). Often the honest answer is "run both small, live, and let production data decide."
• "Why does everyone use gradient boosting for alpha?" — Tabular data, strong interactions, limited samples, and decades of tooling; and it degrades gracefully under regularization (depth, learning rate, subsampling all throttle variance).
• "How do you know your model stopped working, versus a drawdown?" — Pre-registered monitoring: live IC vs. backtest IC, turnover drift, exposure drift. Decide the retirement criteria before deployment, because during a drawdown you'll rationalize anything.

The meta-lesson interviewers want: in markets, ML is 20% modeling and 80% not fooling yourself — the previous lesson's validation discipline, applied under adversarial conditions.