Supervised learning for returns is the practice of training a model on labeled examples — a matrix of predictive features (firm characteristics, technical signals, macro variables) paired with a known future-return target — so it learns a function that maps features to expected returns on data it has never seen. The core tension is that financial returns are mostly noise: the signal-to-noise ratio is far lower than in canonical ML domains like image recognition, the data-generating process is non-stationary, and the flexibility that makes ML powerful is exactly what makes it overfit. The discipline is therefore less about exotic architectures than about honest validation and managing the gap between in-sample fit and out-of-sample reality.

How it's framed and trained

Two problem formulations dominate:

• Regression — predict the continuous future return (e.g. next-month excess return). Evaluated by out-of-sample R², mean squared error, and downstream portfolio performance.
• Classification — predict a discrete label (up/down, or a multi-class outcome). The naive version uses a fixed-horizon sign label; the more rigorous version is López de Prado's triple-barrier method, which labels each observation +1/−1/0 depending on whether a profit-taking, stop-loss, or time barrier is hit first — making the label reflect path and volatility, not just an arbitrary horizon endpoint (López de Prado, Advances in Financial Machine Learning, 2018).

The model menu, roughly in order of flexibility: penalized linear models (LASSO, ridge, elastic net), kernel methods (SVM/SVR), tree ensembles (random forest, gradient-boosted trees / XGBoost, LightGBM), and neural networks (MLPs, and for sequence data LSTMs/temporal CNNs). Training minimizes a loss (MSE for regression, cross-entropy for classification), with regularization (L1/L2, dropout, tree depth/learning-rate limits) as the primary defense against overfitting. Inputs are typically cross-sectionally standardized or rank-transformed because raw characteristics have unstable, fat-tailed scales.

The single most consequential design choice is validation. Standard k-fold cross-validation is invalid here because it trains on the future and because overlapping labels leak information across folds. The accepted fix is a walk-forward / expanding-window split plus López de Prado's purging (drop training samples whose label window overlaps the test set) and embargoing (drop training samples immediately after the test set). Getting this wrong is the difference between a real edge and a fantasy.

How it's used in practice

In quant equity, the canonical application is the cross-section: pool thousands of stocks each month, predict the relative expected return, then form a long-short portfolio (long the top decile of predictions, short the bottom) and weight by conviction or by an optimizer. The model's job is ranking, so rank-correlation and decile-spread matter more than absolute R². At higher frequencies, classifiers feed execution and signal-gating systems.

A pattern worth isolating is meta-labeling: a primary model (or a simple rule) decides side (long/short), and a secondary supervised model — trained on triple-barrier labels — predicts only the probability the primary signal is correct, sizing or vetoing the bet. This separates "which way" from "how much / whether to act," improves precision, and tends to lift the strategy's F1 / risk-adjusted return more reliably than trying to predict direction end-to-end (López de Prado, 2018; corroborated in practitioner studies such as Hudson & Thames).

Adoption, debate & evidence

ML for returns moved from fringe to mainstream after Gu, Kelly & Xiu (2020), "Empirical Asset Pricing via Machine Learning" (Review of Financial Studies), the field's most-cited benchmark. Across ~30,000 U.S. stocks, 1957–2016, they found:

• Flexible models beat linear ones. Their neural networks produced monthly out-of-sample R² in the ~0.33%–0.40% range for individual stocks (the best NN near the top of that range, and higher at the portfolio level) — versus essentially zero or negative for OLS using the full unconstrained predictor set (the constrained OLS-3 benchmark was ~0.16%/month). That number sounds tiny, and it is; it is also economically large when compounded across a broad cross-section.
• A value-weighted long-short decile-spread portfolio sorted on the neural-net forecasts earned an annualized out-of-sample Sharpe ratio of ~1.35 (the equal-weighted version ~2.45; the analogous OLS strategy ~0.61 value-weighted) — far above the linear benchmarks.
• All methods agreed on the same dominant signals: variants of momentum, liquidity, and volatility. The gain came less from new predictors than from modeling non-linearities and interactions among known ones.

These results have been broadly replicated, but the honest landscape is contested on three fronts. First, net-of-cost: much of the headline Sharpe lives in small, illiquid stocks where the long-short spread is widest; transaction costs and shorting frictions erode a large share of paper alpha, and the surviving edge is debated. Second, decay: predictability documented in academic samples often weakens after publication and as more capital chases it. Third, false discoveries — López de Prado's "10 Reasons Most ML Funds Fail" argues that the flexibility of ML, combined with repeated backtesting on the same data, manufactures spurious strategies; he insists every backtest be reported with all trials run (to deflate for multiple testing via tools like the Deflated Sharpe Ratio). The measured reality: real, but small, fragile, and easy to fake.

Strengths & limitations

Works when: the cross-section is wide (many assets average out idiosyncratic noise), features encode genuine non-linear interactions, the horizon is monthly rather than tick-level (more stable), validation is purged/embargoed, and the output is used for relative ranking inside a diversified, cost-aware portfolio.

Fails when: applied to a single asset's price level (near-random-walk, almost no signal), trained with leakage (lookahead features, naive CV, survivorship-biased universes), evaluated only in-sample, or pushed for high accuracy on a noisy target instead of small-but-stable edge. Regime change silently invalidates a model trained on a prior regime.

The #1 misuse: treating it like a Kaggle problem — chasing test-set accuracy on a fixed split, iterating architecture against the same out-of-sample window until it "works." That is backtest overfitting; the apparent skill does not survive deployment. The second-most-common error is ignoring that the relevant metric is risk-adjusted net return of the resulting portfolio, not the model's R² or accuracy.

Sources

• Gu, S., Kelly, B., Xiu, D. (2020). "Empirical Asset Pricing via Machine Learning." Review of Financial Studies 33(5): 2223–2273. NBER w25398; PDF: https://dachxiu.chicagobooth.edu/download/ML.pdf — headline R² (~0.33–0.40%/month NN, ~0.16% OLS-3), dominant predictors, value-weighted long-short Sharpe ~1.35 (equal-weighted ~2.45).
• López de Prado, M. (2018). Advances in Financial Machine Learning, Wiley — triple-barrier labeling, meta-labeling, purged/embargoed CV.
• López de Prado, M. (2018). "The 10 Reasons Most Machine Learning Funds Fail." The Journal of Portfolio Management 44(6): 120–133 (also GARP whitepaper: https://www.garp.org/hubfs/Whitepapers/a1Z1W0000054x6lUAA.pdf) — false discoveries, multiple-testing deflation.
• Bailey, Borwein, López de Prado, Zhu — backtest overfitting / Deflated Sharpe: https://sdm.lbl.gov/oapapers/ssrn-id2507040-bailey.pdf
• Hudson & Thames, "Does Meta Labeling Add to Signal Efficacy?" — independent meta-labeling evaluation.
• Survey context: Frontiers in AI (2025), "AI in financial market prediction"; arXiv 1906.07786 survey of ML in financial forecasting.

Disputes flagged: the exact Sharpe/R² figures vary by specification and sample window — treat the cited values as the widely-reported headline, not a single canonical number. Net-of-cost profitability of cross-sectional ML strategies is genuinely contested; post-publication decay is documented but its magnitude is debated.