Introducing Empirical Asset Pricing via Machine Learning for after-work quants

Takeaway. Gu–Kelly–Xiu run a fair bake-off of machine-learning tools (penalised linear, trees, neural nets, …) for expected returns in the cross-section—showing which nonlinearities help, not crowning a single magic algorithm.

Empirical Asset Pricing via Machine Learning (Shihao Gu, Bryan Kelly, Dacheng Xiu, 2019; NBER w25398) is the after-work ML-for-the-cross-section landmark.

Fit for learners: The cross-section ML bake-off (trees vs nets) every after-work quant should reproduce on CRSP-style features.

This Hub note is a plain-language intro. Not investment advice. Pair with Harvey–Liu–Zhu so you do not treat a leaderboard as a discovery factory.

The problem

Hundreds of characteristics, contested linear factors, and papers that each champion one model. Without a common sample, features, and validation, “our net beats OLS” is not science. Also: predictive R² can be tiny yet still map to sorted-portfolio spreads—or vanish after costs.

The idea / core intuition

• Hold the prediction problem fixed: next-period returns from a large characteristic set.
• Train many ML families under comparable recursive / OOS designs.
• Report both statistical fit and economic sorts; trees often capture interactions OLS misses.
• Complexity is a dial: more flexible models need regularisation and honest splits.

Research bake-off—not a recommended factor ETF.

Why after-work Python learners care

This is the curriculum paper for sklearn on a panel: Pipeline, time-safe CV, and “compare Ridge vs random forest vs tiny MLP” before you write custom PyTorch. Reproducing the comparison design matters more than copying their best chart.

Tiny evening exercise: fake 50 names × 36 months × 8 features; lag features; walk-forward the last 12 months; fit Ridge vs a small decision tree; print OOS R² and a long–short of top vs bottom predicted quintiles on the toy panel—learning only.

One-evening reading checklist

1. Name two model families in the bake-off.
2. Explain why a common validation design beats one-off horse races.
3. Note the difference between R² and sorted-portfolio evaluation.
4. Write one cost / microstructure caveat before treating spreads as cash.

Download PDF

Download PDF

Landing: NBER w25398

Important disclaimer

Not investment advice. This Hub note is educational only. Gu–Kelly–Xiu is academic research—not a recommendation to trade machine-learning expected-return ranks.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not a “AI beats the market” headline

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—functions and files first; then scikit-learn bake-offs are readable instead of intimidating.