ML for alpha is easy to overfit and hard to productionize. This list links to primary sources that stress validation, leakage, and deployment constraints. Read the foundations first, then large-scale empirical work, then nonlinear models, formula alpha, and how to evaluate what you mine.

Foundations and methodology

Start here before tuning hyperparameters. Finance-specific cross-validation and multiple-testing bias are where most ML alpha projects die.

Large-scale return prediction

Canonical empirical work comparing ML methods on equity cross-sections. Useful for understanding what actually works at scale versus in toy backtests.

Deep learning and nonlinear signals

Nonlinear factor models, compression, and learning portfolio weights from latent structure — not just point forecasts on raw features.

An Information Bottleneck Asset Pricing Model

Information-bottleneck constraints compress noisy inputs while preserving pricing-relevant signal in deep factor models. Addresses overfitting in low signal-to-noise equity data.

arXiv

Deep Reinforcement Learning in Factor Investment

CAFPO: conditional autoencoder factors feed a DRL agent for low-frequency long-short weights. Handles changing investable universes via factor compression instead of raw stock panels.

arXiv

Formula alpha and LLM evaluation

Industry-style formula libraries, grammar-guided search, and scoring beyond a single backtest Sharpe — including when LLMs should act as researchers, not stochastic trading agents.

101 Formulaic Alphas

Industry-side alpha construction at scale. Good reality check on signal correlation and turnover.

arXiv

FactorBench: A Portfolio-Aware Benchmark for Automated Factor Mining

~5,000 factors from nine mining methods (GP, RL, diffusion, LLM agents) on five index universes under one execution contract. In-sample IC often fails after style neutralization; LLM pools overlap heavily with Alpha101; dollar-neutral net Sharpe beats long-only headlines. No paradigm wins on all dimensions. Code

arXiv

Before you ship

Walk-forward validation, purged k-fold, and explicit transaction cost models are non-optional. If Sharpe only exists in one in-sample window, fix the pipeline before adding features.

Also follow ongoing preprints at arXiv q-fin and journal pages via DOI.