Introducing From Factor Models to Deep Learning: Machine Learning in Reshaping Empirical Asset Pricing for after-work quants

Takeaway. Treat this survey as a map, not a model: traditional factor / IPCA stories sit next to flexible g(characteristics) (trees, nets), spatio-temporal GNNs, RL weights, plus honest sections on overfitting, missing data, and explainability.

From Factor Models to Deep Learning (Junyi Ye, Bhaskar Goswami, Jingyi Gu, Ajim Uddin, Guiling Wang, 2024; arXiv:2403.06779) is a study-path paper for after-work quants who already met Gu–Kelly–Xiu.

Fit for learners: Survey map from factors to DL/RL pricing — reading guide to plan your SoloQuant study path.

This Hub note is a plain-language intro. Not investment advice.

The problem

CAPM / Fama–French linear β′f cannot ingest a factor zoo, nonlinear return dynamics, or text/image features. Hand-picking factors assumes we already know the cross-section. PCA/IPCA help, but the functional form of expected excess return g(x) is still too rigid for many modern datasets.

The idea / core intuition

• Write y = g(x; θ) + ε and let ML search g (linear → trees → deep nets).
• Split tasks: return prediction, direction / ranking, portfolio weights (supervised vs RL).
• Add spatial structure (graphs of firms) on top of temporal LSTMs/transformers.
• Later chapters: dimension reduction, missing characteristics, alt-data, contrastive / mixture-of-experts tricks—and why regulators still want interpretable risk.

A literature GPS—not a trading system.

Why after-work Python learners care

You do not need to implement every cited net. You need a reading order: factors → ML bake-off → one deep pricing paper → one portfolio/RL paper. The survey’s metric table (RMSE vs Sharpe vs NDCG) is a reminder that loss ≠ PnL.

Tiny evening exercise: write a four-row table: task (predict return / classify up-down / rank / allocate) vs a metric you would refuse to skip. Learning only.

One-evening reading checklist

1. Contrast linear factor models with “learn g(x)”.
2. Name two ML families used for cross-sectional expected returns.
3. Say how RL portfolio papers differ from predict-then-Markowitz.
4. List two adoption challenges (overfitting, explainability, data access).

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. The survey summarises academic ML asset-pricing research—not a recommendation to allocate to any cited model.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not a “AI will replace factors” pitch

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—functions and dictionaries first; then g(x) is just another callable.