FA590 Statistical Learning · Stevens Institute of Technology · Fall 2025

Which model ranks stocks best?

Five models, 890,592 stock-months and a strict time split. Every linear model and the neural net did worse than predicting the average return, yet most of them still ranked stocks usefully. LightGBM was the only model with a positive out-of-sample R², and its top-100 portfolio had the best risk-adjusted return.

See the modelsWhat drives returns

Information coefficient, test set

Spearman rank correlation, 2018–2021
Author
Zhenhan Chen
Data
890,592 monthly observations, 2009–2021 · 94 firm characteristics + 8 macro variables
Split
Train Jan 2009 – Dec 2017 · test Jan 2018 – Nov 2021, no shuffling
Models
OLS · Ridge · Lasso · LightGBM · MLP (32-16)
01

Accuracy versus ranking

R² asks how close the predicted return is; IC asks whether the order of stocks is right. For picking stocks, order is what counts.

Data table

The neural net is the clearest case of overfitting: 15.1% R² in training, −21.3% on the test years. Ridge and Lasso barely move OLS, which suggests the problem is not too many coefficients but a relationship that is not linear.

02

From predictions to a portfolio

Each month: rank every stock by predicted return, buy the top 100 in equal weights, hold one month.

Annualized Sharpe ratio

top-100 equal-weight, test period, before costs

These returns are gross of trading costs and come from an equal-weighted basket that tilts toward small stocks, so the absolute level (about 4.4% a month) overstates what could be earned. The comparison between models is the useful part: LightGBM's edge comes from lower volatility (12.4% vs. about 13.4%), not from a higher average return.

03

What the trees used

LightGBM split gain for the 20 most important of the 50 selected features. Market-wide variables are highlighted.

Macro variableFirm characteristic
macro_ep aggregate earnings-to-pricemacro_tms term spread macro_svar stock market variancemacro_dp aggregate dividend-to-price macro_dfy default yield spreadmacro_ntis net equity issuance mom12m / mom6m / mom1m past 12, 6 and 1-month returnbaspread bid-ask spread

The top three are all macro variables: the trees mostly learned when the cross-section of returns behaves differently, then used momentum and trading-cost characteristics to sort stocks within those regimes. That interaction is exactly what a linear model cannot represent.

04

Limits

What this result does not show.

The best model explains 1.45% of return variance. The training window (2009–2017) contains no crash comparable to 2008, so behaviour in a crisis is untested. Features were screened by their correlation with returns; running that screen on the training years only rules out any look-ahead. Transaction costs are ignored. Next steps would be expanding-window validation, a cost model and ensembles.