Introduction

AlphaGlass is an inherently interpretable machine-learning framework for constructing portfolios that directly optimise investment objectives. Unlike traditional approaches that sequentially predict returns before forming portfolios, AlphaGlass maps firm characteristics into additive signals with sparse pairwise interactions and converts them into long-short portfolios through a differentiable rank-and-mask layer, optimising objectives such as the Sharpe ratio or mean-variance utility end-to-end.

Methodology

AlphaGlass uses a Generalised Additive Model architecture, representing each univariate and pairwise effect as a small feed-forward neural network. Training proceeds in three stages: main effect learning, interaction learning, and joint fine-tuning. Hard portfolio cutoffs are replaced with smooth differentiable soft weights via pairwise logistic comparisons, enabling gradient-based Sharpe ratio optimisation throughout.

Data

Monthly return data are sourced from CRSP and merged with 91 firm characteristics, covering NYSE, NASDAQ, and AMEX stocks. After removing observations with missing data, the dataset comprises 1,656,664 observations over January 2000 to December 2021 (p.9). Data are split into a 10-year training window, 2-year validation window, and 10-year test period.

Empirical results

AlphaGlass achieves a monthly out-of-sample Sharpe ratio of 0.47 (1.62 p.a.) (p.11), outperforming benchmarks: Random Forest (0.14), Neural Network (0.34), and EBM (0.30) (p.11). The in-sample Sharpe ratio is 0.54 (p.11). Out-of-sample Sharpe ratios range from 0.25 to 0.4 for larger size quintiles (p.13), indicating results are not driven by small stocks. When small stocks are excluded, AlphaGlass still achieves a Sharpe ratio of 0.43, compared with 0.12 and 0.13 for the neural net and EBM respectively (p.13).

The four most important univariate characteristics are industry sales concentration (herf), operating leverage dynamics (pchsale_pchxsga), 12-month momentum (mom12m), and industry-adjusted change in employees (chempia) (p.14). The acc × hire interaction ranks among the top 15 effects. When removed individually, maxret or mom12m each reduce the out-of-sample Sharpe ratio from 0.47 to 0.41 (p.17).

Economic insights from interpretability and portfolio properties

Shape functions reveal clear economic patterns: low industry concentration and high momentum increase long signals, while high concentration and low momentum reduce them. Signal decomposition shows herf is the dominant driver of long signals, with S+ = 2.42 (p.25), while the sin stock indicator dominates short signals with a magnitude of 5.60 (p.27). The acc × hire interaction indicates hiring is most informative when accruals are low (p.28).

Mean-variance preferences

AlphaGlass delivers the strongest out-of-sample performance under mean-variance utility at both risk-aversion levels tested. For γ = 1, out-of-sample utility is 1.12, versus 0.99 (EBM), 0.86 (NN), and 0.53 (RF). For γ = 3, AlphaGlass leads at 1.23, followed by EBM (0.87), NN (0.79), and RF (0.34) (p.30).

Theoretical properties

Three theorems establish that: the differentiable rank-and-mask layer faithfully approximates hard quantile sorts; maximising the in-sample Sharpe ratio is argmax-consistent for the population Sharpe ratio; and the out-of-sample Sharpe ratio is near-best-in-class with high probability, with the gap controlled by model complexity and sample length (pp.31–35).

Monte Carlo simulation

Across 100 replications, AlphaGlass attains a mean annualised Sharpe ratio of 1.774 at T = 120 months and 1.970 at T = 240 months, compared with 1.568 and 1.755 (EBM), 1.449 and 1.557 (NN), and 1.411 and 1.438 (RF) (p.38). AlphaGlass captures 74.3% of the ranking-oracle Sharpe ratio at T = 120 and 83.4% at T = 240 (p.38).