Stock selection with macro factors: the case for simple neural networks

Introduction

The intersection of machine learning and macro-quantamental investing offers a promising frontier for systematic equity strategies. In this paper, we investigate whether simple feedforward neural networks can outperform linear models when combining macroeconomic factors for stock selection.

Traditional factor models rely on linear combinations of signals such as value, momentum, and quality. However, the relationship between macroeconomic states and equity returns is inherently nonlinear. Regime changes, policy shifts, and structural breaks in the global economy all suggest that adaptive, nonlinear models may capture return predictability that linear approaches miss.

Data and methodology

We use the JPMaQS dataset, which provides point-in-time macroeconomic indicators across 40+ countries. Our feature set includes growth expectations, inflation differentials, external balances, and monetary policy stance indicators, all timestamped to avoid look-ahead bias.

The neural network architecture is deliberately simple: two hidden layers with ReLU activations, dropout regularization, and batch normalization. We train with a rolling window of 60 months and evaluate out-of-sample over the subsequent 12 months. Cross-validation is performed at the country-sector level to prevent information leakage.

Key findings

Our results demonstrate that even shallow networks achieve meaningful improvements over linear benchmarks. The long-short portfolio generated by the neural network model produces a Sharpe ratio of 0.72 out-of-sample, compared to 0.48 for the equivalent linear factor model. Importantly, the gains are concentrated during periods of macroeconomic regime change, where nonlinear interactions between factors become most relevant.

Feature importance analysis reveals that the network learns to weight growth-inflation interactions and policy divergence signals more heavily than a linear model would, suggesting it captures economically meaningful nonlinearities rather than merely overfitting to noise.

Implications for portfolio construction

These findings suggest that institutional investors using macro-quantamental frameworks for equity selection can benefit from modest increases in model complexity. The key insight is not that deep learning is necessary, but that allowing for nonlinear factor interactions through even a simple network architecture can meaningfully improve risk-adjusted returns.