A Robust Transferable Deep Learning Framework for Cross-sectional Investment Strategy
A Robust Transferable Deep Learning Framework for Cross-sectional Investment Strategy
This paper introduces RIC-NN, a deep learning framework for cross-sectional stock return prediction that addresses the challenge of non-stationarity in financial markets. The framework combines three key innovations: (1) a nonlinear multi-factor approach using a seven-layer neural network with augmented features (20 factors plus lagged and ratio transformations), (2) a stopping criterion based on the Ranked Information Coefficient (rank IC) to prevent overfitting by terminating training when the rank IC on the training window reaches a predefined threshold (0.20), and (3) deep transfer learning across regions to exploit market interdependencies. The model is trained on a rolling 10-year window and used to rank stocks for long and long-short portfolio strategies.
Experiments on MSCI North America (NA) and Asia Pacific (AP) indices from 2005 to 2018 show that RIC-NN consistently outperforms off-the-shelf machine learning methods (LASSO, random forest, standard neural networks) in terms of risk-adjusted returns. Notably, transfer learning from NA to AP significantly improves performance (IR from 1.25 to 1.46, R/R from 1.25 to 1.53), while reverse transfer is ineffective, aligning with the known asymmetric causality from NA to AP markets. The rank IC-based stopping criterion proves more robust than epoch-based stopping, which is sensitive to the choice of epochs.
RIC-NN also surpasses the average performance of top equity investment funds in both regions, even after deducting conservative transaction costs. The framework's simplicity and universality—using identical architecture and stopping values across different markets—make it a practical tool for quantitative investment. Future work may explore more sophisticated portfolio strategies and stateful models like recurrent neural networks with attention mechanisms.
Highlights
- 1Proposes RIC-NN, a deep learning framework for cross-sectional stock return prediction that integrates nonlinear multi-factor modeling, rank IC-based stopping criteria, and transfer learning across regions.
- 2Introduces a novel stopping criterion based on the Ranked Information Coefficient (rank IC) to mitigate overfitting in dynamic financial markets, outperforming epoch-based stopping.
- 3Demonstrates that transfer learning from North America to Asia Pacific significantly improves performance, consistent with asymmetric market causality.
- 4RIC-NN outperforms LASSO, random forest, and standard neural networks in both long and long-short portfolio strategies across MSCI North America and Pacific indices over 14 years.
- 5The strategy achieves higher risk-adjusted returns than the average of top equity investment funds, even after accounting for transaction costs.
Methods
- MSeven-layer feedforward neural network with ReLU activation, dropout, and batch normalization, trained on 20 financial factors augmented with lagged and ratio features.
- MRolling-horizon training with a 10-year sliding window, using Adam optimizer and mini-batch size of 300.
- MStopping criterion based on rank IC threshold (0.20) to prevent overfitting, with initialization from previous time step's weights at rank IC 0.16.
- MTransfer learning by initializing the first four layers of the target region's network with weights trained on the source region.
Results
- RIn MSCI North America, RIC-NN achieves an information ratio of 0.30 (long) and R/R of 0.49 (long-short), outperforming LASSO (0.11, 0.21), RF (0.13, 0.15), and NN (0.18, 0.22).
- RIn MSCI Pacific, RIC-NN with transfer from NA achieves IR of 1.46 and R/R of 1.53, surpassing all baselines and the non-transfer version.
- RTransfer learning from NA to AP yields a 17% improvement in IR (1.25 to 1.46) and 22% in R/R (1.25 to 1.53), while reverse transfer is ineffective.
- RRIC-NN's rank IC-based stopping consistently outperforms epoch-based stopping across different epoch choices in both markets.
- RAfter transaction costs, RIC-NN achieves annualized returns of 7.79% (NA) and 9.44% (AP), exceeding the average fund returns of 5.90% and 7.88%, respectively.
Analyze Paper
Generate insights from "A Robust Transferable Deep Learning Framework for Cross-sect...".