Deep Alpha-CNN Model Testing

Article BigQuant

Summary

This document describes a competition built around Deep Alpha-CNN, an equity prediction model using a seven-layer one-dimensional convolutional neural network with residual connections. The report’s stated findings favor the deeper network over a two-layer version, suggest that smaller kernel, batch, and feature-map settings performed better, and claim that rolling tests showed more resilience across bull and bear markets than a deep neural network baseline. The competition invited users to alter model structure and parameters and compare results using benchmark and rolling-test scores.

The post provides reported performance figures for a baseline model and an eight-year rolling test, including returns, Sharpe ratios, volatility, and drawdown. These results are presented without enough detail here to assess data construction, trading costs, validation design, or the scoring formulas’ suitability; in particular, the scoring description includes terms whose sign or scaling is unclear. The findings are therefore specific to the reported experiment and do not establish that the architecture will generalize to other markets or periods.

Key ideas

The model uses a seven-layer one-dimensional CNN with residual connections.
The report says the deeper CNN outperformed a two-layer network in its tests.
It reports better performance with smaller kernel, batch, and feature-map settings.
Rolling tests are presented as evidence of greater regime resilience than a DNN baseline.
The reported results lack sufficient methodological detail to establish generalization.

Deep Alpha-CNN: Testing a Residual Seven-Layer Model Across Market Regimes

Summary

Key ideas

Tags