Testing a Deep Alpha CNN with Multiple Factor Time Slices

Summary

This competition entry modifies a Deep Alpha convolutional neural network by supplying five time slices for each of 98 factors, rather than using a single slice. The network structure and most parameters remain close to the original model because training time was limited. The author reports that the expanded input performed worse than the single-slice version and suggests that further hyperparameter tuning may be needed.

The document gives aggregate performance statistics for a baseline and yearly results from rolling tests across eight years, including returns, Sharpe ratios, and maximum drawdowns. The variation across years includes both strong and negative outcomes, underscoring that the reported result is not uniform across periods. These are backtest results from a competition entry, not evidence of live performance; the article gives limited detail on data construction, costs, validation safeguards, or parameter search. It also warns that its platform resources describe an older version and may no longer apply to the current platform.

Key ideas

• The model input was changed from one to five time slices for each of 98 factors.
• Most of the network design and parameters were retained because training and competition time were limited.
• The author reports worse performance for the expanded time-slice input and identifies tuning as a possible next step.
• Yearly rolling-test results vary, including negative performance in some periods.
• The results are historical backtests with limited methodological detail, and the platform material is marked as outdated.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.