DeepAlpha-DNN Equity Backtests Across Market Regimes
Summary
The article evaluates a deep-learning stock-ranking model across broad-market and filtered Chinese equity universes. It uses rolling windows of three years for training and one year for evaluation, then compares portfolio sizes, rebalance intervals, weighting choices, and universes such as highly active stocks, limit-up leaders, index constituents, institutional favorites, and recent listings. One reported setup holds ten stocks equally weighted and rebalances every two trading days; the author also says changing the relative-return label improved results.
The reported backtests include annual returns, Sharpe ratios, alpha, and drawdowns. Several activity-based universes show stronger aggregate results than the index-restricted universe, while 2022 is weak or negative for many configurations. The author presents the model as a complement to other signals and notes that substantial changes to labeling, normalization, and portfolio rules are essential; the original model alone did not reproduce these outcomes. The evidence is historical and author-reported, with no independent validation or detailed transaction-cost analysis, so claims of robustness and live suitability remain uncertain.
Key ideas
- The study compares DeepAlpha-DNN stock selection across multiple equity universes and rebalance schedules.
- It evaluates with rolling three-year training and one-year test periods.
- A two-day rebalance of ten equally weighted stocks is among the tested configurations.
- Filtering for highly active stocks produced stronger reported results than restricting selection to index constituents.
- The author attributes results partly to changes in labels, normalization, and portfolio construction, limiting attribution to the original model.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.