Interpretable Factor Design for Stock Ranking Models
Summary
The article compares a StockRanker stock-selection strategy with a Deepalpha deep neural network using the same 17 short-term factors, training data, and backtest setup. The strategy selects the top-ranked A-share stock each trading day and sells it the following day. The factors comprise volume-price and moving-average measures. The author reports strong historical and simulated-live results for StockRanker, while stating that switching to the neural network did not directly improve the strategy. The live observation period for the neural network was considered too short to support a reliable conclusion.
The broader argument is that repeated searches across factors and model configurations can find attractive historical paths that may not persist. The author favors interpretable factors and trading rules, grouping stocks by trading context and matching them with relevant signals. Candidate factors are tested individually and in combinations; human trading knowledge narrows the search space. After selecting a model, the author adds controls based on its behavior, such as reducing trading frequency when liquidity is poor. These are the author’s process and reported results, not independent validation; extensive search itself can make backtest results fragile.
Key ideas
- The comparison holds the factors, training data, and backtest setup constant while changing the model algorithm.
- The strategy ranks A-shares daily, buys one leading candidate, and sells it the next day.
- The author reports no direct performance improvement from replacing StockRanker with a deep neural network.
- Interpretable factors and trading context are used to narrow model searches.
- The author proposes adding risk controls that reflect model behavior and market liquidity.
- The short live observation period limits conclusions about the neural network’s effectiveness.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.