Testing Neural Network Stock Ranks with Sector and Market Factors
Summary
The author investigates whether the highest-ranked stocks from a convolutional neural network strategy deliver the best results. The experiments apply a model trained on a small number of stocks to broader selections, then backtest separate groups of ranked stocks. Inputs include sector-level averages and stock returns relative to their sector, alongside a market-wide return measure for the small-cap universe. The author also describes training on daily selections and exiting the following day.
The reported tests suggest Sharpe ratios vary systematically across rank ranges rather than randomly, motivating the idea of selecting a rank band with stronger historical results. The evidence is exploratory: the document gives no detailed charts or full statistical validation, and the tests use a particular model, Chinese small-cap universe, and historical split. The author notes uneven stock counts across industries and warns that concept-group data may lose relevance over time. The described strategy omits risk controls, so its observations should not be taken as evidence of robust live performance.
Key ideas
- The top-ranked stock in a neural-network model may not be the best-performing choice.
- Sector aggregates and stock returns relative to sector performance can provide model features.
- The author compares backtests across rank bands and observes that Sharpe ratios appear to vary systematically.
- A selected rank interval could be tested as a way to reconstruct a portfolio from model predictions.
- Uneven sector membership, changing concept groups, and the absence of risk controls limit the conclusions.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.