Factor timing with panel-data CNNs: evidence from China’s A-share market

This paper proposes a modification to CNN-based technical analysis: instead of feeding each asset’s historical returns in isolation, we stack cross-sectional vectors into a two-dimensional market-state matrix, allowing the model to utilize both temporal and cross-sectional information. Applied to factor timing in China’s A-share market, the strategy achieves a monthly long-short return of 1.01% with risk-adjusted alphas that remain significant after controlling for common factors. The panel-data design consistently outperforms a pure time-series benchmark on identical factor-level data.