PERFORMANCE OPTIMIZATION OF DEEPSEEK MOE ARCHITECTURE IN MULTI-SCALE PREDICTION OF STOCK RETURNS

Stock market data has significant multi-scale characteristics. High-frequency data (such as minute-level price fluctuations) contains rich but noise-intensive short-term information, while low-frequency data (such as daily trend) reflects long-term market dynamics but has response delays. Traditional time-series models (such as LSTM or Transformer) have inherent limitations in processing multi-scale features: the recursive structure of LSTM is difficult to efficiently process high-frequency noise, and the self-attention mechanism of Transformer is insufficient in capturing local features and has a large number of parameters. This study proposes a dynamic routing optimization framework based on DeepSeek MoE (Mixture of Experts), which realizes effective decoupling and fusion of multi-scale features through a hierarchical processing architecture, intelligent routing mechanism, and efficient parallel computing technology. Experimental results show that on the Shanghai-Shenzhen 300 constituent stocks (2018-2024) dataset, the high-frequency prediction error of the model is reduced by 32.7% compared with traditional methods, and the maximum drawdown rate under extreme market conditions is reduced by 41%. Gradient attribution analysis reveals the dominant role of liquidity factors (such as turnover rate) in the prediction results, providing an interpretable intelligent decision-making framework for quantitative investment.