In the domain of quantitative trading, the creation of stock selection strategies often relies on factors, also known as influencers. Factors are typically generated by transforming historical data through various computational methods into signals that can indicate market trends, known as factor values. In this study, we utilized finance, statistics, and econometric models (such as GARCH and ARIMA) to construct a comprehensive factor dataset from various frequency dimensions based on daily and minute-level historical data, totaling 67 factors. Based on this dataset, we employed XGBoost and LightGBM models from the ensemble learning strategy to fit the next period’s return rates and integrated the concept of incremental learning for rolling predictions, thus generating optimized multi-factor combination results to aid in the creation of stock selection strategies. The experimental results show that under our constructed factor library and fitting strategy, the composite factor values generated in the CSI 300 stock pool can achieve an average RankIC of 0.0871, indicating strong predictive performance.