Abstract:

The violent fluctuations of the stock market pose a threat to financial stability and have a significant impact on a country's economic development. Therefore， understanding and predicting stock market fluctuations play a crucial role in evaluating a country's economic performance. Stock returns exhibit characteristics such as non-stationarity， nonlinearity， and volatility aggregation. As a result， stock return forecasting has garnered substantial interest among scholars. However， most existing studies solely rely on historical stock price sequences for prediction， which often leads to subpar results. The weekly frequency of fund position changes holds significant value in determining future market trends. Increasing fund positions can drive stock market upswings， while individual retail investors tend to follow and mimic these position changes， thereby influencing future stock market movements. Recognizing the information gain effect of fund position changes on the stock market and the intricate relationship between these two types of data， a novel model is proposed that combines the reverse mixed data sampling model （R-MIDAS） with machine learning algorithms. The model is applied to predict the index return rate and investment strategy for 27 industries.The empirical results demonstrate several key findings. Firstly， the performance of the R-MIDAS-GWO-SVR algorithm surpasses that of other benchmark models， such as R-MIDAS-SVR， R-MIDAS-CNN， and R-MIDAS-LSTM. In particular， the R-MIDAS-GWO-SVR model outperforms the LR model in 19 industries. Secondly， the proposed model exhibits excellent performance in single-industry investment strategies， as indicated by risk-adjusted performance indicators based on the forecasted results. Lastly， when considering multi-industry portfolio investments， the R-MIDAS-GWO-SVR model consistently outperforms other models for various values of k （specifically， 5， 7， and 9）. The combination of the R-MIDAS model and machine learning methods shows promising potential in predicting mixed frequency data. These findings contribute to the literature by introducing a new approach to stock return forecasting and highlighting the importance of incorporating fund position changes into prediction models. The proposed model has significant implications for investors， regulators， and policy makers in making informed decisions and formulating effective investment strategies in the stock market.

Key words: stocks return, time series forecasting, reverse mixed frequency data sampling, public offering of fund position, investment strategy