Abstract:

Accurately predicting future stock returns can effectively enhance the out-of-sample performance of a investment portfolio， yet a single prediction model may struggle to robustly forecast a vast number of stock assets， each with distinct data characteristics. To address this issue， a mixed ensemble algorithm based on machine learning and ensemble learning is proposed to predict a wide variety of stock assets. Specifically， the proposed mixed ensemble algorithm is based on a mix of five single models and five ensemble models： including Multiple Linear Regression （MLR）， Support Vector Regression （SVR）， Generalized Regression Neural Network （GRNN）， Convolutional Neural Network （CNN）， Long Short-Term Memory （LSTM）， Simple Average， Linear Weighted， Random Forest （RF）， Gradient Boosting Decision Tree （GBDT）， and Entropy Weighting. By generating investor views through the proposed mixed ensemble algorithm and pre-selecting a portion of high-quality stocks， the Black-Litterman （BL） model is further incorporated to construct a more effective portfolio strategy. In the empirical section of this study， 105 constituent stocks from the CSI 300 index were selected as research samples， covering 242 trading days from January 1， 2022， to December 30， 2022， with daily data frequency. Empirical results show that， the mixed ensemble algorithm significantly reduces the forecasting error of future stock returns compared to benchmark models. After stock pre-selection the out-of-sample performance of all portfolio strategies significantly surpasses the CSI 300 Index. Notably， the out-of-sample performance of the constructed portfolio strategy by using the mixed ensemble forecasting algorithm and BL model markedly is superior to traditional portfolio strategies such as minimum variance， maximum Sharpe ratio， maximum expected return， equal weight， and equal risk weight， as well as other benchmark forecasting algorithms and BL model-constructed portfolio strategies. Finally， a robustness test is conducted by varying the rolling window's length， which further verifies the robustness of this approach. The proposed portfolio optimization model can significantly improve the out-of-sample performance of a vast number of stocks and provide theoretical guidance for stock investment in the real market.

Key words: portfolio strategy, ensemble learning, Black-Litterman model, stock return prediction, out-of-sample test