Black-Litterman Dynamic Portfolio Based on Hybrid Learning Model

Stock portfolio management is one of the core issues of quantitative trading research, which is significant to improving a portfolio’s return. As one of the main methods of portfolio construction, the Black-Litterman framework has two difficulties: first, how to design a model to give investor future views on stocks when faced with nonlinear and high-noise financial data; second, how to make stock dynamic portfolio selection based on the investor views, and perform backtesting of performance in actual trading scenarios. To solve the problems, we propose a hybrid learning model that integrates ARIMA, GM, and LSTM for multi-period forecasting of stock prices and takes the prediction results as investor views. When performing multi-period forecasting, we dynamically design the dataset for training and forecasting and use the single-period forecast results as part of the multi-period forecast results. Finally, we use the multi-period forecast results for stock portfolio selection and backtest the performance of the proposed model. The experimental results show that the model proposed is better than the existing popular portfolio models in the four performance metrics. Meanwhile, the ablation experiments also confirm the effectiveness of the proposed portfolio method.