金融工程研究报告：多元时序预测在行业轮动中的应用

Quantitative Models and Construction Methods 1. Model Name: Multivariate CNN-LSTM - Model Construction Idea: The model leverages the advantages of CNN and LSTM in different scenarios to predict multiple parallel financial time series by considering the correlation between them[12][14]. - Detailed Construction Process: - General Structure: The model consists of an input layer, a one-dimensional convolutional layer, a pooling layer, an LSTM hidden layer, and a fully connected layer to produce the final prediction results[14]. - Formula: ${\hat{x}}_{k,t+h}=f_{k}(x_{1,t},\dots,x_{k,t},\dots,x_{1,t-1},\dots,x_{k,t-1},\dots)$ This formula indicates that each variable depends not only on its past values but also on the past values of other variables[11]. - Hyperparameters: - Number of convolution filters: 64 - Convolution kernel size: 2 - Use of padding: Yes - Pooling layer window size: (2,2) - Number of hidden units in the first LSTM layer: 128 - Number of hidden units in the second LSTM layer: 128 - Activation method between LSTM layers: ReLU - Time series look-back window: 10 - Number of training epochs: 100[20] - Evaluation Metric: Root Mean Square Error (RMSE) $RMSE={\sqrt{\frac{1}{n}\sum_{i}({\hat{y_{i}}}-y_{i}\,)^{2}}}$ where $y_i$ represents the standardized index price, and $\hat{y_i}$ represents the CNN-LSTM prediction value[21]. - Model Evaluation: The model achieved good tracking and high accuracy in predicting multiple parallel financial time series, similar to the performance in predicting stock indices in the Asia-Pacific market[14][17]. 2. Model Name: Grouped Multivariate CNN-LSTM - Model Construction Idea: To improve prediction accuracy, the industry indices are grouped based on investment attributes, and a separate prediction model is constructed for each group[26][27]. - Detailed Construction Process: - Grouping: The industry indices are divided into six groups: Consumer and Medicine, Upstream Resources and Materials, High-end Manufacturing, Real Estate and Infrastructure, Big Tech, and Big Finance[27]. - Model Structure: Each group of industry indices is predicted using a separate CNN-LSTM model, as shown in the general structure diagram[28]. - Evaluation Metric: The prediction accuracy is evaluated using RMSE, similar to the original model[33]. - Model Evaluation: Grouping and training different CNN-LSTM sub-models for each industry group improved the prediction accuracy, especially for industries with previously low prediction accuracy[30][32]. Model Backtesting Results 1. Multivariate CNN-LSTM Model - Prediction Error (Training Phase): 1.52% to 3.18%[23] - Prediction Error (Testing Phase): 1.56% to 3.30%[23][25] 2. Grouped Multivariate CNN-LSTM Model - Prediction Error (Training Phase): 1.49% to 2.60%[33] - Prediction Error (Testing Phase): 1.61% to 2.82%[33] Quantitative Factors and Construction Methods 1. Factor Name: Weekly Industry Rotation Signal - Factor Construction Idea: Use the predicted values from the multivariate CNN-LSTM model to estimate the future weekly returns of industry indices and select the top five industries with the highest expected returns for equal-weight allocation[3]. - Detailed Construction Process: - Prediction: Predict the future weekly returns of industry indices using the multivariate CNN-LSTM model[34]. - Allocation: Every five trading days, select the top five industries with the highest expected returns for equal-weight allocation[35]. - Training: Retrain the model at the beginning of each quarter using an extended window of historical data from March 2014 to the training point[35]. - Factor Evaluation: The annualized return of the industry rotation portfolio reached 15.6%, with an annualized excess return of approximately 11.6%, and the risk-return characteristics significantly improved compared to the benchmark[3][35]. Factor Backtesting Results 1. Weekly Industry Rotation Signal - Annualized Return: 15.6%[38] - Annualized Volatility: 25.6%[38] - Maximum Drawdown: -27.1%[38] - Sharpe Ratio: 0.7[38] - Longest Drawdown Recovery Time: 248 days[38]