1. Report Industry Investment Rating - Not provided in the content 2. Core Viewpoints of the Report - The report introduces the LSTM time - series model to improve the prediction of Treasury bond futures inter - period spreads without expanding the original factor set, finding that introducing time - series modeling can improve return stability and risk control for some varieties, but the sensitivity to time - series information varies among different contracts [38][39][40] - The LSTM model outperforms the DNN model on T and TF contracts, while the DNN model shows stronger prediction ability on TS contracts; the performance of both models on the TL contract is limited due to the short listing time [38][39] - Using the LSTM model's prediction probability for position weighting can enhance the return - risk characteristics of strategies on T, TF, and TS contracts, but the effect is not obvious on the TL contract [38][39] 3. Summary by Relevant Catalogs 3.1 Inter - period Spread Influencing Factors and Indicator Selection - Traditional machine - learning models and DNNs have limitations in time modeling as they assume sample independence in the time dimension and rely on manual feature engineering to introduce time information, and DNNs lack explicit sequence structure [7] - LSTM is an improved form of RNN, introducing a memory unit to decouple long - term information storage and current state output, and using gate - control structures to control information flow, which can better capture long - term dependencies [7] - Treasury bond futures inter - period spreads show trend characteristics in some stages, which are difficult to capture by previous models. Therefore, the LSTM time - series model is introduced to enhance the description of the time - series structure [12] 3.2 Recurrent Neural Network Testing Process 3.2.1 Data Processing and Sample Construction Process - The data includes fundamental factors of T, TF, TS, and TL contracts, and the features are constructed by aligning and introducing capital - related indicators and creating derived spread variables. The label is the first - order difference of the inter - period spread [16][17] - Data pre - processing involves removing early - listing samples and some end - of - month trading days, filtering small - amplitude spread changes, filling missing values, and standardizing data [18] - Time - series samples are constructed using a sliding window with a 5 - trading - day historical window, and samples are divided into training, validation, and test sets by strict time segmentation [19][20] - The model is trained using a weighted cross - entropy loss function and the Adam optimizer, with learning - rate decay and early - stopping mechanisms based on validation - set loss [22] 3.2.2 Parameter Setting and LSTM Network Structure - The task is a binary - classification prediction for the next - trading - day direction change of the inter - period spread, with input as a sequence of factor values over 5 consecutive trading days and output as binary logits [24] - The LSTM network has 3 layers, a hidden - state dimension of 8, and a Dropout ratio of 0.3. It uses the hidden state at the last time - step for classification [25] - The weighted cross - entropy loss is used to address class imbalance, and the Adam optimizer with learning - rate decay is applied for parameter updates [26][27] 3.3 Model Test Results 3.3.1 Comparison of Out - of - Sample Tests between LSTM and DNN Models - On T and TF contracts, the LSTM model has higher cumulative returns, Sharpe ratios, and better drawdown control compared to the DNN model; on the TS contract, the DNN model performs better; on the TL contract, the performance of both models is limited [28][30] 3.3.2 Probability - Weighted Backtesting of the LSTM Model - The prediction probability of the LSTM model is used as the position signal, and the position is normalized and limited to control risks [33] - Probability - weighted strategies improve the out - of - sample performance on most contracts, enhancing returns and Sharpe ratios without significantly increasing drawdowns, but the effect is not obvious on the TL contract [35] 3.4 Conclusion - Introducing the LSTM time - series model can improve the prediction of Treasury bond futures inter - period spreads for some varieties, but the effectiveness depends on contract characteristics, sample coverage, and spread structure [38][39][40] - Using prediction probability for position weighting has potential in improving strategy performance, and future research can further test the applicability of time - series modeling under more complex conditions [39][40]