1. Report Industry Investment Rating No relevant content provided. 2. Core Viewpoints of the Report - The report innovatively proposes a dual - channel deep - learning model (LSTM and GRU) that integrates daily - frequency and minute - frequency data, which can effectively capture market information on different time scales, significantly improve the prediction accuracy and stability of the strategy outside the sample (especially during market downturns), and provide a new idea with strong generalization ability for reconstructing the quantitative timing system of the bond market [2]. - The dual - channel model shows excellent generalization ability and robustness in out - of - sample tests, and can maintain a high winning rate in bear markets, effectively making up for the shortcoming of traditional factors failing in market downturns [3]. - In the multi - factor timing framework, the weight of deep - learning factors should be controlled at a relatively low proportion, and machine - learning factors should play a supplementary role to achieve the unity of interpretability and performance improvement [43][44]. 3. Summary by Relevant Catalogs 3.1 Deep - Learning Model Introduction - Traditional quantitative factors in the bond market have declined in performance in recent years, and there is a need to reconstruct and re - mine bond - market quantitative factors. Deep - learning methods can be used to find complex relationships in data, and RNN, LSTM, and GRU are considered suitable for the timing task of Treasury bond futures [7][8]. - RNN can process time - series data but has the problem of gradient disappearance when dealing with long time - series [9]. - LSTM solves the gradient - disappearance problem through a cell state and three gating units, enabling it to learn long - range dependencies in sequences [15]. - GRU simplifies the structure of LSTM, reduces the number of learnable parameters, and has high parameter efficiency and fast training speed [19]. - A dual - channel model is designed to process daily - frequency and minute - frequency data simultaneously to extract features on different time scales and predict the daily - frequency returns of Treasury bond futures, which can reduce the over - fitting risk [22]. 3.2 Treasury Bond Futures Timing Test 3.2.1 Back - testing Settings - The target variable is the open - to - open return of 10 - year Treasury bond futures, and the back - testing time interval is from January 2016 to August 2025, with daily rebalancing, 100% margin, 1 - time leverage, and a bilateral handling fee of 0.01% [25][26][27]. 3.2.2 Daily - frequency Channel Model - The single - daily - frequency channel model based on daily - frequency features performs well within the sample but poorly outside the sample, with obvious over - fitting [33]. 3.2.3 Dual - channel Model - The dual - channel model fuses multi - frequency time - series information. The addition of minute - frequency information significantly improves the prediction effect of the model outside the sample, enhances the generalization ability and stability, and maintains a relatively high winning rate in both long and short positions [40][41][42]. 3.3 Deep - Learning Allocation in the Multi - factor Framework - Deep - learning factors in the multi - factor timing framework have high performance but also have over - fitting risks and lack of interpretability. The weight of deep - learning factors should be controlled at a relatively low proportion, and machine - learning factors should play a supplementary role [43][44]. 3.4 Conclusion - The report explores the application of deep - learning models in Treasury bond futures quantitative timing and proposes a dual - channel deep - learning framework based on multi - frequency data fusion, which can effectively improve the performance of multi - factor strategies [45].