Core Insights - The article focuses on optimizing a timing model based on price and volume factors, addressing issues of instability, signal volatility, and the reliability of single signals [1][2]. Factor Selection - The model employs a dual standard of group IC and threshold settings to tackle the challenge of unstable effectiveness, ensuring that selected factors can consistently predict outcomes across different value ranges [2]. Model Training and Signal Generation - A strategy of random grouping and independent training is used to filter noise and balance signal robustness. The signal generation process involves rolling smoothing and multi-group voting to ensure accurate and stable timing signals [3]. - Backtesting from 2019 to September 2025 shows significant improvements over benchmarks, with a 1-day signal yielding an annualized return of 3.61% and a Sharpe ratio of 1.12, outperforming the benchmark [3].

Report Industry Investment Rating No relevant content provided. Core Viewpoints of the Report - The timing model constructed in a "scenario-based combat" approach in 2025 performed averagely due to factors such as unstable effectiveness, multicollinearity, and high volatility. A new volume-price factor timing model based on a grouping algorithm was reconstructed, focusing on optimizing three major issues: unstable effectiveness, large signal fluctuations, and insufficient reliability of single signals [4][7]. - By using double standards of grouped IC and thresholds in factor screening, factors that can stably play a predictive role in both high and low value intervals were selected, ensuring the effectiveness of model information from the source [4][10]. - Through the strategies of "random grouping + independent training" and "rolling smoothing + multi - group voting", noise was filtered, effective information was aggregated, and accurate and robust timing signals were generated [4]. - The back - test results showed that the model significantly outperformed the benchmark, especially the 1 - day signal, which demonstrated strong stable timing ability both within and outside the sample [4]. Summary According to the Directory 1. Factor Screening: Double Standards Determine Stability and Effectiveness - **Factor Reserve**: 87 factors covering intraday patterns, price fluctuations, trading volume statistics, trends, momentum, overbought and oversold conditions were constructed around the volume - price characteristics of Treasury bond futures, comprehensively covering core dimensions for timing [11]. - **Screening Standard**: An annual rolling back - test framework was adopted, with a 3 - year data window each year. Factors were sorted and divided into 5 groups. Two conditions were set: at least 4 groups of IC should maintain the same direction, and the average absolute value of the same - direction IC should be no less than 0.05. This mechanism could identify factors with stable cross - sectional prediction ability and adapt to market changes [13][14]. 2. Model Building and Signal Generation: Group Voting Based on Random Grouping and Cross - Validation - **Model Building**: A deep - learning architecture of bidirectional multi - layer GRU + attention mechanism was used, focusing on short - term timing requirements for T + 1 and T + 5 day price movements. Techniques such as Dropout and layer normalization were introduced to avoid overfitting, and the "rolling window" training mode was adopted [16]. - **Random Grouping and Signal Generation** - **Random Grouping of Factors and Parallel Training**: After selecting effective factors based on grouped IC values each year, they were randomly divided into multiple groups. Each group of factors was used to train an independent GRU sub - model, generating diverse prediction results [20]. - **Signal Generation**: The final timing signal was generated through "rolling smoothing + multi - group voting". Rolling smoothing was used to filter noise according to different prediction periods, and multi - group voting was used to confirm the signal direction, reducing the influence of single - sub - model prediction errors. A comprehensive signal was also generated by integrating 1 - day and 5 - day prediction results [21][22]. 3. Strategy Back - test Design and Parameter Selection: Balancing Robustness and Adaptability - **Strategy Back - test Design**: A long - short full - position trading mode was set. When the final signal was 1, a full - position long was taken; when it was - 1, a full - position short was taken; when it was 0, the position remained unchanged. Transaction costs and slippage were ignored, and the closing price of the 10 - year Treasury bond futures main contract was used as the benchmark [25]. - **Parameter Selection**: Three major dimensions of parameters were focused on: training window, number of groups, and number of factors in each group. Different parameter combinations were tested from 2019 to September 2025, and the optimal parameter combinations were selected based on the back - test results, emphasizing overall return - capturing ability and stable advantages over the benchmark in most years [26]. - **Back - test Performance**: The 1 - day signal performed outstandingly both within and outside the sample. Within the sample, the 1 - day signal had an average annualized return of 3.61% and a Sharpe ratio of 1.12; outside the sample, from the National Day holiday in 2025 to the present, the 1 - day signal had a cumulative return of 0.99%, a Sharpe ratio of 5.98, and a maximum drawdown of only 0.13%. The 5 - day signal's adaptability decreased outside the sample [30][34].