1. Report Industry Investment Rating No information provided in the report. 2. Core Viewpoints - Different contracts of the same futures variety have significantly different effective factors for basis, which confirms the differences in trading groups and preferences on each contract [4]. - The effective factors and the correlations of the same factors vary in different time periods, indicating that the factors influencing the basis differ in different market environments [4]. - Combining cross - sectional factor models with time - series models can improve the prediction effect to some extent [4]. 3. Summary by Directory 3.1 Basis Prediction Model Extension and Optimization - Based on the previous research report, the prediction process is expanded and improved in four aspects: adding new factors and adjusting factor selection criteria, changing fixed factors to rolling - screened factors, extending from single - contract to full - contract prediction, and adding time - series models to build a multi - model fusion prediction system [10]. 3.2 Factor Addition and Screening - New factors are added from the perspective of capital behavior, including using different types of ETF shares and market values to represent different market styles, and using CFTI seat classification data to represent different customer groups. Most ETF shares and market values have significant correlations with the basis, and there are certain rules in the direction of the correlations. For example, aggressive ETFs are negatively correlated with the basis, while stable ETFs are positively correlated. Some CFTI factors also show significant correlations, such as retail net positions being positively correlated with the basis and institutional net positions being negatively correlated [11][14][16]. - As the number of candidate factors expands, a factor correlation matrix is constructed to screen highly correlated factor pairs and retain factors more highly correlated with the basis to improve model robustness [17]. 3.3 Rolling - Screened Factors - The static fixed factor selection method is changed to a rolling - window dynamic factor screening method to capture effective factors with stronger explanatory and predictive power for the basis in different periods, enhancing the model's adaptability to different market environments [18]. - During the rolling test, the correlations of some factors change significantly. For example, the correlations of excess - related factors with the annualized basis weaken over time, while the influence of margin balance and other factors increases. The correlations of volatility - related factors change from negative to positive, and the positive correlation of scale - index ETF factors strengthens [20][21]. 3.4 Full - Contract Prediction - The prediction object is extended from a single contract to all contracts. Different contracts have different effective factors, which verifies the differences in trading groups and characteristics on each contract. Near - month contracts are more speculative, affected by short - term market fluctuations, while far - month contracts are more for medium - to - long - term hedging, related to market investment styles and hedging demand [25][27]. 3.5 Multi - Model Fusion - An AR time - series model is introduced on the basis of the original cross - sectional factor model to capture the time - series dynamics and mean - reversion law of the basis, forming a dual - prediction logic of cross - sectional information and time - series characteristics. The model training and back - testing use a rolling - window framework, and rolling factor screening and dynamic parameter adjustment are performed simultaneously [28]. - After optimization, the direction prediction accuracy of the linear regression model is significantly improved, and the time - series model has better overall direction accuracy. The random forest model performs better in terms of prediction accuracy. The 4 - model combination has an average prediction MSE of about 0.0374 and an average direction accuracy of about 61% for IC contracts. For other varieties (IH, IF, IM), the 4 - model combination also shows certain improvements in prediction accuracy and direction judgment [32][36][38]. 3.6 Summary - The report expands and improves the basis prediction model in four aspects. Taking IC as an example, different contracts have different correlations with factors. Near - month contracts are more speculative, and far - month contracts are more for hedging. The correlations of factors change over time, and the rolling - screening method has different effects on different contracts and varieties. Time - series models have better direction accuracy, and machine - learning models are better in prediction accuracy [65].

Core Insights - JD Finance has launched an AI wealth management tool named "Jing Xiaobei," which utilizes multi-model fusion and multi-agent collaboration to enhance investment opportunities and provide a smarter wealth management experience [1] Group 1: Multi-Model Fusion - Jing Xiaobei integrates a hybrid large model system that combines general capabilities with specialized financial models, enhancing financial analysis and decision-making [2] - The tool accesses vast amounts of structured and unstructured financial data, including real-time market data and macroeconomic indicators, to respond quickly to market dynamics [2] - For example, in fund diagnostics, Jing Xiaobei can analyze market trends and generate multi-dimensional quantitative evaluation reports to support user decision-making [2] Group 2: Risk Mitigation - Jing Xiaobei addresses AI "hallucination" risks by utilizing data from the JD Finance platform and employing methods like position analysis and risk preference tracking [3] - The system establishes a dynamic data lineage tracking system to ensure data reliability and automatically monitors asset portfolio deviations, triggering risk alerts when necessary [3] Group 3: Comprehensive Service Innovation - The product features five core functions: investment opportunities, intelligent analysis, asset optimization, risk alerts, and growth support, all integrated into the JD Finance app [4] - Jing Xiaobei offers personalized recommendations based on user profiles and historical preferences, enhancing user experience and operational efficiency [4] - The AI tool continuously learns from user interactions, improving service accuracy over time and transforming complex financial decision-making into a closed-loop experience [4] - JD Finance aims to deepen the integration of AI technology with financial scenarios, iterating on model capabilities and expanding service boundaries to reshape wealth management [4]