主观投资框架验证与个股决策Agent

Quantitative Models and Construction Methods 1. Model Name: Investment Chain of Thought (CoT) Dynamic Generation Mechanism - Model Construction Idea: Utilize large language models (LLMs) to extract, track, and validate dynamic investment logic chains from unstructured data like research reports, enabling the identification of core pricing logic in complex markets[2][3][14] - Model Construction Process: 1. Data Input: Research reports are classified by industry and weighted by quality metrics (e.g., report length, analyst team authority)[24] 2. Logic Chain Generation: LLMs generate monthly investment logic chains by identifying key drivers, constructing causal pathways, and scoring signal strength[26] 3. Time-Series Processing: Logic chains are merged and updated semi-annually, ensuring they only include past data and adapt to signal decay[26] 4. Validation: Rolling backtesting is conducted to verify the effectiveness of each logic chain in the current market environment[3][34] 5. Output: High-quality logic chains are selected for investment strategies[3][24] - Model Evaluation: The CoT framework effectively captures complex pricing logic and provides differentiated alpha opportunities, but individual logic chains may face temporary or permanent invalidation risks[3][34] 2. Model Name: Multi-Logic Chain Voting Strategy - Model Construction Idea: Combine multiple high-quality CoT logic chains to enhance the robustness and stability of stock selection strategies[4][46] - Model Construction Process: 1. Logic Chain Selection: Top 1/3 CoT logic chains are selected based on information ratio (IR) performance[45] 2. Voting Mechanism: Each selected logic chain votes for stocks that meet its criteria, with a threshold of three votes required for inclusion in the portfolio[46] 3. Rebalancing: Portfolios are rebalanced monthly, with empty positions maintained during periods of low effective logic coverage[45][46] - Model Evaluation: The strategy demonstrates superior excess returns and stability compared to single-chain models, though it faces challenges during industry downturns[57] 3. Model Name: Weighted CoT Stock Selection Strategy - Model Construction Idea: Assign weights to stocks based on the performance of the CoT logic chains selecting them, improving risk control and return consistency[58] - Model Construction Process: 1. Weight Assignment: Stocks selected by higher-performing logic chains receive higher scores, with scores derived from the information ratio of the chains during validation[58] 2. Threshold Filtering: Stocks with cumulative scores above a set threshold (e.g., 3.5) are included in the portfolio[58] 3. Rebalancing: Portfolios are adjusted monthly, with signals carried over during periods of low effective logic coverage[58][62] - Model Evaluation: The strategy achieves better risk control and higher information ratios compared to the voting strategy, though it struggles with signal mismatches during rapid market shifts[62] --- Model Backtesting Results 1. Investment Chain of Thought (CoT) Dynamic Generation Mechanism - Annualized excess return: Ranges from -23.81% to 26.32% depending on the logic chain[35] - Tracking error: Ranges from 11.07% to 37.39%[35] - Information ratio (IR): Ranges from -1.09 to 1.84[35] 2. Multi-Logic Chain Voting Strategy - Annualized excess return: 17.16% (relative to analyst equal-weight benchmark)[57] - Annualized volatility: 35.87%[57] - Maximum drawdown: 43.99%[57] - Information ratio (IR): 0.48[57] 3. Weighted CoT Stock Selection Strategy - Annualized excess return: 16.21% (relative to analyst equal-weight benchmark)[61] - Annualized volatility: 31.94%[61] - Maximum drawdown: 44.17%[61] - Information ratio (IR): 0.51[61] --- Quantitative Factors and Construction Methods 1. Factor Name: Signal Strength (CoT Signal) - Factor Construction Idea: Quantify the effectiveness of each logic chain in predicting stock performance[26][29] - Factor Construction Process: 1. Driver Identification: Key drivers (triggers) are identified as the starting point of each logic chain[29] 2. Causal Pathway Construction: Drivers are linked to investment conclusions through a series of reasoning steps[29] 3. Signal Scoring: Each chain is assigned a signal strength score based on its backtesting performance[29] - Factor Evaluation: Signal strength effectively filters out weak or outdated logic chains, ensuring the robustness of the CoT framework[29] --- Factor Backtesting Results 1. Signal Strength (CoT Signal) - Top-performing logic chain: "New Product Development and Commercialization Progress" with an IR of 1.84[35] - Lowest-performing logic chain: "Technology Empowerment and Service Upgrades" with an IR of -1.09[35]