Quantitative Models and Construction Methods - **Model Name**: IRCF Factor **Construction Idea**: The IRCF factor is designed to capture the asymmetric behavior of institutional and retail investors under different market conditions, focusing on "institutional accumulation" and "retail panic" signals[9][48][49] **Construction Process**: 1. **Small Order Factor**: Calculate "small order sell count/small order buy count" and only take values during stock downturns to capture retail panic signals[48][49] 2. **Large Order Factor**: Calculate "large order buy count/large order sell count" and only take values during stock upturns to filter noise and identify institutional accumulation signals[48][49] 3. **Statistical Features**: Derive three statistical features (mean, standard deviation, 90th percentile) for both small and large order factors over a 40-day rolling window[48][49] 4. **Differentiated Filtering**: Exclude large order indicators for stocks with daily average turnover in the top 1/3 of the market to mitigate algorithmic trading interference[49] 5. **Normalization and Aggregation**: Standardize the six derived indicators and aggregate them to form the IRCF factor[49] **Evaluation**: The IRCF factor demonstrates strong predictive power and stability, effectively capturing micro-level trading dynamics[48][49][52] - **Model Name**: Context-Feature Factor System **Construction Idea**: This framework integrates market context and behavioral features to enhance signal precision and reduce noise[59][60] **Construction Process**: 1. **Context Definition**: Classify market states based on stock price movements, trading volume, amplitude, and intraday returns[60] 2. **Behavioral Features**: Monitor small/large order buy and sell counts and amounts to track trading footprints[60] 3. **Aggregation**: Apply statistical methods (mean, standard deviation, percentile) to refine raw sequences into actionable factors[60] **Evaluation**: The system significantly improves signal reliability by aligning behavioral features with specific market contexts[59][60] Model Backtesting Results - **IRCF Factor**: - Annualized long-short return: 25.8% - Annualized long-only excess return: 9.6% - Long-only IR: 2.13 - Monthly IC mean: 7.1% - ICIR: 3.29 - IC win rate: 85.2%[50][51][52] - **Context-Feature Factor System**: - Annualized long-short return: 18.5%-23.0% (depending on specific factors) - Annualized long-only excess return: 7.8%-9.7% - Long-only IR: 1.82-2.26 - IC mean: 6.3%-7.5% - ICIR: 2.77-3.54 - IC win rate: 75.9%-85.2%[61][62][63] Quantitative Factors and Construction Methods - **Factor Name**: Small Order Sell Count Factor **Construction Idea**: Focus on retail panic during market downturns to identify reversal signals[33][38][48] **Construction Process**: 1. Calculate "small order sell count/small order buy count" during stock downturns[33][38][48] 2. Derive statistical features (mean, standard deviation, 90th percentile) over a 40-day rolling window[44][46][48] **Evaluation**: Exhibits strong predictive power in downturn scenarios, with IC mean reaching 6.3%-7.4%[38][46][48] - **Factor Name**: Large Order Buy Count Factor **Construction Idea**: Track institutional accumulation during market upturns[37][48][49] **Construction Process**: 1. Calculate "large order buy count/large order sell count" during stock upturns[37][48][49] 2. Derive statistical features (mean, standard deviation, 90th percentile) over a 40-day rolling window[44][45][48] **Evaluation**: Demonstrates balanced performance across different market conditions, with IC mean around 5.1%-6.4%[37][45][48] Factor Backtesting Results - **Small Order Sell Count Factor**: - Annualized long-short return: 7.0%-7.8% - Annualized long-only excess return: 3.5%-7.8% - IC mean: 6.3%-7.4% - ICIR: 2.77-3.54[38][46][48] - **Large Order Buy Count Factor**: - Annualized long-short return: 5.9%-6.3% - Annualized long-only excess return: 2.3%-6.3% - IC mean: 5.1%-6.4% - ICIR: 2.10-3.13[37][45][48]