金融工程专题报告：股票涨跌情境中机构与散户的逆向资金流

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]