海量Level 2数据因子挖掘系列（一）-安宁宁-专题-2024-07-15

Quantitative Models and Construction Methods Model Name: Large Order Ratio Factor - Construction Idea: The factor is based on Level 2 tick-by-tick order data, focusing on the ratio of large orders to total orders[2][10][24] - Construction Process: - Aggregate tick-by-tick order data to reflect the original order size[24] - Define large orders as those with transaction volumes greater than the mean plus N standard deviations[27] - Construct large buy order ratio factors (BigBuy_1p0, BigBuy_1p5, BigBuy_2p0) and large sell order ratio factors (BigSell_1p0, BigSell_1p5, BigSell_2p0) using different standard deviation thresholds[27] - Formula: $ \text{BigBuy}_1p0 = \frac{\text{Large Buy Orders}}{\text{Total Orders}} $[27] - Evaluation: The factor shows significant performance in predicting stock returns, especially in short-term trading windows[28][30] - Test Results: - 5-day window: RankIC mean 5.4%, win rate 64%, annualized return 33.15%, max drawdown 13.43%, Sharpe ratio 1.83[31] - 20-day window: RankIC mean 7.9%, win rate 70%, annualized return 28.08%, max drawdown 9.18%, Sharpe ratio 1.80[33] Model Name: Time-Dimension Decoupled Large Order Ratio Factor - Construction Idea: Decouple large order ratio factors based on different time periods within the trading day[10][38] - Construction Process: - Define large orders within specific time windows (e.g., first 15 minutes, first 30 minutes, last 15 minutes, last 30 minutes)[38] - Construct factors such as BigBuy_1p0_09301000 for large buy orders in the first 30 minutes[38] - Formula: $ \text{BigBuy}_1p0_09301000 = \frac{\text{Large Buy Orders (09:30-10:00)}}{\text{Total Orders}} $[38] - Evaluation: Time-dimension decoupled factors provide more stable returns and lower drawdowns compared to non-decoupled factors[40][42] - Test Results: - 5-day window: RankIC mean 3.4%, win rate 63%, annualized return 29.78%, max drawdown 13.02%, Sharpe ratio 1.63[44] - 20-day window: RankIC mean 4.5%, win rate 70%, annualized return 26.63%, max drawdown 8.37%, Sharpe ratio 1.73[48] Model Name: Order-Dimension Decoupled Large Order Ratio Factor - Construction Idea: Decouple large order ratio factors based on the buy and sell attributes of orders[10][56] - Construction Process: - Define factors based on combinations of large and small buy and sell orders (e.g., BigBuy_BigSell_1p0, BigBuy_SmallSell_1p0)[56] - Formula: $ \text{BigBuy_BigSell}_1p0 = \frac{\text{Large Buy Orders + Large Sell Orders}}{\text{Total Orders}} $[56] - Evaluation: Order-dimension decoupled factors show improved performance in multi-directional trading strategies[57][60] - Test Results: - 5-day window: RankIC mean 6.7%, win rate 72%, annualized return 31.01%, max drawdown 15.44%, Sharpe ratio 1.67[60] - 20-day window: RankIC mean 9.8%, win rate 82%, annualized return 28.57%, max drawdown 11.53%, Sharpe ratio 1.80[62] Model Name: Multi-Dimension Decoupled Large Order Ratio Factor - Construction Idea: Combine time and order dimension decoupling to construct more refined large order ratio factors[10][68] - Construction Process: - Define factors based on both time and order attributes (e.g., BigBuy_BigSell_1p0_09301000, BigBuy_BigSell_1p0_14301457)[68] - Formula: $ \text{BigBuy_BigSell}_1p0_09301000 = \frac{\text{Large Buy Orders + Large Sell Orders (09:30-10:00)}}{\text{Total Orders}} $[68] - Evaluation: Multi-dimension decoupled factors provide the most stable and high-performing returns across different market conditions[69][71] - Test Results: - 5-day window: RankIC mean 3.3%, win rate 66%, annualized return 29.02%, max drawdown 14.70%, Sharpe ratio 1.56[72] - 20-day window: RankIC mean 4.9%, win rate 75%, annualized return 26.56%, max drawdown 9.82%, Sharpe ratio 1.68[76] Model Backtest Results Large Order Ratio Factor - 5-day window: RankIC mean 5.4%, win rate 64%, annualized return 33.15%, max drawdown 13.43%, Sharpe ratio 1.83[31] - 20-day window: RankIC mean 7.9%, win rate 70%, annualized return 28.08%, max drawdown 9.18%, Sharpe ratio 1.80[33] Time-Dimension Decoupled Large Order Ratio Factor - 5-day window: RankIC mean 3.4%, win rate 63%, annualized return 29.78%, max drawdown 13.02%, Sharpe ratio 1.63[44] - 20-day window: RankIC mean 4.5%, win rate 70%, annualized return 26.63%, max drawdown 8.37%, Sharpe ratio 1.73[48] Order-Dimension Decoupled Large Order Ratio Factor - 5-day window: RankIC mean 6.7%, win rate 72%, annualized return 31.01%, max drawdown 15.44%, Sharpe ratio 1.67[60] - 20-day window: RankIC mean 9.8%, win rate 82%, annualized return 28.57%, max drawdown 11.53%, Sharpe ratio 1.80[62] Multi-Dimension Decoupled Large Order Ratio Factor - 5-day window: RankIC mean 3.3%, win rate 66%, annualized return 29.02%, max drawdown 14.70%, Sharpe ratio 1.56[72] - 20-day window: RankIC mean 4.9%, win rate 75%, annualized return 26.56%, max drawdown 9.82%, Sharpe ratio 1.68[76] Selected Large Order Factor Portfolio Performance All Market - Annualized Return: 36.61%[89] - Max Drawdown: 17.52%[89] - Sharpe Ratio: 2.03[89] - Excess Annualized Return: 33.07%[89] CSI 300 - Annualized Return: 12.24%[97] - Max Drawdown: 14.51%[97] - Sharpe Ratio: 0.75[97] - Excess Annualized Return: 13.40%[97] CSI 500 - Annualized Return: 22.55%[103] - Max Drawdown: 9.08%[103] - Sharpe Ratio: 1.12[103] - Excess Annualized Return: 18.67%[103] CSI 800 - Annualized Return: 18.54%[114] - Max Drawdown: 7.22%[114] - Sharpe Ratio: 1.14[114] - Excess Annualized Return: 18.95%[114] CSI 1000 - Annualized Return: 24.61%[123] - Max Drawdown: 10.43%[123] - Sharpe Ratio: 1.36[123] - Excess Annualized Return: 17.39%[123] ChiNext - Annualized Return: 36.20%[131] - Max Drawdown: 25.15%[131] - Sharpe Ratio: 1.59[131] - Excess Annualized Return: 25.07%[131]