Quantitative Models and Construction Methods Model Name: DBD-GRU Model - **Model Construction Idea**: Combining factor slicing theory with deep learning to enhance information extraction and prediction capabilities[4][25] - **Model Construction Process**: - Assume input data x contains features A and B, where feature A is the slicing indicator[4] - Use the median of feature A in the time series as the threshold to construct two masks: mask_Ahigh and mask_Alow[4] - Input the masked data into two branch networks: GRU_high and GRU_low[4] - Take the difference between the outputs of the last time step of the two networks as the input to the output layer[4] - Formula: $ \text{DBD-GRU} = \text{GRU}_\text{high} - \text{GRU}_\text{low} $[4] - **Model Evaluation**: The DBD-GRU model provides significant information increment compared to the original slicing theory factors and baseline GRU model factors[5][55] - **Model Testing Results**: - Ideal Amplitude-DBD RankIC: -10.33%[5] - Ideal Reversal-DBD RankIC: -10.31%[5] - Active Buy-Sell-DBD RankIC: -9.81%[5] Quantitative Factors and Construction Methods Factor Name: Ideal Reversal Factor - **Factor Construction Idea**: Improve traditional reversal factors by slicing the time series data based on transaction amount[14][15] - **Factor Construction Process**: - Retrieve past 20 days of data for selected stocks[17] - Calculate the average transaction amount per trade for each day[17] - Sum the price changes for the 10 days with the highest transaction amounts, denoted as M_high[18] - Sum the price changes for the 10 days with the lowest transaction amounts, denoted as M_low[18] - Ideal Reversal Factor M = M_high - M_low[18] - **Factor Evaluation**: The slicing process effectively distinguishes between strong and weak reversal periods, enhancing the factor's stability and predictive power[24] - **Factor Testing Results**: - RankIC: -6.06%[40] - RankICIR: -2.98[40] - Annualized long-short return: 24.26%[40] - Annualized long-short volatility: 9.38%[40] - Long-short return volatility ratio: 2.59[40] - Maximum long-short drawdown: 7.80%[40] - Monthly win rate: 78.57%[40] Factor Name: Ideal Amplitude Factor - **Factor Construction Idea**: Measure the difference in amplitude information between high and low price states of stocks[78] - **Factor Construction Process**: - Retrieve past 20 days of data for selected stocks[80] - Calculate the daily amplitude (highest price/lowest price - 1)[80] - Calculate the average amplitude for the 25% of days with the highest closing prices, denoted as Amplitude_high[80] - Calculate the average amplitude for the 25% of days with the lowest closing prices, denoted as Amplitude_low[80] - Ideal Amplitude Factor = Amplitude_high - Amplitude_low[80] - **Factor Evaluation**: The factor effectively captures the difference in amplitude information between high and low price states, providing a stable and predictive measure[24] - **Factor Testing Results**: - RankIC: -7.00%[40] - RankICIR: -3.47[40] - Annualized long-short return: 21.02%[40] - Annualized long-short volatility: 10.53%[40] - Long-short return volatility ratio: 2.00[40] - Maximum long-short drawdown: 17.67%[40] - Monthly win rate: 76.98%[40] Factor Name: Active Buy-Sell Factor - **Factor Construction Idea**: Measure retail investors' trading behavior in a declining market environment[79] - **Factor Construction Process**: - Retrieve past 20 days of data for selected stocks[79] - Calculate the daily stock price change and small order inflow intensity[79] - Formula for small order inflow intensity: $ \frac{\text{Active Buy Amount (small orders)} - \text{Active Sell Amount (small orders)}}{\text{Active Buy Amount (small orders)} + \text{Active Sell Amount (small orders)}} $[79] - Calculate the average small order inflow intensity for the 25% of days with the lowest closing prices to obtain the Active Buy-Sell Factor[79] - **Factor Evaluation**: The factor effectively captures retail investors' trading behavior in a declining market, providing a stable and predictive measure[24] - **Factor Testing Results**: - RankIC: -3.39%[40] - RankICIR: -1.27[40] - Annualized long-short return: 10.20%[40] - Annualized long-short volatility: 12.57%[40] - Long-short return volatility ratio: 0.81[40] - Maximum long-short drawdown: 25.26%[40] - Monthly win rate: 70.63%[40] Factor Backtesting Results DBD-GRU Model Factors - **Ideal Amplitude-DBD**: - RankIC: -10.33%[47] - RankICIR: -3.68[47] - Annualized long-short return: 34.31%[52] - Annualized long-short volatility: 15.17%[52] - Long-short return volatility ratio: 2.26[52] - Maximum long-short drawdown: 17.98%[52] - Monthly win rate: 76.98%[52] - **Ideal Reversal-DBD**: - RankIC: -10.31%[47] - RankICIR: -3.57[47] - Annualized long-short return: 37.62%[52] - Annualized long-short volatility: 12.55%[52] - Long-short return volatility ratio: 3.00[52] - Maximum long-short drawdown: 8.96%[52] - Monthly win rate: 80.95%[52] - **Active Buy-Sell-DBD**: - RankIC: -9.81%[47] - RankICIR: -3.63[47] - Annualized long-short return: 33.33%[52] - Annualized long-short volatility: 13.32%[52] - Long-short return volatility ratio: 2.50[52] - Maximum long-short drawdown: 13.82%[52] - Monthly win rate: 75.40%[52] DBD-Combine Factor - **Performance in Major Broad-Based Indices**: - **CSI 300**: - RankIC: -5.76%[63] - RankICIR: -1.87[61] - Annualized long-short return: 14.9%[63] - Annualized excess return: 7.64%[67] - Excess IR: 1.84[61] - Maximum excess drawdown: 3.37%[61] - **CSI 500**: - RankIC: -7.40%[68] - RankICIR: -2.58[65] - Annualized long-short return: 17.5%[69] - Annualized excess return: 7.23%[70] - Excess IR: 1.37[65] - Maximum excess drawdown: 6.43%[65] - **CSI 1000**: - RankIC: -9.84%[75] - RankICIR: -3.48[71] - Annualized long-short return: 30.8%[72] - Annualized excess return: 11.8%[76] - Excess IR: 2.21[71] - Maximum excess drawdown: 3.94%[71]

Quantitative Factors and Construction Methods 1. Factor Name: Original ACT Factor - **Construction Idea**: The ACT factor measures the "active net buying intensity" by comparing the active buying amount with the active selling amount, reflecting the future price movement expectations of different types of traders [13][14] - **Construction Process**: - The ACT factor is calculated as follows: $$ \mathrm{ACT} = \frac{\text{Active Buy Amount} - \text{Active Sell Amount}}{\text{Active Buy Amount} + \text{Active Sell Amount}} $$ where "Active Buy Amount" and "Active Sell Amount" represent the total amounts of active buying and selling transactions, respectively [13] - The factor value for each stock is computed as the average ACT value over the past 20 trading days at the end of each month [14] - Stocks with insufficient trading days, suspension, or ST status are excluded from the sample [14] - **Evaluation**: The IC values of the original ACT factor are low, and its stock selection ability does not meet expectations, which has led to reduced research interest in this factor in recent years [14] 2. Factor Name: ACT Factor with Segmentation - **Construction Idea**: The segmentation method divides the ACT factor based on different market conditions (e.g., high-return and low-return days) to better capture the nuanced behavior of different trader groups [6][15] - **Construction Process**: - Calculate the daily ACT value over the past 20 trading days [20] - Identify the highest-return days as "high-return days" and the lowest-return days as "low-return days" [20] - Compute the average ACT value for high-return days (ACT_high) and low-return days (ACT_low) [20] - **Evaluation**: - Large and medium-sized orders show strong positive stock selection effects on high-return days, while small orders exhibit strong negative stock selection effects on low-return days [6][20] - The segmentation approach aligns with intuition, as institutional investors dominate large and medium orders, while retail investors dominate small orders [21] 3. Factor Name: ACT Positive and ACT Negative Factors - **Construction Idea**: Based on the segmentation analysis, ACT Positive focuses on large and medium orders with positive stock selection effects, while ACT Negative focuses on small orders with negative stock selection effects [7][24] - **Construction Process**: - ACT Positive: $$ \mathrm{ACT\ Positive} = \frac{\text{Active Buy Amount (Large + Medium Orders)} - \text{Active Sell Amount (Large + Medium Orders)}}{\text{Active Buy Amount (Large + Medium Orders)} + \text{Active Sell Amount (Large + Medium Orders)}} $$ - ACT Negative: $$ \mathrm{ACT\ Negative} = \frac{\text{Active Buy Amount (Small Orders)} - \text{Active Sell Amount (Small Orders)}}{\text{Active Buy Amount (Small Orders)} + \text{Active Sell Amount (Small Orders)}} $$ - The factor values are averaged over the past 20 trading days [26] - **Evaluation**: - ACT Positive demonstrates superior stock selection ability, particularly in smaller segmentation ratios (e.g., λ=10%), with a high return-to-volatility ratio [7][26] - ACT Negative shows stable return-to-volatility ratios but declining returns in recent years, reflecting the increasing dominance of large funds [7][31] --- Factor Backtesting Results 1. Original ACT Factor - **IC Values**: Gradually increase from negative to positive as the order size increases (small → medium → large → extra-large) [14] - **Stock Selection Ability**: Poor, as evidenced by the unsatisfactory multi-long-short net value curve [14] 2. ACT Positive Factor - **λ=10%**: - Multi-long-short return-to-volatility ratio: 3.06 - Multi-long return-to-volatility ratio: 0.87 [26][28] - **Performance in Neutralized Conditions**: - Multi-long-short return-to-volatility ratio: 2.40 - Multi-long return-to-volatility ratio: 0.66 [38] 3. ACT Negative Factor - **λ=10%**: - Multi-long-short return-to-volatility ratio: 0.82 - Multi-long return-to-volatility ratio: 0.73 [33] 4. ACT Positive Factor in Different Sample Spaces - **CSI 300**: - λ=20%: Multi-long-short return-to-volatility ratio: 1.32; Multi-long return-to-volatility ratio: 0.63 [44] - **CSI 500**: - λ=20%: Multi-long-short return-to-volatility ratio: 1.78; Multi-long return-to-volatility ratio: 0.80 [44]