开源量化评论（94）：从高频股价形态到追涨杀跌因子

• Model Name: Full-period Momentum Reversal Factor; Model Construction Idea: Identify momentum reversal patterns in stock prices to measure the proportion of retail investor trading and construct a negative alpha factor[2][10] - Model Construction Process: 1. Select 1-minute frequency trading data for the past 20 trading days[12] 2. Calculate the excess returns for the current minute at time t and the next minute at time t+1, denoted as $R_t$ and $R_{t+1}$ respectively[12] 3. Compute the cosine similarity between $R_t$ and $R_{t+1}$ as the factor value[12] 4. Formula for cosine similarity: $\rho = \frac{\sum_{t=1}^{n-1} R_t R_{t+1}}{\sqrt{\sum_{t=1}^{n-1} R_t^2 \sum_{t=1}^{n-1} R_{t+1}^2}}$[14] - Model Evaluation: The factor constructed using excess returns shows better performance compared to using absolute returns, with a monthly RankIC improvement rate of 64%[15] - Model Name: End-of-day Momentum Factor; Model Construction Idea: Decompose the full-period momentum reversal factor by direction and time, and identify the end-of-day momentum factor as a stronger source of negative alpha[20][21] - Model Construction Process: 1. Select 1-minute frequency trading data for the last 90 minutes of each trading day over the past 20 trading days[27] 2. Calculate the excess returns for each minute, selecting those with positive excess returns[27] 3. Compute the cosine similarity between the positive excess return minute and the next minute's excess return as the factor value[27] 4. Formula for cosine similarity: $\rho = \frac{\sum_{t=1}^{n-1} R_t R_{t+1}}{\sqrt{\sum_{t=1}^{n-1} R_t^2 \sum_{t=1}^{n-1} R_{t+1}^2}}$[14] - Model Evaluation: The end-of-day momentum factor shows a RankIC mean of -9.2% and a RankICIR of -4.04, with a strong negative alpha[28] - Model Name: End-of-day Momentum Deviation Factor; Model Construction Idea: Measure the difference between end-of-day and full-period momentum characteristics[4] - Model Construction Process: 1. Subtract the full-period momentum factor from the end-of-day momentum factor to obtain the deviation factor[4] - Model Evaluation: The deviation factor shows a RankIC mean of -4.9% and a RankICIR of -4.1, with a stable performance[57] - Model Name: End-of-day Momentum Autoregressive Coefficient Factor; Model Construction Idea: Replace the cosine similarity in the end-of-day momentum factor with the slope of the autoregressive coefficient[5] - Model Construction Process: 1. Select the last 30 minutes of excess returns and calculate the slope of the autoregressive coefficient as the factor value[5] - Model Evaluation: The autoregressive coefficient factor shows a RankIC mean of -4.4% and a RankICIR of -3.3, with a stable performance[68] Factor Backtesting Results - Full-period Momentum Reversal Factor: - IC Mean: -4.95% - ICIR: -2.11 - RankIC Mean: -7.93% - RankICIR: -2.98 - Annualized Return: 20.13% - Annualized Volatility: 8.13% - Annualized IR: 2.48 - Maximum Drawdown: 13.74% - Monthly Win Rate: 72.95%[19] - End-of-day Momentum Factor: - RankIC Mean: -9.2% - RankICIR: -4.04 - Annualized Return: 25.3% - Annualized IR: 3.4 - Maximum Drawdown: 7.6% - Monthly Win Rate: 81%[28][31] - End-of-day Momentum Deviation Factor: - RankIC Mean: -4.9% - RankICIR: -4.1 - Annualized Return: 14.1% - Annualized IR: 3.3 - Maximum Drawdown: 7.5% - Monthly Win Rate: 82%[57][62] - End-of-day Momentum Autoregressive Coefficient Factor: - RankIC Mean: -4.4% - RankICIR: -3.3 - Annualized Return: 10.8% - Annualized IR: 2.7 - Maximum Drawdown: 6.2% - Monthly Win Rate: 79%[68][72]