多因子ALPHA系列报告之(三十四)：基于多期限的选股策略研究

Quantitative Models and Factor Construction Multi-Horizon Factor - Factor Name: Multi-Horizon Factor - Construction Idea: This factor captures short-term reversal, medium-term momentum, and long-term reversal effects by analyzing moving average (MA) data across multiple time horizons [2][14][21] - Construction Process: - Calculate moving averages for different time horizons $L = [3, 5, 10, 20, 30, 60, 90, 120, 180, 240, 270, 300]$ using the formula: $A_{j t,L} = \frac{P_{j,\,d-L+1}^{t} + \cdots + P_{j,d}^{t}}{L}$ where $P_{j,d}^t$ represents the price of stock $j$ at time $t$ [21] - Standardize the moving average factor: $\tilde{A}_{j t,\,L} = \frac{A_{j t,\,L}}{P_{j}^{t}}$ [22] - Perform cross-sectional regression of stock returns on lagged standardized moving average factors: $r_{j,t} = \beta_{0,t} + \Sigma_{i}\beta_{i,t}\tilde{A}_{j t-1,L_{i}} + \epsilon_{j,t}$ [23] - Predict next-period regression coefficients by averaging the past 25 weeks' coefficients: $E\left[\beta_{i,\,t+1}\right] = \frac{1}{25}\,\sum_{m=1}^{25}\,\beta_{i,t+1-m}$ [24] - Use predicted coefficients and new factor values to estimate next-period returns: $E\left[r_{j,t+1}\right] = \Sigma_{i}\,E\left[\beta_{i,\,t+1}\right]\tilde{A}_{j t,\,L_{i}}$ [25] - Rank stocks by predicted returns and construct long-short portfolios [26] - Evaluation: The factor demonstrates strong predictive power for stock returns across different market segments, with positive IC values dominating [30][32] LLT Trend Factor - Factor Name: LLT Trend Factor - Construction Idea: To address the lagging sensitivity of MA, the LLT (Low-Lag Trendline) indicator is used as a replacement. LLT reduces delay and better captures momentum and reversal effects [14][76] - Construction Process: - LLT is calculated using a second-order linear filter with the recursive formula: $LLT = \begin{cases} P(T), & T=1,2 \\ (2-2\alpha)LLT(T-1) - (1-\alpha)^2LLT(T-2) + \left(\alpha-\frac{\alpha^2}{4}\right)P(T) \\ + \left(\frac{\alpha^2}{2}\right)P(T-1) - \left(\alpha-\frac{3}{4}\alpha^2\right)P(T-2), & \text{else} \end{cases}$ where $\alpha = \frac{2}{1+N}$ and $N$ is the smoothing parameter [76] - Replace MA with LLT in the multi-horizon factor construction process [76] - Evaluation: LLT-based factors outperform MA-based factors in terms of IC mean, positive IC ratio, and predictive power for asset returns [82][84] --- Backtesting Results Multi-Horizon Factor - Annualized Return: 25.40% [3][48] - Annualized Volatility: 14.12% [48] - Maximum Drawdown: 13.31% [48] - IR: 1.81 [48] LLT Trend Factor - Annualized Return: 29.58% [4][103] - Annualized Volatility: 10.46% [103] - Maximum Drawdown: 11.57% [103] - IR: 2.51 [103]