序列相似度的应用：DTW预期收益率因子

Quantitative Models and Construction Methods - Model Name: DTW Expected Return Factor Model Construction Idea: The model uses the Dynamic Time Warping (DTW) algorithm to measure the similarity between return sequences over a specific time range. It identifies sequences with high similarity and calculates their returns as a proxy for expected future returns. This approach serves as an alternative to traditional momentum factors, especially when the number of cross-sectional assets is limited [4][30][33] Model Construction Process: 1. For each asset, take the past 20-day return sequence ( s ) [33] 2. Compute DTW distances between ( s ) and the 20-day return sequences of all assets over a rolling window of 5–120 days [33] 3. Select the top 10% of sequences with the smallest DTW distances [33] 4. Calculate the average return of these sequences over the next 5 days as the factor value [33] 5. Constraints: - Minimum lookback window of 5 days to avoid overly correlated sequences due to close time intervals [33] - Maximum lookback window of 120 days to prevent factor failure caused by market regime shifts [33] Formula: DTW cost function: $\operatorname{cost}(i,j) = d(i,j) + \operatorname*{min} \begin{cases} \operatorname{cost}(i-1,j) \\ \operatorname{cost}(i,j-1) \\ \operatorname{cost}(i-1,j-1) \end{cases}$ where ( d(i,j) ) represents the pointwise distance between two sequences [17] Model Evaluation: The DTW Expected Return Factor demonstrates strong trend-following characteristics and provides significant advantages over traditional momentum factors, especially in scenarios with fewer cross-sectional assets [4][30][33] Factor Backtesting Results - ETF Rotation Portfolio: - IC Mean: 0.034 - ICIR: 0.128 - Maximum Drawdown: 16.2% - Turnover Rate: 73% [35] - STAR 50 Constituents: - IC Mean: 0.053 - ICIR: 0.251 - Maximum Drawdown: 51.0% - Turnover Rate: 72% [40] Quantitative Factors and Construction Methods - Factor Name: DTW Expected Return Factor Factor Construction Idea: The factor leverages DTW to identify sequences with high similarity across assets and uses their subsequent returns as a proxy for expected returns. This approach captures price inertia while addressing limitations of traditional momentum factors [4][30][33] Factor Construction Process: 1. Extract the past 20-day return sequence for each asset [33] 2. Compute DTW distances between the sequence and all other assets' 20-day return sequences over a rolling window of 5–120 days [33] 3. Select the top 10% of sequences with the smallest DTW distances [33] 4. Calculate the average return of these sequences over the next 5 days as the factor value [33] Formula: DTW cost function: $\operatorname{cost}(i,j) = d(i,j) + \operatorname*{min} \begin{cases} \operatorname{cost}(i-1,j) \\ \operatorname{cost}(i,j-1) \\ \operatorname{cost}(i-1,j-1) \end{cases}$ where ( d(i,j) ) represents the pointwise distance between two sequences [17] Factor Evaluation: The factor exhibits strong trend-following characteristics and performs well in scenarios with fewer cross-sectional assets. It also shows robustness when combined with trend-based timing strategies [4][30][33] Factor Backtesting Results - ETF Rotation Portfolio: - IC Mean: 0.034 - ICIR: 0.128 - Maximum Drawdown: 16.2% - Turnover Rate: 73% [35] - STAR 50 Constituents: - IC Mean: 0.053 - ICIR: 0.251 - Maximum Drawdown: 51.0% - Turnover Rate: 72% [40]