深度学习因子2月超额1.50%，本周热度变化最大行业为钢铁、环保：市场情绪监控周报（20260224-20260227）-20260301

Quantitative Models and Construction 1. Model Name: DecompGRU-based Multi-Stock Long Portfolio - Model Construction Idea: The model is based on the DecompGRU architecture, which integrates trend decomposition with end-to-end time-series and cross-sectional modeling for stock scoring[8] - Model Construction Process: - Select the top 200 stocks with the highest integrated scores from the DecompGRU model - Equally weight the selected stocks in the portfolio - Rebalance weekly on the first trading day of the week based on updated factor values from the previous Friday's close - Exclude stocks that are limit-up, limit-down, or suspended from trading - The stock selection universe is the CSI All Share Index - No transaction costs are considered - Benchmark: CSI All Share Equal Weight Index[8] - Model Evaluation: The model demonstrates strong performance in terms of cumulative returns and risk-adjusted metrics[10] 2. Model Name: DecompGRU-based ETF Rotation Portfolio - Model Construction Idea: The model aggregates individual stock scores from the DecompGRU model to construct an ETF rotation strategy[11] - Model Construction Process: - ETF pool includes industry and thematic ETFs - Retain only the ETF with the highest 5-day average trading volume if multiple ETFs track the same index - ETFs must meet liquidity criteria: 5-day average daily turnover > 20 million CNY and 20-day average daily turnover > 10 million CNY - Weekly rebalancing with 2-6 ETFs in the portfolio - Benchmark: Wind Thematic ETF Index[11] - Model Evaluation: The model shows robust cumulative returns and moderate drawdowns, indicating its effectiveness in capturing ETF-level opportunities[13] --- Model Backtesting Results 1. DecompGRU-based Multi-Stock Long Portfolio - Cumulative Absolute Return: 84.37%[10] - Cumulative Excess Return (vs. CSI All Share Equal Weight Index): 43.11%[10] - Maximum Drawdown: 10.08%[10] - Weekly Win Rate: 65.96%[10] - Monthly Win Rate: 100%[10] - February 2026 Absolute Return: 5.41%[10] - February 2026 Excess Return: 1.50%[10] 2. DecompGRU-based ETF Rotation Portfolio - Cumulative Absolute Return: 53.40%[13] - Cumulative Excess Return (vs. Wind Thematic ETF Index): 17.47%[13] - Maximum Drawdown: 7.82%[13] - Weekly Win Rate: 62.50%[13] - Monthly Win Rate: 72.73%[14] - February 2026 Absolute Return: 9.51%[14] - February 2026 Excess Return: 8.18%[14] --- Quantitative Factors and Construction 1. Factor Name: Sentiment Heat Factor - Factor Construction Idea: The factor is derived from user behavior data (e.g., browsing, watchlist additions, and clicks) to measure stock-level sentiment heat, which is aggregated to broader indices, industries, and concepts[15] - Factor Construction Process: - Calculate the total heat for each stock as the sum of browsing, watchlist, and click counts - Normalize the total heat as a percentage of the market total and multiply by 10,000 - Aggregate normalized heat values to broader levels (e.g., indices, industries, concepts) to derive sentiment heat proxies[15] - Factor Evaluation: The factor effectively captures sentiment-driven mispricing and behavioral biases, particularly at the stock level[15] 2. Factor Name: Broad-Based Index Heat Momentum - Factor Construction Idea: Constructed using the weekly heat change rate (smoothed with a 2-week moving average) of major broad-based indices (CSI 300, CSI 500, CSI 1000, CSI 2000, and others)[16][19] - Factor Construction Process: - Group all A-shares into CSI 300, CSI 500, CSI 1000, CSI 2000, and "Others" - Sum the sentiment heat for each group - Calculate the weekly heat change rate for each group and smooth using a 2-week moving average (MA2)[16][19] - Factor Evaluation: The factor is used to construct a rotation strategy, which has shown positive annualized returns since 2017[24] 3. Factor Name: Concept Heat Momentum - Factor Construction Idea: Similar to the Broad-Based Index Heat Momentum factor, this factor tracks weekly heat change rates for various market concepts[35] - Factor Construction Process: - Calculate the weekly heat change rate for each concept - Construct two portfolios: - High Heat Portfolio (TOP): Select the top 5 concepts with the highest heat change rates, exclude the bottom 20% of stocks by market cap, and equally weight the top 10 stocks by total heat within each concept - Low Heat Portfolio (BOTTOM): Select the same concepts but equally weight the bottom 10 stocks by total heat within each concept[40] - Factor Evaluation: The BOTTOM portfolio has historically delivered consistent excess returns, with an annualized return of 15.71% and a maximum drawdown of 28.89%[41] --- Factor Backtesting Results 1. Sentiment Heat Factor - Broad-Based Index Heat Momentum Strategy Annualized Return (2017-2026): 8.74%[24] - Maximum Drawdown: 23.5%[24] - 2026 YTD Return: 5.6%[24] 2. Concept Heat Momentum Factor - BOTTOM Portfolio Annualized Return: 15.71%[41] - BOTTOM Portfolio Maximum Drawdown: 28.89%[41] - 2026 YTD Return (BOTTOM Portfolio): 1.4%[41]