Quantitative Models and Construction 1. Model Name: GRU - **Model Construction Idea**: The GRU model is used to predict future stock returns based on historical data and incorporates various technical and fundamental factors[3][4][5] - **Model Construction Process**: The GRU model is trained on historical data to predict future returns. It uses a recurrent neural network structure, specifically the Gated Recurrent Unit (GRU), to capture sequential dependencies in time-series data. The model is applied to different stock pools (e.g., CSI 300, CSI 500, CSI 1000) and is evaluated based on its long-short portfolio returns[3][4][5] - **Model Evaluation**: The GRU model demonstrates strong performance in predicting returns, with positive long-short portfolio returns in most cases. However, its performance varies across different stock pools and time horizons[3][4][5] 2. Model Name: Barra5d - **Model Construction Idea**: The Barra5d model predicts future returns by incorporating short-term technical factors and ensuring style neutrality[6][25] - **Model Construction Process**: The Barra5d model uses a combination of short-term technical indicators (e.g., 5-day momentum) and applies style-neutral constraints to ensure that the predictions are not biased by market-wide factors. The model is tested on various stock pools, including CSI 300, CSI 500, and CSI 1000[6][25] - **Model Evaluation**: The Barra5d model shows strong performance, particularly in the CSI 500 and CSI 1000 stock pools, with weekly long-short portfolio returns exceeding 3% in some cases[6][25] 3. Model Name: Open1d - **Model Construction Idea**: The Open1d model focuses on short-term price movements and is designed to capture immediate market reactions[19][21][23] - **Model Construction Process**: The Open1d model uses one-day price changes as its primary input and applies machine learning techniques to predict short-term returns. It is evaluated based on its ability to generate excess returns in long-short portfolios[19][21][23] - **Model Evaluation**: The Open1d model has shown strong performance year-to-date, with cumulative excess returns of 4.24% relative to the CSI 1000 index[19][21][23] --- Model Backtesting Results 1. GRU Model - Weekly long-short portfolio return: Positive in most cases, with variations across stock pools[3][4][5] - CSI 500 stock pool: Weekly long-short return > 3%[5] - CSI 1000 stock pool: Performance is mixed, with some models (e.g., Barra1d, Barra5d) performing well[6][25] 2. Barra5d Model - Weekly long-short portfolio return: > 3% in the CSI 500 stock pool[6][25] - Strong performance in the CSI 1000 stock pool, particularly in predicting style-neutral future returns[6][25] 3. Open1d Model - Year-to-date excess return: 4.24% relative to the CSI 1000 index[19][21][23] - Weekly long-short portfolio return: Mixed, with some weeks showing slight negative returns[19][21][23] --- Quantitative Factors and Construction 1. Factor Name: Beta - **Factor Construction Idea**: Measures the historical sensitivity of a stock's returns to market returns[15] - **Factor Construction Process**: Beta is calculated as the slope of the regression line between a stock's returns and market returns over a specified historical period[15] 2. Factor Name: Momentum - **Factor Construction Idea**: Captures the tendency of stocks with strong past performance to continue performing well[15] - **Factor Construction Process**: Momentum is calculated as the mean of historical excess returns over a specified period[15] 3. Factor Name: Volatility - **Factor Construction Idea**: Measures the variability of a stock's returns over time[15] - **Factor Construction Process**: $ \text{Volatility} = 0.74 \times \text{Historical Excess Return Volatility} + 0.16 \times \text{Cumulative Excess Return Deviation} + 0.1 \times \text{Residual Return Volatility} $ - Historical Excess Return Volatility: Standard deviation of excess returns - Cumulative Excess Return Deviation: Deviation of cumulative excess returns from the mean - Residual Return Volatility: Standard deviation of residual returns after removing market effects[15] 4. Factor Name: Liquidity - **Factor Construction Idea**: Measures the ease of trading a stock based on turnover rates[15] - **Factor Construction Process**: $ \text{Liquidity} = 0.35 \times \text{Monthly Turnover Rate} + 0.35 \times \text{Quarterly Turnover Rate} + 0.3 \times \text{Annual Turnover Rate} $ - Turnover Rate: Ratio of trading volume to total shares outstanding[15] --- Factor Backtesting Results 1. Beta Factor - Weekly long-short portfolio return: Strong performance in recent weeks[16] 2. Momentum Factor - Weekly long-short portfolio return: Positive for long-term momentum (e.g., 120-day), negative for short-term momentum (e.g., 20-day)[18][23][25] 3. Volatility Factor - Weekly long-short portfolio return: Positive for long-term volatility (e.g., 120-day), mixed for short-term volatility (e.g., 20-day)[18][23][25] 4. Liquidity Factor - Weekly long-short portfolio return: Strong performance, particularly in high-turnover stocks[16]

Quantitative Models and Construction 1. Model Name: GRU Model - **Model Construction Idea**: The GRU model is a machine learning-based quantitative model designed to capture complex patterns in stock price movements and factor relationships[3][19][23] - **Model Construction Process**: The GRU (Gated Recurrent Unit) model is trained on historical stock data, incorporating various financial and technical indicators as input features. The model uses a recurrent neural network structure to process sequential data, enabling it to predict stock price trends and factor performance. Specific details on the training parameters or architecture were not provided in the report[3][19][23] - **Model Evaluation**: The GRU model demonstrated strong performance in multi-factor strategies and outperformed other models in certain market conditions, particularly in the small-cap stock universe[7][29][33] 2. Model Name: Barra1d Model - **Model Construction Idea**: The Barra1d model is a factor-based quantitative model that utilizes the Barra risk model framework to analyze and predict stock returns[3][19][23] - **Model Construction Process**: The Barra1d model incorporates style factors such as size, value, momentum, and volatility, along with industry and country factors. It uses historical data to estimate factor exposures and risk premiums, which are then applied to construct long-short portfolios[3][19][23] - **Model Evaluation**: The Barra1d model experienced significant drawdowns in certain market conditions, particularly in the CSI 500 and CSI 1000 stock universes, indicating sensitivity to market volatility[5][26][29] 3. Model Name: Open1d Model - **Model Construction Idea**: The Open1d model focuses on short-term price movements and factor dynamics, leveraging daily open prices as a key input[3][19][29] - **Model Construction Process**: The Open1d model uses daily open prices and other technical indicators to construct long-short portfolios. It emphasizes short-term momentum and volatility factors to capture rapid market movements[3][19][29] - **Model Evaluation**: The Open1d model achieved strong performance, with its excess returns reaching new highs for the year, particularly in the CSI 1000 stock universe[6][29][33] 4. Model Name: Close1d Model - **Model Construction Idea**: The Close1d model is similar to the Open1d model but focuses on daily closing prices to capture end-of-day market dynamics[3][19][29] - **Model Construction Process**: The Close1d model uses daily closing prices and technical indicators to construct long-short portfolios. It emphasizes factors such as closing momentum and volatility to predict stock movements[3][19][29] - **Model Evaluation**: The Close1d model demonstrated strong performance, particularly in the CSI 1000 stock universe, with consistent positive excess returns[6][29][33] --- Model Backtesting Results 1. GRU Model - Weekly excess return: 0.46%-1.43% relative to the CSI 1000 index[7][33][34] - Year-to-date excess return: Not explicitly provided 2. Barra1d Model - Weekly excess return: 0.46%[33][34] - Year-to-date excess return: 2.10%[34] 3. Open1d Model - Weekly excess return: 1.43%[33][34] - Year-to-date excess return: 3.90%[34] 4. Close1d Model - Weekly excess return: 1.38%[33][34] - Year-to-date excess return: 1.87%[34] 5. Multi-Factor Strategy - Weekly excess return: 1.01%[33][34] - Year-to-date excess return: 2.15%[34] --- Quantitative Factors and Construction 1. Factor Name: Momentum - **Factor Construction Idea**: Momentum factors are designed to capture the tendency of stocks with strong past performance to continue performing well in the short term[15][16][19] - **Factor Construction Process**: - Historical excess return mean: $0.74 \times \text{volatility of excess returns} + 0.16 \times \text{cumulative excess return deviation} + 0.1 \times \text{residual return volatility}$[15] - **Factor Evaluation**: Momentum factors showed strong performance in the current market, particularly in high-volatility environments[3][19][23] 2. Factor Name: Valuation - **Factor Construction Idea**: Valuation factors aim to identify undervalued stocks based on financial ratios such as price-to-book (P/B) and price-to-earnings (P/E)[15][16][19] - **Factor Construction Process**: - Valuation factor: $1 / \text{P/B ratio}$[15] - **Factor Evaluation**: Valuation factors demonstrated strong performance in the small-cap stock universe, particularly in the CSI 1000 index[6][29][33] 3. Factor Name: Growth - **Factor Construction Idea**: Growth factors focus on identifying stocks with high earnings and revenue growth potential[15][16][19] - **Factor Construction Process**: - Growth factor: $0.24 \times \text{earnings growth rate} + 0.47 \times \text{revenue growth rate}$[15] - **Factor Evaluation**: Growth factors performed well across multiple stock universes, particularly in high-growth environments[3][19][23] 4. Factor Name: Volatility - **Factor Construction Idea**: Volatility factors measure the risk associated with stock price fluctuations, often used to identify low-risk investment opportunities[15][16][19] - **Factor Construction Process**: - Volatility factor: $0.74 \times \text{historical return volatility} + 0.16 \times \text{cumulative return deviation} + 0.1 \times \text{residual return volatility}$[15] - **Factor Evaluation**: Volatility factors showed mixed performance, with low-volatility stocks underperforming in certain market conditions[5][26][29] --- Factor Backtesting Results 1. Momentum Factor - Weekly excess return: 0.89%[17][19][23] - Year-to-date excess return: 42%[17][19][23] 2. Valuation Factor - Weekly excess return: 1.68%[17][19][23] - Year-to-date excess return: 1.14%[17][19][23] 3. Growth Factor - Weekly excess return: 1.20%[17][19][23] - Year-to-date excess return: 4.03%[17][19][23] 4. Volatility Factor - Weekly excess return: 0.16%[17][19][23] - Year-to-date excess return: 8.10%[17][19][23]

- The report discusses the application of stock pair trading ideas in factor strategies, specifically focusing on reversal factors which have historically shown strong performance[1] - Traditional reversal factors include "N-month price reversal," "highest price length," and "volume ratio," which capture the trend that stocks with low past returns tend to perform better in the future and vice versa[1][2] - The report introduces a pair reversal factor that captures reversal opportunities between individual stocks within the same industry, differing from traditional pair trading by using periodic closing instead of stop-loss conditions[2][3] - The pair reversal factor is tested using a hedging strategy with a monthly rebalancing frequency, using the CSI 800 index constituents as the stock pool, and achieving an annualized excess return of 8% from 2007 to 2016[3][4] - The pair reversal factor is also applied to enhance multi-factor portfolios with weekly rebalancing, showing improved returns even after considering transaction costs, with a benchmark multi-factor portfolio return of 424.40% and a pair rebalancing portfolio return of 501.59% during the sample period from 2007 to 2016[4][5] Quantitative Models and Construction Methods 1. **Model Name**: Pair Reversal Factor - **Construction Idea**: Capture reversal opportunities between individual stocks within the same industry, similar to pair trading but with periodic closing instead of stop-loss conditions[2][3] - **Construction Process**: 1. Perform cointegration regression on the log prices of two assets to check for cointegration relationship[43][44] 2. Calculate the spread and standard deviation of the spread during the learning period[45][46] 3. Use the spread and standard deviation to determine the opening threshold and execute trades accordingly[46][49] 4. Rebalance the portfolio monthly by closing all positions and reopening new ones based on the updated spread and standard deviation[51][53] - **Evaluation**: The pair reversal factor effectively captures stock price reversals and mean reversion of price spreads, providing significant excess returns at the individual stock level[69] Model Backtest Results 1. **Pair Reversal Factor**: - **Annualized Return**: 31.17% (2007), 50.85% (2008), 51.19% (2009), 21.39% (2010), 14.26% (2011), 14.75% (2012), 25.75% (2013), 9.10% (2014), 59.01% (2015), 17.05% (2016), 1246.06% (full sample)[63] - **Maximum Drawdown**: 4.44% (2007), 4.62% (2008), 4.61% (2009), 2.97% (2010), 2.64% (2011), 2.23% (2012), 2.57% (2013), 4.99% (2014), 5.48% (2015), 4.07% (2016), 5.48% (full sample)[63] - **Win Rate**: 58.38% (2007), 60.57% (2008), 59.02% (2009), 58.26% (2010), 58.20% (2011), 59.66% (2012), 59.66% (2013), 51.02% (2014), 59.84% (2015), 59.43% (2016), 58.27% (full sample)[63] Quantitative Factors and Construction Methods 1. **Factor Name**: N-month Price Reversal - **Construction Idea**: Measure the price change over a fixed time window to capture the reversal effect[30][33] - **Construction Process**: 1. Calculate the price change over the past N months: $(\text{Current Price} - \text{Price N months ago}) / \text{Price N months ago}$[33] - **Evaluation**: Reversal factors have shown strong performance in historical studies, with high IC values and good performance in various metrics such as LS return, LS win rate, LS IR, IC IR, and IC P[33][35] Factor Backtest Results 1. **N-month Price Reversal**: - **IC**: -5.72% (1-month), -4.75% (3-month), -4.10% (6-month), -3.55% (12-month)[35] - **LS Return**: 21.84% (1-month), 20.33% (3-month), 18.13% (6-month), 17.66% (12-month)[35] - **LS Win Rate**: 64.41% (1-month), 59.32% (3-month), 56.78% (6-month), 61.02% (12-month)[35] - **LS IR**: 0.99 (1-month), 0.81 (3-month), 0.77 (6-month), 0.83 (12-month)[35] - **IC IR**: 0.72 (1-month), 0.92 (3-month), 0.78 (6-month), 0.83 (12-month)[35] - **IC P**: 0.0% (1-month), 0.2% (3-month), 0.5% (6-month), 1.1% (12-month)[35]