Quantitative Models and Construction GAN_GRU Model - **Model Name**: GAN_GRU - **Construction Idea**: The GAN_GRU model combines Generative Adversarial Networks (GAN) for feature generation and Gated Recurrent Units (GRU) for time-series feature encoding to create a stock selection factor[4][13][22] - **Construction Process**: 1. **GAN Component**: - **Generator**: Generates realistic data samples from random noise using the loss function: $$L_{G}\,=\,-\mathbb{E}_{z\sim P_{z}(z)}[\log(D(G(z)))]$$ where \(z\) represents random noise, \(G(z)\) is the generated data, and \(D(G(z))\) is the discriminator's output probability that the generated data is real[24][25][26] - **Discriminator**: Distinguishes real data from generated data using the loss function: $$L_{D}=-\mathbb{E}_{x\sim P_{data}(x)}[\log\!D(x)]-\mathbb{E}_{z\sim P_{z}(z)}[\log(1-D(G(z)))]$$ where \(x\) is real data, \(D(x)\) is the discriminator's output probability for real data, and \(D(G(z))\) is the output probability for generated data[27][29][30] - **Training Process**: Alternating training of the generator and discriminator until convergence[30][34] 2. **GRU Component**: - Two GRU layers (GRU(128,128)) followed by an MLP (256,64,64) to encode time-series features and predict future returns[22] - Input features include 18 price-volume metrics (e.g., closing price, turnover rate) sampled over 40 days to predict cumulative returns for the next 20 trading days[14][18][19] - Data preprocessing involves outlier removal, normalization, and cross-sectional standardization[18] - Training uses semi-annual rolling windows with hyperparameters such as batch size equal to the number of stocks, Adam optimizer, learning rate of \(1e-4\), and IC-based loss function[18][22] 3. **Feature Generation**: - GAN's generator processes raw price-volume time-series features (Input_Shape=(40,18)) and outputs transformed features with preserved time-series properties[37] - **Evaluation**: The model effectively combines GAN's feature generation capabilities with GRU's time-series encoding, providing robust predictive power for stock selection[4][22][37] --- Model Backtesting Results GAN_GRU Model Performance Metrics - **IC Mean**: 0.1095*** - **ICIR (Non-Annualized)**: 0.88 - **Turnover Rate**: 0.82X - **Recent IC**: 0.1008*** - **One-Year IC Mean**: 0.0514*** - **Annualized Return**: 36.03% - **Annualized Volatility**: 21.87% - **IR**: 1.55 - **Max Drawdown**: 27.29% - **Annualized Excess Return**: 21.87%[41][42][45] Industry-Level Performance - **Top 5 Industries by Recent IC**: - Media: 0.4279*** - Coal: 0.2355*** - Retail: 0.2003*** - Food & Beverage: 0.1701*** - Chemicals: 0.1395***[41][42][45] - **Top 5 Industries by One-Year IC Mean**: - Media: 0.1304*** - Steel: 0.1212*** - Retail: 0.1191*** - IT: 0.1064*** - Food & Beverage: 0.0988***[41][42][45] - **Top 5 Industries by Recent Excess Return**: - Media: 4.57% - Agriculture: 3.26% - Construction Materials: 3.19% - Light Manufacturing: 2.53% - Coal: 2.22%[45][46][48] - **Top 5 Industries by One-Year Average Excess Return**: - Real Estate: 1.83% - Retail: 1.41% - Consumer Services: 1.39% - Automotive: 1.18% - Utilities: 1.07%[45][46][48] --- Quantitative Factors and Construction GAN_GRU Factor - **Factor Name**: GAN_GRU - **Construction Idea**: Derived from the GAN_GRU model, this factor leverages GAN for feature generation and GRU for time-series encoding to predict stock returns[4][13][22] - **Construction Process**: - Input features include 18 price-volume metrics sampled over 40 days[14][18][19] - GAN generates transformed features while preserving time-series properties[37] - GRU encodes these features and outputs predicted returns as the factor[22][37] - Factor values undergo industry and market-cap neutralization and standardization[22] - **Evaluation**: The factor demonstrates strong predictive power across multiple industries and time periods, with significant IC values and excess returns[4][22][37] --- Factor Backtesting Results GAN_GRU Factor Performance Metrics - **IC Mean**: 0.1095*** - **ICIR (Non-Annualized)**: 0.88 - **Turnover Rate**: 0.82X - **Recent IC**: 0.1008*** - **One-Year IC Mean**: 0.0514*** - **Annualized Return**: 36.03% - **Annualized Volatility**: 21.87% - **IR**: 1.55 - **Max Drawdown**: 27.29% - **Annualized Excess Return**: 21.87%[41][42][45] Industry-Level Performance - **Top 5 Industries by Recent IC**: - Media: 0.4279*** - Coal: 0.2355*** - Retail: 0.2003*** - Food & Beverage: 0.1701*** - Chemicals: 0.1395***[41][42][45] - **Top 5 Industries by One-Year IC Mean**: - Media: 0.1304*** - Steel: 0.1212*** - Retail: 0.1191*** - IT: 0.1064*** - Food & Beverage: 0.0988***[41][42][45] - **Top 5 Industries by Recent Excess Return**: - Media: 4.57% - Agriculture: 3.26% - Construction Materials: 3.19% - Light Manufacturing: 2.53% - Coal: 2.22%[45][46][48] - **Top 5 Industries by One-Year Average Excess Return**: - Real Estate: 1.83% - Retail: 1.41% - Consumer Services: 1.39% - Automotive: 1.18% - Utilities: 1.07%[45][46][48]

Quantitative Models and Factor Construction Factor Name: DELTAROE - **Construction Idea**: Measures the change in return on equity (ROE) over a specific period, reflecting the company's profitability dynamics[6][19][46] - **Construction Process**: - Calculate the difference in ROE between the current period and the previous period - Formula: $\Delta ROE = ROE_{current} - ROE_{previous}$[19][46] - **Evaluation**: Demonstrates strong performance across multiple indices, indicating its effectiveness in capturing profitability trends[6][46][49] Factor Name: Standardized Unexpected Earnings (SUE) - **Construction Idea**: Quantifies the deviation of actual earnings from expected earnings, standardized by the standard deviation of forecast errors[19] - **Construction Process**: - Formula: $SUE = \frac{E_{actual} - E_{expected}}{\sigma_{forecast\ errors}}$ - Where $E_{actual}$ is the actual earnings, $E_{expected}$ is the consensus forecast, and $\sigma_{forecast\ errors}$ is the standard deviation of forecast errors[19] - **Evaluation**: Effective in identifying earnings surprises, with positive performance in various market conditions[6][19] Factor Name: Trend - **Construction Idea**: Captures momentum by comparing short-term and long-term exponential weighted moving averages (EWMA) of stock prices[14] - **Construction Process**: - Two variations: - $Trend_{120} = \frac{EWMA_{halflife=20}}{EWMA_{halflife=120}}$ - $Trend_{240} = \frac{EWMA_{halflife=20}}{EWMA_{halflife=240}}$ - Where $EWMA$ is the exponentially weighted moving average with specified half-life[14] - **Evaluation**: Exhibits strong positive returns, indicating market preference for momentum strategies[9][11] Factor Name: Liquidity - **Construction Idea**: Measures stock liquidity through turnover rates and regression-based liquidity beta[14] - **Construction Process**: - Average log turnover over 243 days - Regression of individual stock turnover against market turnover to derive liquidity beta[14] - **Evaluation**: Mixed performance, with sensitivity to market conditions[10][14] Factor Name: Volatility - **Construction Idea**: Captures price fluctuations using various measures of historical volatility[14] - **Construction Process**: - Standard deviation of daily returns over 243 days - Range-based volatility: $Range = \frac{High_{243} - Low_{243}}{Low_{243}}$ - Maximum and minimum return averages over six days within 243 days[14] - **Evaluation**: Underperforms in high-volatility environments, indicating reduced investor appetite for risk[10][14] --- Factor Backtesting Results DELTAROE - **Performance**: - CSI 300: Weekly return 0.41%, monthly return 1.59%[6][22] - CSI 500: Weekly return 0.95%, monthly return 1.19%[6][26] - CSI 800: Weekly return 1.08%, monthly return 1.62%[6][30] - CSI 1000: Weekly return 1.79%, monthly return 1.54%[6][34] - CNI 2000: Weekly return 4.71%, monthly return 7.07%[6][37] - CSI All Share: Weekly return 1.84%, monthly return 2.41%[6][46] Standardized Unexpected Earnings (SUE) - **Performance**: - CSI 500: Weekly return 1.20%, monthly return 1.59%[6][26] - CSI 800: Weekly return 0.34%, monthly return 0.98%[6][30] - CNI 2000: Weekly return 1.65%, monthly return 14.68%[6][37] - CSI All Share: Weekly return 0.96%, monthly return 1.14%[6][46] Trend - **Performance**: - Weekly return 1.15%, monthly return 4.58%, annualized return 19.73%[9][11] Liquidity - **Performance**: - Weekly return -0.43%, monthly return -3.25%, annualized return 37.53%[10][11] Volatility - **Performance**: - Weekly return -0.95%, monthly return -2.71%, annualized return 31.87%[10][11] --- MFE Portfolio Construction - **Optimization Model**: - Objective: Maximize single-factor exposure - Formula: $\begin{array}{ll}max&f^{T}w\\ s.t.&s_{l}\leq X(w-w_{b})\leq s_{h}\\ &h_{l}\leq H(w-w_{b})\leq h_{h}\\ &w_{l}\leq w-w_{b}\leq w_{h}\\ &b_{l}\leq B_{b}w\leq b_{h}\\ &0\leq w\leq l\\ &1^{T}w=1\\ &\Sigma|w-w_{0}|\leq to_{h}\end{array}$[61][62] - Constraints: - Style and industry exposure limits - Stock weight deviation limits - Turnover rate limits[64][65] - **Backtesting**: - Monthly rebalancing - Transaction cost: 0.3% per side - Metrics: Historical returns, risk-adjusted performance[65]