Core Insights - The report integrates trading volume and trading time intervals to construct an IT factor that captures the "uniform order" behavior of informed traders, demonstrating strong predictive power for returns in the A-share market [5][4]. - The IT factor's construction is innovative, combining volume (V) and time intervals (Δt) to effectively identify informed trading behavior [5]. - The IT factor's predictive ability is validated with a bi-weekly RankIC average of approximately 0.064, indicating its effectiveness in forecasting stock returns [5][49]. - A long position portfolio based on the IT factor (Top 100) achieved an annualized return of 25.9%, significantly outperforming the Wind All A index [5][78]. - The IT factor can generate stable excess returns when applied to index enhancement strategies, particularly within the CSI 1000 index [5][94]. Informed Trading Behavior - Informed traders tend to exhibit "uniform order" characteristics to minimize information leakage and execution costs, leading to a dual uniformity in trading volume distribution and time intervals [7][11]. - The necessity to conceal trading intentions arises from the potential market impact of large orders, which can trigger adverse reactions from other investors [9][10]. Factor Construction and Validation - The IT factor is constructed by standardizing the trading clustering indicator, which combines both trading volume and time intervals to assess the uniformity of informed trading behavior [21][27]. - The factor's performance is robust, with a high win rate of 82.43% and a multi-strategy annualized return of 34.81% [49][50]. Portfolio Construction - The Top 100 portfolio based on the IT factor has shown consistent annualized returns, with a cumulative return of 704.3% over the backtesting period [80][81]. - The portfolio's performance is stable across different years, with most years outperforming the benchmark [82][84]. Incremental Information - The IT factor maintains significant predictive power even when controlling for liquidity and volatility factors, indicating its unique contribution to stock selection [60][68]. - The correlation between the IT factor and other level 2 factors is relatively low, suggesting that it captures distinct trading behaviors [62]. Sector and Market Capitalization Analysis - The IT factor performs particularly well in small and mid-cap stocks, as evidenced by its higher RankIC in the CSI 1000 index compared to larger indices [66][67]. - The factor's performance remains strong even after industry and market capitalization neutralization, with an improved win rate [69][75]. Application in Index Enhancement - The IT factor has demonstrated superior performance in enhancing index strategies, particularly within the CSI 1000, yielding an annualized excess return of 9.3% [94][111]. - Combining the IT factor with the BABR factor enhances the overall performance of index strategies, indicating a complementary relationship between the two factors [106][110].

Quantitative Models and Construction Methods 1. Model Name: GRU+GAT_SA Weighted Model - **Model Construction Idea**: Combines GRU for time-series information extraction and GAT for cross-sectional information extraction, with SA weighting to optimize factor performance[24][28][32] - **Model Construction Process**: - GRU is used to extract time-series features from input data, replacing LSTM due to its superior performance in terms of speed and factor effectiveness[24] - GAT is applied to capture cross-sectional relationships among stocks, using three types of networks: industry, financial, and capital flow networks[25][28] - SA weighting is introduced by adding an MLP layer during training, with Barra style factor returns as input to determine the weights of the three networks[32] - Financial indicators are incorporated at the output layer before the final factor output, enhancing multi-head performance[36] - **Model Evaluation**: The model demonstrates superior performance in multi-head returns and information ratio compared to simpler models like GRU or GAT alone[31][32] 2. Model Name: GRU+GAT_SA Weighted Model with Financial Consideration - **Model Construction Idea**: Enhances the GRU+GAT_SA Weighted Model by incorporating financial indicators to improve factor performance[36] - **Model Construction Process**: - Financial indicators are categorized into nine dimensions, including growth, profitability, quality, solvency, capital structure, turnover, goodwill, R&D, and valuation[36][37] - These indicators are standardized and integrated into the model at the output layer, combined with GRU and GAT outputs[36] - **Model Evaluation**: The inclusion of financial indicators significantly improves multi-head performance, especially in long-only strategies[36][39] --- Model Backtesting Results GRU+GAT_SA Weighted Model - **10-day RankIC**: 11.5%[44] - **Annualized RankICIR**: 5.9[44] - **Long-Short Annualized Return**: 58.3%[44] - **Long-Short Information Ratio**: 5.3[44] - **Maximum Drawdown**: -5.8%[44] - **Long-Only Annualized Return**: 22.1%[44] - **Long-Only Information Ratio**: 2.8[44] GRU+GAT_SA Weighted Model with Financial Consideration - **10-day RankIC**: 11.7%[44] - **Annualized RankICIR**: 5.7[44] - **Long-Short Annualized Return**: 58.9%[44] - **Long-Short Information Ratio**: 5.1[44] - **Maximum Drawdown**: -4.8%[44] - **Long-Only Annualized Return**: 24.1%[44] - **Long-Only Information Ratio**: 3.0[44] --- Quantitative Factors and Construction Methods 1. Factor Name: PV (Price-Volume) - **Factor Construction Idea**: Derived from basic market data, including open, high, low, close prices, average price, and trading volume[19] - **Factor Construction Process**: - Basic market data is directly used as input features for GRU+GAT_SA Weighted Model[19] - **Factor Evaluation**: Demonstrates strong performance in both long-short and long-only strategies[44] 2. Factor Name: G (Technical Indicators and K-line State Variables) - **Factor Construction Idea**: Derived from technical indicators and K-line state variables based on basic market data[19][45] - **Factor Construction Process**: - Technical indicators and K-line state variables are calculated and encoded as input features for the model[45][46] - **Factor Evaluation**: Shows incremental performance improvement when combined with financial indicators[49] 3. Factor Name: C (Capital Flow) - **Factor Construction Idea**: Based on large and small order capital flow data, including original data, derived indicators, and state variables[19][52] - **Factor Construction Process**: - Capital flow data is processed into state variables, such as net buy/sell and active buy/sell proportions, and used as input features[54][55] - **Factor Evaluation**: Outperforms traditional manually constructed capital flow factors[58] 4. Factor Name: HF (High-Frequency Features) - **Factor Construction Idea**: Derived from high-frequency data, aggregated into daily features[19][59] - **Factor Construction Process**: - High-frequency data, such as minute-level returns and trading volume, is aggregated and used as input features for the model[59] - **Factor Evaluation**: Demonstrates strong performance in long-short and long-only strategies[62] 5. Factor Name: DP (Genetic Algorithm Factors) - **Factor Construction Idea**: Derived from genetic algorithm-based factor mining, further enhanced using deep learning[19][60] - **Factor Construction Process**: - Genetic algorithm is used to generate 185 factors, of which 48 are selected based on historical performance and completeness[60] - These factors are used as input features for the GRU+GAT_SA Weighted Model[65] - **Factor Evaluation**: Deep learning significantly improves the performance of genetic algorithm factors[66] 6. Factor Name: ML_C (Composite Deep Learning Factor) - **Factor Construction Idea**: Combines multiple dimensions of factors using SA weighting and multi-dimensional optimization[69] - **Factor Construction Process**: - Factors from different dimensions (e.g., PV, G, C, HF, DP) are combined using SA weighting and long-only return weighting to create a composite factor[69] - **Factor Evaluation**: Achieves the best overall performance among all tested factors[72] --- Factor Backtesting Results PV Factor - **10-day RankIC**: 11.7%[68] - **Annualized RankICIR**: 5.7[68] - **Long-Short Annualized Return**: 58.9%[68] - **Long-Short Information Ratio**: 5.1[68] - **Maximum Drawdown**: -4.8%[68] - **Long-Only Annualized Return**: 24.1%[68] - **Long-Only Information Ratio**: 3.0[68] G Factor - **10-day RankIC**: 11.0%[68] - **Annualized RankICIR**: 5.8[68] - **Long-Short Annualized Return**: 59.9%[68] - **Long-Short Information Ratio**: 6.2[68] - **Maximum Drawdown**: -2.5%[68] - **Long-Only Annualized Return**: 23.3%[68] - **Long-Only Information Ratio**: 3.3[68] C Factor - **10-day RankIC**: 10.6%[68] - **Annualized RankICIR**: 5.1[68] - **Long-Short Annualized Return**: 56.4%[68] - **Long-Short Information Ratio**: 5.2[68] - **Maximum Drawdown**: -4.4%[68] - **Long-Only Annualized Return**: 19.5%[68] - **Long-Only Information Ratio**: 2.8[68] HF Factor - **10-day RankIC**: 11.6%[68] - **Annualized RankICIR**: 5.9[68] - **Long-Short Annualized Return**: 57.5%[68] - **Long-Short Information Ratio**: 5.8[68] - **Maximum Drawdown**: -5.2%[68] - **Long-Only Annualized Return**: 19.1%[68] - **Long-Only Information Ratio**: 2.6[68] DP Factor - **10-day RankIC**: 11.4%[68] - **Annualized RankICIR**: 6.2[68] - **Long-Short Annualized Return**: 49.2%[68] - **Long-Short Information Ratio**: 4.4[68] - **Maximum Drawdown**: -4.7%[68] - **Long-Only Annualized Return**: 20.3%[68] - **Long-Only Information Ratio**: 2.8[68] ML_C Factor - **10-day RankIC**: 14.2%[72] - **Annualized RankICIR**: 6.3[72] - **Long-Short Annualized Return**: 72.7%[72] - **Long-Short Information Ratio**: 6.1[72] - **Maximum Drawdown**: -4.8%[72] - **Long-Only Annualized Return**: 26.1%[72] - **Long-Only Information Ratio**: 3.1[72]

Core Insights - The report presents a systematic framework for automated factor research based on Large Language Models (LLM) and Prompt Engineering, aiming to explore the potential applications of AI in the entire quantitative investment chain [1] - The framework was first applied to low-frequency price-volume factors, optimizing the classic Alpha158 factor library and transitioning from an "optimization" paradigm to a "generation" paradigm [1] - AI demonstrated strong factor discovery capabilities in both fundamental and high-frequency data domains, successfully generating new factors and enhancing traditional factor libraries [1] - The report also explores AI's application in unstructured text analysis, utilizing the Gemini model to interpret sentiment from extensive research memos, creating unique sentiment indicators that effectively integrate into stock selection strategies [1] Group 1: Low-Frequency Price-Volume Factor Optimization - The framework was initially applied to the optimization of low-frequency price-volume factors, using the Alpha158 factor library as a foundation for optimization experiments [1] - AI identified logical flaws in original factors and proposed effective improvements, with optimization effects being consistent across multiple time windows from 5 to 60 days [1] - New factors generated by AI, with low correlation to sample factors, showed robust out-of-sample performance, with some factors achieving an Information Coefficient Information Ratio (ICIR) above 1.0 [1] Group 2: Fundamental and High-Frequency Factor Discovery - In the fundamental dimension, AI not only generated enhanced versions of classic factors but also innovatively expanded value, quality, and growth factors from novel perspectives [1] - In the high-frequency dimension, AI was empowered to directly generate Python code, uncovering a set of novel and high-performing high-frequency factors, with some strong signal factors achieving annualized returns exceeding 60% [1] - Integrating the AI-generated high-frequency factor library into the AGRU neural network model significantly improved annualized excess returns from 18.24% to 25.28% [1] Group 3: Alternative Data Processing and Sentiment Analysis - The report investigates AI's potential in processing alternative data, analyzing nearly one million words of research memos using the Gemini 2.5 Pro model [1] - A weekly sentiment factor was constructed, revealing unique asymmetric predictive capabilities, where negative sentiment strongly predicted future price declines, achieving annualized excess returns of 8.26% [1] - This sentiment factor exhibited low correlation with traditional price-volume and fundamental factors, serving as an independent and effective supplementary information source [1] Group 4: Comprehensive Strategy Development - A multi-dimensional information fusion strategy was developed, integrating AI-discovered high-frequency factors with low-frequency market data into the AGRU neural network to form a core Alpha [1] - The final strategy, enhanced by AI sentiment factors for risk adjustment, improved annualized excess returns from 11.15% to 11.81% while maintaining turnover rates [1] - The strategy demonstrated a significant increase in the information ratio from 2.18 to 2.31, validating AI's potential to empower quantitative research across multiple stages and achieve a "1+1>2" effect [1]

Quantitative Factors and Construction Factor Name: KeyPeriod_ret_zero - **Construction Idea**: This factor focuses on the return characteristics during horizontal trading periods within key intraday timeframes, leveraging Level 2 tick data to refine minute-frequency factors[7][25][41] - **Construction Process**: - Identify horizontal trading periods based on minimal price fluctuations - Calculate returns during these periods using tick-level data - Aggregate and smooth the data over different time horizons (e.g., 5-day, 20-day)[25][27] - **Evaluation**: Demonstrates strong predictive power for stock selection, with high IC stability and win rates[7][25] Factor Name: KeyPeriod_ret_low5pct - **Construction Idea**: This factor captures return characteristics during significant downward price movements within key intraday timeframes[7][25][64] - **Construction Process**: - Identify periods where returns fall within the bottom 5% of all intraday returns - Calculate and aggregate these returns over different time horizons - Apply smoothing techniques to enhance signal stability[25][27] - **Evaluation**: Exhibits robust performance in identifying underperforming stocks, with high IC values and win rates[7][25] Factor Name: KeyPeriod_price_low5pct - **Construction Idea**: This factor focuses on price levels during periods of low prices (bottom 5%) within key intraday timeframes[7][25][88] - **Construction Process**: - Identify periods where prices fall within the bottom 5% of all intraday prices - Aggregate and smooth the data over different time horizons - Incorporate buy/sell distinctions for further refinement[25][32] - **Evaluation**: Effective in capturing undervalued stocks, with strong IC performance and high win rates[7][25] Factor Name: KeyPeriod_amount_top30pct - **Construction Idea**: This factor targets periods of high transaction amounts (top 30%) within key intraday timeframes[7][25][110] - **Construction Process**: - Identify periods where transaction amounts are in the top 30% of all intraday amounts - Aggregate and smooth the data over different time horizons - Differentiate between buy and sell transactions for enhanced granularity[25][35] - **Evaluation**: Demonstrates strong predictive power for high-liquidity stocks, with high IC values and win rates[7][25] Factor Name: KeyPeriod_amount_low50pct - **Construction Idea**: This factor captures periods of low transaction amounts (bottom 50%) within key intraday timeframes[7][25][133] - **Construction Process**: - Identify periods where transaction amounts are in the bottom 50% of all intraday amounts - Aggregate and smooth the data over different time horizons - Incorporate buy/sell distinctions for further refinement[25][35] - **Evaluation**: Useful for identifying low-liquidity stocks, though performance is less consistent compared to other factors[7][25] Factor Name: KeyPeriod_sync_low50pct - **Construction Idea**: This factor measures volume-price divergence during periods of low synchronization (bottom 50%) within key intraday timeframes[7][25][155] - **Construction Process**: - Identify periods where volume and price movements are least synchronized - Aggregate and smooth the data over different time horizons - Differentiate between buy and sell transactions for enhanced granularity[25][38] - **Evaluation**: Effective in capturing unique market dynamics, with strong IC performance and high win rates[7][25] --- Backtesting Results KeyPeriod_ret_zero - **IC Mean**: -5.36% (20-day horizon)[27] - **Win Rate**: 85.1% (20-day horizon)[27] - **IR**: 1.34 (2020-2025)[55] KeyPeriod_ret_low5pct - **IC Mean**: 5.47% (20-day horizon)[27] - **Win Rate**: 84.1% (20-day horizon)[27] - **IR**: 1.41 (2020-2025)[77] KeyPeriod_price_low5pct - **IC Mean**: 5.59% (20-day horizon)[32] - **Win Rate**: 85.3% (20-day horizon)[32] - **IR**: 2.22 (2020-2025)[97] KeyPeriod_amount_top30pct - **IC Mean**: 11.23% (20-day horizon)[35] - **Win Rate**: 84.8% (20-day horizon)[35] - **IR**: 1.37 (2020-2025)[123] KeyPeriod_amount_low50pct - **IC Mean**: -10.50% (20-day horizon)[35] - **Win Rate**: 75.0% (20-day horizon)[35] - **IR**: 0.77 (2020-2025)[145] KeyPeriod_sync_low50pct - **IC Mean**: 6.00% (20-day horizon)[38] - **Win Rate**: 81.5% (20-day horizon)[38] - **IR**: 1.44 (2020-2025)[172]