开源量化评论（109）：深度学习赋能技术分析

Quantitative Models and Construction Methods 1. Model Name: LSTMtech - Model Construction Idea: Directly use LSTM to extract factors from stock price data and technical indicators without incorporating financial data[13][15] - Model Construction Process: 1. Input layer includes stock price data (open, high, low, close, volume) and technical indicators generated using the talib library[15] 2. Training process uses a rolling window: 6 years for training, 2 years for validation, updated annually[15] 3. Factor performance metrics: RankIC of 7.42%, RankICIR of 4.25, annualized return of 24.02% for 10 long-short groups[15] 2. Model Name: LSTMdeap_tech - Model Construction Idea: Combine effective factors extracted by genetic algorithms with original technical indicators, then use LSTM for further factor mining[3][24][26] - Model Construction Process: 1. Genetic algorithm is applied to original technical indicators to extract effective factors using a sample period from 2010 to 2016[24] 2. Effective factors are combined with original technical indicators and input into the LSTM framework[26] 3. Factor performance metrics: RankIC of 9.27%, RankICIR of 4.54, annualized return of 32.44% for 10 long-short groups[26] 3. Model Name: LSTMgraph - Model Construction Idea: Use LSTM to extract factors based on manually defined graphical states of stock price patterns and technical indicators[32][41] - Model Construction Process: 1. Graphical states are manually defined based on K-line patterns and technical indicator positions (e.g., relative positions of moving averages)[33] 2. States are input into the LSTM framework for training, including synthesized K-line patterns from 1 to 20 days[41] 3. Factor performance metrics: RankIC of 9.01%, RankICIR of 4.70, annualized return of 32.25% for 10 long-short groups[41][44] 4. Model Name: LSTMdeap_tech_graph - Model Construction Idea: Combine LSTMdeap_tech and LSTMgraph factors equally to enhance performance[5][47][49] - Model Construction Process: 1. Combine LSTMdeap_tech and LSTMgraph factors with equal weights[47] 2. Factor performance metrics: RankIC of 10.89%, RankICIR of 4.99, annualized return of 37.28% for 10 long-short groups[49] 5. Model Name: LSTMpro_combined - Model Construction Idea: Combine LSTMdeap_tech_graph with a factor derived from broader data sources (e.g., minute-level aggregated daily indicators, financial data)[6][54] - Model Construction Process: 1. Combine LSTMdeap_tech_graph with the broader factor using equal weights[54] 2. Factor performance metrics: RankIC of 11.93%, annualized return of 39.85% for 10 long-short groups[54] --- Model Backtesting Results 1. LSTMtech - RankIC: 7.42% - RankICIR: 4.25 - Annualized return (10 long-short groups): 24.02%[15] 2. LSTMdeap_tech - RankIC: 9.27% - RankICIR: 4.54 - Annualized return (10 long-short groups): 32.44%[26] 3. LSTMgraph - RankIC: 9.01% - RankICIR: 4.70 - Annualized return (10 long-short groups): 32.25%[41][44] 4. LSTMdeap_tech_graph - RankIC: 10.89% - RankICIR: 4.99 - Annualized return (10 long-short groups): 37.28%[49] 5. LSTMpro_combined - RankIC: 11.93% - Annualized return (10 long-short groups): 39.85%[54] --- Quantitative Factors and Construction Methods 1. Factor Name: Tech_K_similarity - Factor Construction Idea: Manually define graphical states and calculate historical similarity to derive future returns[33][35] - Factor Construction Process: 1. Define states based on K-line patterns and technical indicators (e.g., MACD, volume)[33] 2. Identify historical periods with similar states and calculate average excess returns over the next 20 days[35] 3. Factor performance metrics: RankIC of 5.10%, RankICIR of 3.09, annualized return of 19.25% for 10 long-short groups[39] 2. Factor Name: Tech_similarity_combined - Factor Construction Idea: Combine Tech_K_similarity with a similar factor derived from technical indicators[40] - Factor Construction Process: 1. Combine Tech_K_similarity and the technical indicator-based factor equally[40] 2. Factor performance metrics: RankIC of 5.89%, RankICIR of 3.25, annualized return of 25.97% for 10 long-short groups[40][43] --- Factor Backtesting Results 1. Tech_K_similarity - RankIC: 5.10% - RankICIR: 3.09 - Annualized return (10 long-short groups): 19.25%[39] 2. Tech_similarity_combined - RankIC: 5.89% - RankICIR: 3.25 - Annualized return (10 long-short groups): 25.97%[40][43]