使用多代理強化學習進行投資組合管理.pdf

1、2024 Databricks Inc.All rights reservedSTOCK PORTFOLIO STOCK PORTFOLIO MANAGEMENT USING MANAGEMENT USING AIAIAyesha NasimAyesha NasimAI&ML EngineerAI&ML Engineer12024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reserved Experienced in Developing Developing&OperatingOperatingTech

2、nology Agnostic AIAI-Based Based Enterprise Scale Products Masters in Data Science and Working Allegis Group as MLOps EngineerABOUT THE SPEAKERABOUT THE SPEAKER22024 Databricks Inc.All rights reserved3POWER OF DATA POWER OF DATA FUELED AIFUELED AI-BASED BASED INVESTMENT INVESTMENT DECISIONSDECISIONS

3、2024 Databricks Inc.All rights reservedDeveloped an Automated Stock Portfolio Management System and overall net gain in terms of a dollar amount,using:Multi-Agent Reinforcement LearningTop 10 Stable&Volatile Stocks each from S&PCombination of Leading,Lagging,and Coincident Technical IndicatorsSentim

4、ents from News Publications4VISIONVISIONPortfolio Management Using MultiPortfolio Management Using Multi-Agent Reinforcement LearningAgent Reinforcement Learning2024 Databricks Inc.All rights reservedDevelopmentDevelopment EnvironmentEnvironmentLibrariesLibraries5TOOLS&TECHNOLOGIESTOOLS&TECHNOLOGIES

5、Google CollabVS CodePandasTexBlobOpenAI GymStable Baselines3Data SourceData SourceYahoo Finance(Stock Prices)Kaggle(News)FRED2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedARCHITECTUREARCHITECTURE2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights res

6、ervedIMPLEMENTATIONIMPLEMENTATION7KickKick-offoffTime Period SelectionTime Period Selection-Idea Inception-Reviewed Pre-Existing WorkExplored GDP&Stocks Historical Data How effective can a Multi-Agent Reinforcement Learning Algorithm be for Automated Stock Portfolio Management?2024 Databricks Inc.Al

7、l rights reserved2024 Databricks Inc.All rights reservedIMPLEMENTATIONIMPLEMENTATION8Portfolio SelectionPortfolio SelectionData Collection&PreData Collection&Pre-ProcessingProcessingSelect 5 most and 5 least volatile stocks-Found News Publication Dataset&Conduct Sentiment Classification-Figured Out

8、Technical Indicators-Refined&Pre-Processed Stocks Dataset-Pre-Processed the Technical Indicators Data Set2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedIMPLEMENTATIONIMPLEMENTATION9Model ConstructionModel ConstructionFiguredoutthe ReinforcementLearningStrategy,the exis

9、ting Packages&Libraries2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedIMPLEMENTATIONIMPLEMENTATION10Environment SetupEnvironment SetupSetup the Environment,Observation&Action Space,and associated methodsAgents SelectionAgents SelectionSelect Deep RL Agents&their respec

10、tive hyperparameters2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedIMPLEMENTATIONIMPLEMENTATION11Ensemble Strategy Ensemble Strategy DevelopmentDevelopmentDevelopment of Multi-agent strategyA Higher Sharpe Ratio Indicates Better Risk-Adjusted Performance When Comparing

11、 Similar Portfolios2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedRESULTSRESULTS125 Most&5 Least Volatile Stocks5 Most&5 Least Volatile StocksBacktest Stats&Backtest Stats&Account SummaryAccount SummaryCumulative return improved by approximately 35.34%35.34%compared to

12、 baseline2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedRESULTSRESULTS13Backtest Stats&Backtest Stats&Account SummaryAccount Summary5 Most Volatile Stocks5 Most Volatile StocksCumulative return decreased by approximately 8.56%8.56%compared to baseline2024 Databricks In

13、c.All rights reserved2024 Databricks Inc.All rights reservedRESULTSRESULTS14Backtest Stats&Backtest Stats&Account SummaryAccount Summary5 Least Volatile Stocks5 Least Volatile StocksCumulative return improved by approximately 53%53%compared to baseline2024 Databricks Inc.All rights reserved2024 Data

14、bricks Inc.All rights reserved15GENERATIVE AI GENERATIVE AI BASED STOCK BASED STOCK RECOMMENDATION RECOMMENDATION SYSTEMSYSTEM2024 Databricks Inc.All rights reservedUI/UX was developed on Streamlit16VISIONVISIONGenerative AI Based Stock Recommendation SystemGenerative AI Based Stock Recommendation S

15、ystemTo utilize Generative AI for Stock Investment DecisionsDevelopedaRecommendationSystemConstitutingMultipleModulesChainedTogether by LangchainRAG&ReAct Based Recommendations toeliminate hallucinations and provide qualityoutputs2024 Databricks Inc.All rights reserved17WORKFLOWWORKFLOWUser PromptGe

16、t the Stock TickerFetch Historic Data on StockGet Financial StatementsScrape the Web for Stock NewsLLM ReAct based Verbal AnalysisOutput Recommendation:BuyBuy,SellSell,or Hold Hold with Justification2024 Databricks Inc.All rights reserved18DEMODEMO2024 Databricks Inc.All rights reserved2024 Databric

17、ks Inc.All rights reserved There are multiple AI-Powered Techniques forStock Portfolio Management,Both Predictive&Generative.Combination of PredictivePredictive&GenerativeGenerative AIAImodels can be used to leverage strengths ofbothintheAutonomousStockPortfolioManagement SystemCONCLUSIONCONCLUSION19Databricks Databricks offers a robust suite of tools and capabilities e.g.,ML Runtime&ML Runtime&MLflowMLflow to empower the development and scaling of such AI products.AI products.2024 Databricks Inc.All rights reserved20Q/AQ/A Me On: