Research on Multi Source Data Driven Modeling of Individual Investment Decision Behavior and Machine Learning Prediction Mechanism

Abstract

This study develops a multi-source data-driven framework for modeling individual investment decision behavior and predicting investor actions through machine learning. The research focuses on how trading history, portfolio structure, market exposure, information attention, and user profile variables jointly shape next-day buy, sell, and hold decisions. The dataset is constructed from brokerage transaction records, investor account profiles, daily stock-market quotation data, and financial-news interaction logs, with all records aligned at the user-day level. A fixed feature system is built through rolling-window trading indicators, portfolio-risk indicators, market-state variables, attention features, and demographic-account variables. The prediction mechanism is implemented through an XGBoost multi-class classifier, supported by standardized feature processing, stratified sample splitting, probability output generation, and SHAP-based interpretation. The results show that the model achieves stable prediction performance on the test set, with macro-F1 of 0.759±0.008 and multi-class AUC of 0.864±0.004. Trading frequency, unrealized return, cash proportion, market volatility, and news-click intensity are the strongest behavioral predictors.