AI-Driven Predictive Analytics in Finance

Abstract

Machine learning models are increasingly deployed for financial prediction tasks -- credit scoring, fraud detection, stock return forecasting, and portfolio optimisation -- yet systematic comparisons across prediction horizons, market regimes, and regulatory constraints remain scarce. This study presents a controlled evaluation of seven ML methods -- logistic regression (baseline), random forest, gradient-boosted trees (XGBoost), feedforward neural networks, LSTM networks, temporal fusion transformers (TFT), and graph neural networks for inter-asset dependencies -- across four financial prediction tasks: credit default prediction (Lending Club, 2.26M loans), transaction fraud detection (IEEE-CIS, 590K transactions), equity return forecasting (S&P; 500 constituents, daily 2010-2021), and portfolio risk estimation (covariance forecasting for 50-asset portfolios). All models were evaluated under realistic temporal train-validation-test splits to prevent look-ahead bias, with performance measured during both stable and crisis market regimes. XGBoost achieved the highest AUC on credit scoring (0.784 +- 0.006) and fraud detection (0.942 +- 0.004). The temporal fusion transformer achieved the lowest forecasting error on equity returns (RMSE = 0.0184 +- 0.0012) and best Sharpe ratio when integrated into a mean-variance portfolio (0.84 +- 0.14 vs. 0.62 +- 0.18 for equal-weight benchmark). However, all models exhibited significant performance degradation during the March 2020 COVID-19 market crash: mean AUC on credit scoring dropped 6.8% and return forecasting RMSE increased 124%. Model interpretability analysis revealed that tree-based models satisfied regulatory explainability requirements under the EU AI Act framework more readily than neural approaches. A practical deployment framework balancing predictive performance, regime robustness, and regulatory compliance is proposed.