LTR-Net: A deep learning-based approach for financial data prediction and risk evaluation in enterprises

Area of Science:

• Financial analytics
• Machine learning
• Time-series analysis

Background:

• Financial data prediction and risk assessment are complex multi-task problems.
• Traditional models struggle with temporal dependencies, global information, and nonlinear relationships in financial data.
• Limited prediction accuracy hinders effective financial decision-making.

Purpose of the Study:

• To propose LTR-Net, a novel deep learning model for enhanced financial data prediction and risk assessment.
• To address the limitations of traditional models in capturing complex financial data dynamics.
• To improve the accuracy, stability, and robustness of financial forecasting.

Main Methods:

• Developed LTR-Net, a multi-module deep learning architecture integrating LSTM, Transformer, and ResNet.
• Incorporated temporal dependency modeling, global information capture, and deep feature extraction modules.
• Evaluated LTR-Net on Kaggle Financial Distress Prediction and Yahoo Finance Stock Market datasets.

Main Results:

• LTR-Net significantly outperformed LSTM, GRU, Transformer, and DeepAR on financial datasets.
• Achieved higher accuracy, stability, and robustness across metrics like MSE, RMSE, MAE, and AUC.
• Ablation studies confirmed the crucial contributions of LSTM, Transformer, and ResNet modules.

Conclusions:

• LTR-Net offers superior performance for financial data prediction and risk assessment.
• The model demonstrates strong generalization capabilities applicable to other data analysis domains.
• LTR-Net represents a significant advancement in handling complex financial time-series data.