An Adaptive Sharpe Ratio-Based Temporal Fusion Transformer for Financial Forecasting

Abstract

Financial time series forecasting plays a critical role in investment decision-making, risk management, and portfolio optimization. Traditional forecasting models, including statistical approaches and deep learning-based architectures, often focus on maximizing predictive accuracy but fail to incorporate dynamic risk-adjusted performance metrics. The Sharpe ratio, a widely used measure of risk-adjusted return, is typically applied post-forecasting, limiting its potential to guide predictive models during training. To address this limitation, this study proposes an adaptive Sharpe ratio-based temporal fusion transformer (AS-TFT) that integrates risk-aware forecasting mechanisms to optimize financial predictions while considering return volatility.

Experimental evaluations on stock indices, cryptocurrency prices, and commodity markets demonstrate that the AS-TFT outperforms conventional forecasting methods in terms of both predictive accuracy and portfolio returns. The results highlight the importance of integrating risk-adjusted financial performance metrics within deep learning-based forecasting architectures, offering a practical and scalable solution for financial decision-making.