Self-Optimizing Algorithms For Continuous Learning In Financial Time-Series Forecasting

Abstract

Self-optimizing algorithms are increasingly essential for continuous learning systems, especially in financial time-series forecasting, where data distributions are highly dynamic and non-stationary. Traditional machine learning and deep learning models rely on static training procedures, limiting their adaptability to rapid market fluctuations and leading to model drift and degraded predictive performance. The research proposes a self-optimizing continuous learning based on a Modified Osprey Attention-Gated Recurrent Unit (ModO-Att-GRU) model. The objective is to enable autonomous adaptation to evolving financial patterns through the integration of online learning, reinforcement-based optimization, and meta-learning strategies. Financial data is collected from a publicly available Kaggle stock market dataset. Min-Max normalization is employed as a preprocessing technique to stabilize training and improve convergence. For feature extraction, the Relative Strength Index (RSI) is utilized to capture momentum and trend-related information embedded in the time-series data.The proposed ModO-Att-GRU model incorporates an osprey-inspired optimization mechanism for dynamic weight adjustment, along with an attention gating mechanism to selectively focus on relevant temporal features. The proposed ModO-Att-GRU model was implemented and evaluated using Python. Experimental evaluation is conducted on financial datasets using metrics such as prediction accuracy (93.85%), and test Receiver Operating Characteristic Area under the Curve (ROC-AUC) (95.08%). Results demonstrate that the proposed method improves adaptability, reduces model drift, and enhances long-term forecasting performance compared to conventional methods.