Dynamic Optimisation of Window Sizes for Enhanced Time-Series Forecasting

Accurately forecasting future trends in complex, high-dimensional time series is challenging, as the predictive accuracy of models hinges on selecting appropriate historical data windows. While existing research in fields such as finance and energy forecasting has advanced from simple linear models to sophisticated machine learning and deep learning architectures, most current approaches rely on static, arbitrarily chosen window sizes. This limitation prevents models from adapting to rapidly evolving conditions and may dilute the impact of recent, highly informative data. To overcome these shortcomings, this work introduces a dynamic window sizing framework that adjusts window lengths in real time, guided by changes in volatility. Using cryptocurrency markets as an illustrative case study -- where high volatility and complex dynamics demand agile methodologies -- the proposed approach identifies optimal window sizes for distinct volatility levels and integrates them into a hybrid deep learning Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural network (LSTM-GRU) model. This enables the model to align its training perspective with evolving market states, capturing both transient and persistent patterns more effectively than static-window counterparts. Experimental evaluations show demonstrate that dynamic window sizing consistently outperforms all tested static window approaches by yielding lower mean squared error (MSE) and mean absolute error (MAE), along with higher directional accuracy. Specifically, the dynamic model attained an MSE of 0.1749, an MAE of 0.2281, and a directional accuracy of 56.77%, outperforming the best static model, which recorded an MSE of 0.1933, an MAE of 0.2883, and a directional accuracy of 49.11%. By offering a robust mechanism for aligning forecasting models with current data characteristics, the proposed adaptive strategy addresses the limitations of static window sizing and sets a new benchmark for handling rapidly changing data. Moreover, this approach establishes a generalisable framework that can enhance decision-making and risk mitigation in diverse domains where adaptability and precision are critical.These results highlight the effectiveness of adopting an adaptive approach in scenarios characterised by frequent changes in data patterns, thereby underscoring the importance of dynamic methodologies over fixed-window configurations.