Identifying Optimal Window Size Configurations for Big Data Time Series Forecasting

Optimal window sizing in time series forecasting emerges as a pivotal factor for enhancing predictive accuracy, particularly in the volatile cryptocurrency market. While traditional models often rely on static window sizes, resulting in compromised forecasting performance, this research explores optimal window configurations across various market volatilities. By employing a hybrid Long Short-Term Memory and Gated Recurrent Unit (LSTM-GRU) model, the study systematically identifies the most effective window sizes for high, medium, and low volatility conditions. Results demonstrate that smaller windows are preferable in highly volatile environments to capture rapid market shifts, whereas larger windows are more suitable for stable conditions to incorporate a broader historical context. By identifying the predetermined optimal window sizes for each volatility segment, this study offers valuable insights for researchers aiming to enhance the adaptability and efficacy of predictive models. These results are especially useful for exploring dynamic window sizing techniques across various domains, particularly in fields where data volatility significantly impacts model performance.