In this study, we propose a real-time time-series prediction system that integrates dynamic window sizing with advanced neural network architectures. This approach addresses the limitations of static window sizing, which often fails to adapt to rapidly changing market conditions. By tailoring the input window size to volatility levels, dynamic window sizing ensures that the forecasting model is provided with contextually relevant information, enhancing its adaptability and accuracy. The system integrates dynamic window sizes and deploying advanced neural network architectures, namely LSTM GRU hybrid model and a modified Transformer encoder, for real-time forecasting. In the case study of cryptocurrency price prediction, the data is collected from high-frequency datasets from Coinspot and Reddit, containing price, volume, technical indicators, and associated sentiment. Window sizes are optimised through a rigorous process to enhance predictive accuracy over static window sizing approaches. The results show that dynamic window sizing enhances predictive accuracy compared to static window sizing approaches. Overall, this research contributes to the advancement of dynamic forecasting methodologies and underscores the importance of incorporating dynamic window sizes into real-time data processing pipelines for enhanced predictive performance.