Enhancing Ecological Forecasting with LSTM Models: The Impact of Partition-Based Data Shuffling on Predictive Accuracy

This study explores the impact of a novel partition-based shuffling technique on the predictive accuracy of Long Short-Term Memory (LSTM) models in ecological forecasting. By utilizing datasets on mosquito population dynamics and environmental variables, we demonstrate that partition-based shuffling-which randomizes inter-year sequences while preserving intra-year temporal patterns-significantly enhances model performance. Experimental results show that models trained with this method achieve a notable reduction in root mean square error (RMSE) compared to those trained on non-shuffled data, underscoring its ability to balance temporal integrity with generalization. The study offers a robust methodological framework that improves the applicability of LSTM models in ecological and epidemiological research by addressing biases caused by sequential data dependencies. The findings advocate for the adoption of partition-based shuffling as a crucial preprocessing step to enhance the reliability and robustness of time-series forecasting models in similar domains.