Fazeldehkordi et al. (2025) The impact of window size on the performance and accuracy of time series forecasting models for meteorological drought prediction

Identification

• Journal: Stochastic Environmental Research and Risk Assessment
• Year: 2025
• Date: 2025-10-17
• Authors: Leila Fazeldehkordi, Jie‐Lun Chiang
• DOI: 10.1007/s00477-025-03098-7

Research Groups

• Department of Civil Engineering, National Pingtung University of Science and Technology, Pingtung, Taiwan
• Department of Soil and Water Conservation, National Pingtung University of Science and Technology, Pingtung, Taiwan

Short Summary

This study investigates the impact of input window size on the predictive performance of SARIMAX, MLP, Seq2Seq-LSTM, and BiLSTM models for meteorological drought forecasting. It finds that optimal window sizes significantly improve forecasting accuracy, with a 12-month window enabling all models to accurately predict conditions for the first six months of 2024.

Objective

• To explore and evaluate the influence of different input lengths (window sizes) on the performance of SARIMAX, MLP, Seq2Seq-LSTM, and BiLSTM forecasting models for meteorological drought prediction.
• To incorporate observed data from 2024 for post-forecast evaluation to assess model performance under recent, actual conditions.
• To determine how input length affects model performance and prediction accuracy, identify optimal window sizes, and assess the reliable forecast horizon.

Study Configuration

• Spatial Scale: Tainan City and its 34 districts, Tainan Basin, Taiwan (2232 square kilometers).
• Temporal Scale:
  • Data Period: Monthly rainfall data from 1993 to 2023.
  • Drought Index: Standardized Precipitation Index (SPI) with a 3-month timescale (SPI-3).
  • Forecast Horizon: 1 to 12 months (T+1 to T+11) for model evaluation, and 12 months for future forecasting (2024).
  • Input Window Sizes: Short-term (3, 6 months), medium-term (12, 15 months), and long-term (18, 24, 30, 36 months).

Methodology and Data

• Models used:
  • Seasonal Autoregressive Integrated Moving Average with exogenous variables (SARIMAX)
  • Multi-Layer Perceptron (MLP)
  • Sequence-to-Sequence Long Short-Term Memory (Seq2Seq-LSTM)
  • Bidirectional Long Short-Term Memory (BiLSTM)
• Data sources:
  • Monthly rainfall data from 22 weather stations (1993-2023) obtained from the Central Weather Bureau (CWB) website.
  • JRC Yearly Water Classification History, v1.4, for water surface maps.
  • Observed SPI-3 data for 2024 for post-forecast evaluation.

Main Results

• Varying input window sizes significantly improved forecasting accuracy, with each model performing optimally at specific window sizes.
• All models, when trained with a 12-month window, accurately forecasted SPI-3 conditions for the first six months of 2024.
• The MLP model performed best with 3- and 6-month window sizes for short-term forecasts, and reliably with a 12-month window for short-to-medium term.
• The Seq2Seq model excelled with 12-, 15-, and 18-month window sizes, demonstrating superior performance for forecasting horizons of 3 to 11 months.
• The BiLSTM model provided stable performance with 12- and 15-month window sizes across all horizons, effectively capturing long-term dependencies.
• Window sizes of 30 and 36 months increased the risk of overfitting in the models, indicating a need for retuning.
• Models trained without specific window sizes (first scenario) failed to accurately capture future drought events in 2024, despite acceptable evaluation metrics (e.g., SARIMAX R² test = 0.83, MLP R² test = 0.86).
• Post-forecast evaluation highlighted the critical importance of continuous model maintenance through regular updates and retraining to adapt to new shifts in climate patterns (e.g., unusual typhoon season in 2024).

Contributions

• Addressed a significant gap in drought forecasting literature by systematically evaluating the impact of input window size on the performance and accuracy of various time series models.
• Provided specific optimal window sizes for SARIMAX, MLP, Seq2Seq-LSTM, and BiLSTM models for meteorological drought prediction in the Tainan Basin.
• Enhanced the practical value of drought forecasting research by incorporating actual observed data from 2024 for a rigorous post-forecast evaluation, demonstrating the limitations of relying solely on traditional evaluation metrics (e.g., R²) for future predictive accuracy.
• Emphasized the necessity of continuous model updates and retraining for reliable long-term drought forecasting in dynamic climatic conditions.

Funding

• National Science and Technology Council (NSTC) of Taiwan (ROC) under Grant number [NSTC 114-2321-B-020-001].

Citation

@article{Fazeldehkordi2025impact,
  author = {Fazeldehkordi, Leila and Chiang, Jie‐Lun},
  title = {The impact of window size on the performance and accuracy of time series forecasting models for meteorological drought prediction},
  journal = {Stochastic Environmental Research and Risk Assessment},
  year = {2025},
  doi = {10.1007/s00477-025-03098-7},
  url = {https://doi.org/10.1007/s00477-025-03098-7}
}