How much historical data do we need? The role of data recency and training period length in LSTM-based rainfall-runoff modeling

Rainfall-runoff models based on Long Short-Term Memory (LSTM) networks typically require extensive datasets for training. While increasing the number of watersheds generally improves model performance, the appropriate temporal scope of training data—especially under potential non-stationarity from climate change—remains unclear, and the importance of data recency remains underexplored. This study investigates whether decades of historical data are necessary when training with large numbers of watersheds, and whether shorter, more recent training periods can match the performance of extended records. Using hydrometeorological data from 1374 North American watersheds (spanning 1950–2023), we systematically evaluate the effect of data recency on LSTM performance through three complementary experimental designs: (1) backward-expanding training periods (fixed recent data + progressively older blocks), (2) forward-expanding periods (fixed old data + progressively newer blocks), and (3) sliding-window periods (fixed-length windows moving forward). Results reveal that newer data universally improves predictions, while older data contributes marginally or negatively to model performance. Notably, the benefit of increasing the number of watersheds is conditional on data recency, as Prediction in Ungauged Basins (PUB) performance improves most significantly with recent data. These findings provide the empirical evidence for the importance of data recency, demonstrating that data recency—not just data volume—is critical for LSTM-based streamflow prediction, underscoring the need for recency-aware training strategies.