New Regularization Prevents Suboptimal Collapse in Time Series Models

Time series forecasting (TSF) is critical across many applications, and Time Series Foundation Models (TSFMs) have shown strong generalization capabilities after pretraining on large datasets. Reinforcement learning (RL) post-training is increasingly used to further enhance TSFM performance on specific downstream tasks. However, a significant challenge, termed "suboptimal collapse," has been observed where RL post-training can inadvertently shift TSFM output distributions away from the true ground truth in certain forecast regions, thereby limiting overall performance. The core issue identified is the difficulty in initially sampling high-quality trajectories close to the ground truth. To counteract this, the researchers propose Ground-Truth Neighborhood Regularization (GTN-R). GTN-R uses the actual ground truth as a reference point, actively guiding the model's probability mass towards this "ground-truth neighborhood." This strategy increases the likelihood of sampling high-quality trajectories, effectively mitigating suboptimal collapse and leading to improved forecasting accuracy. GTN-R is designed to be flexible and can be integrated into various existing RL methods for TSFMs, with extensive experiments demonstrating its effectiveness.