Detecting the Future: All-at-Once Event Sequence Forecasting with Horizon Matching

Long-horizon events forecasting is a crucial task across various domains, including retail, finance, healthcare, and social networks. Traditional models for event sequences often extend to forecasting on a horizon using an autoregressive (recursive) multi-step strategy, which has limited effectiveness due to typical convergence to constant or repetitive outputs. To address this limitation, we introduce DEF, a novel approach for simultaneous forecasting of multiple future events on a horizon with high accuracy and diversity. Our method optimally aligns predictions with ground truth events during training by using a novel matching-based loss function. We establish a new state-of-the-art in long-horizon event prediction, achieving up to a 50% relative improvement over existing temporal point processes and event prediction models. Furthermore, we achieve state-of-the-art performance in next-event prediction tasks while demonstrating high computational efficiency during inference. Code - https://github.com/ivan-chai/hotpp-benchmark * Corresponding author. ing purchases for the next month or making long-term medical prognoses (Xue et al. 2022) . This task presents unique challenges that differ from traditional next-event prediction. The conventional approach typically relies on autoregressive models, which predict the next event step by step (Xue et al. 2024; Xiao et al. 2018) . While these models are effective for immediate next-event forecasting, their performance tends to deteriorate as the prediction horizon extends (Karpukhin, Shipilov, and Savchenko 2024). The same is true for horizon prediction models, including GAN (Xiao et al. 2018) and diffusion (Zhou et al. 2025), which predict multiple future events at once but use pairwise losses between events on corresponding positions. In this study, we identify significant limitations of pairwise losses in the context of long-horizon prediction. To address these challenges, we propose DEF (Detection-based Event Forecasting),which detects multiple future events in parallel and employs a novel horizon matching loss, which dynamically aligns predictions with the closest ground-truth events, as illustrated in Fig. 1 . This loss function enables the model to capture the full distribution of events within the horizon while remaining robust to outlier events. We demonstrate that our approach establishes a new state-of-the-art in long-horizon prediction, surpassing both autoregressive and horizon prediction approaches in terms of accuracy and prediction diversity. Additionally, our method exhibits high computational efficiency during inference, ranking among the fastest methods.