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

TL;DR

Long-horizon event forecasting is challenging due to error accumulation and limited output diversity in autoregressive approaches. The authors present DEF, a horizon-matching, multi-head framework that predicts K future events in parallel and aligns them with ground truth via a Hungarian-based loss. DEF achieves state-of-the-art performance on long-horizon metrics across diverse datasets and also excels in next-event prediction, while maintaining efficient, parallel inference. The approach combines a probabilistic prediction head, a calibration mechanism, and a parameter-efficient conditional head design, enabling robust deployment across domains such as retail, healthcare, and social networks.

Abstract

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.

Figures (7)

• Figure 1: (a) A typical next-event or pairwise loss compares events at corresponding positions, often resulting in incorrect matching. (b) The proposed matching loss calculates the loss function between the closest events, leading to a more robust and balanced error measure. (c) The proposed DEF method enhances the diversity of predictions. We demonstrate 3 example sequences from the Amazon dataset generated by the autoregressive IFTPP method, Diffusion, and the proposed approach. Each label type is depicted using a distinct shape and color combination. The precise timestamps are omitted for simplicity.
• Figure 2: The proposed DEF simultaneously predicts $\mathrm{K}$ future events. Each prediction head outputs occurrence probability $\hat{o}$, time $\hat{t}$, and labels distribution $\hat{p}(l)$. During training, a novel matching loss aligns predictions with the ground truth sequence and evaluates its likelihood.
• Figure 3: Our conditional prediction head.
• Figure 4: Extended horizons prediction on the Transactions dataset.
• Figure 5: Next event prediction errors: MAE for time and error rate for type.