Title: HEPA: A Self-Supervised Horizon-Conditioned Event Predictive Architecture for Time Series

Abstract:

Accurate forecasting of critical events within multivariate time series—ranging from cardiac arrhythmias to turbine malfunctions—is essential, yet hindered by the scarcity of labeled data due to the infrequency and high annotation costs of such occurrences. To address this, we present HEPA (Horizon-conditioned Event Predictive Architecture), grounded in two fundamental strategies. Initially, we employ a causal Transformer encoder that undergoes self-supervised pretraining using a Joint-Embedding Predictive Architecture (JEPA). In this framework, a horizon-conditioned predictor is trained to anticipate future representations instead of raw values, compelling the encoder to extract predictable temporal patterns exclusively from unlabeled datasets. Subsequently, we keep the encoder weights fixed and fine-tune only the predictor to align with the specific target event, thereby generating a monotonic survival cumulative distribution function (CDF) across various horizons. Demonstrating robust generalizability with consistent architecture and optimizer hyperparameters, HEPA addresses challenges in water contamination monitoring, cyberattack identification, and volatility regime detection, alongside eight other event categories spanning 11 distinct domains. The model outperforms state-of-the-art time-series architectures, including PatchTST, iTransformer, MAE, and Chronos-2, on 10 out of 14 benchmark tasks. Notably, HEPA achieves this superior performance while requiring an order of magnitude fewer tuned parameters and, particularly on lifecycle datasets, an order of magnitude less labeled data.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC