HEPA

Summary

HEPA is a self-supervised horizon-conditioned event predictive architecture for multivariate time series. It pretrains a causal encoder and predictor to forecast future representations, freezes the encoder, and fine-tunes only the predictor plus an event head to produce a monotonic probability of an event within each future horizon.

Method Contract

• Input: multivariate observation history up to time $t$.
• Self-supervised target: representation of a future interval at horizon $\Delta t$.
• Stabilizer: SIGReg with a shared target encoder; no EMA or stop-gradient target.
• Downstream adaptation: predictor plus event head; encoder frozen.
• Output: horizon-indexed survival CDF $p(t,\Delta t)$.
• Current transfer claim: architecture and recipe transfer across domains; weights do not.

Official Artifacts

• Preprint: https://arxiv.org/abs/2605.11130v4
• Official project page: https://www.forgis.com/papers/hepa.html
• Official code: https://github.com/Forgis-Labs/HEPA
• FMSD @ ICML 2026 Spotlight listing: https://icml-structured-fm-workshop.github.io/accepted-papers/
• Raw paper Markdown: paper_hepa-2026.md

Role In The Wiki

HEPA is the direct time-series event-prediction bridge between LeNEPA-style latent self-supervision and operational rare-event prediction. It is also the method component of the Forgis stack: FactoryNet supplies multi-embodiment industrial trajectories, and FactoryBench supplies state/intervention/counterfactual/decision evaluation.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in HEPA rather than duplicating verdict rows here.

At the entity level, HEPA is important because it retains and specializes the predictive head rather than treating the pretrained encoder as the only reusable object. The missing step is a single cross-domain or multi-embodiment checkpoint with streaming state, calibrated uncertainty, and candidate-control conditioning.

Caveats

• Each dataset receives its own pretrained encoder.
• The code release changes the C-MAPSS fine-tuning stopping rule relative to the paper.
• Results are strongest for extended precursor dynamics and not uniform across short-window anomalies.
• No action, control-input, intervention, or counterfactual rollout interface is evaluated.

Related Pages

• HEPA source
• FactoryNet
• FactoryBench
• LeNEPA
• Joint Embedding Predictive Architecture
• Latent-Space Predictive Learning
• Foundation Time-Series Model Research Agenda