Phongsakon/heimdall-forecasters

Heimdall forecasters — DK1 imbalance prices (private development checkpoint)

Pre-release research checkpoints for the Heimdall thesis project (BSc Software Engineering, SDU Sønderborg; industrial partner Danfoss A/S). These weights predict 15-minute imbalance prices for the Nordic mFRR EAM in bidding zone DK1 under the post-2025-03-04 single-price regime.

The checkpoints are versioned by sprint day; this push corresponds to 2026-05-09 and contains the F0 / F3 / F7 / F8 forecaster zoo entries trained per the frozen seed set [13, 42, 137, 1729, 31415].

Motivation

The thesis's central object is a focal verifier-guarded market-maker whose every bid passes a two-stage verifier: physical feasibility (deterministic) and conformal worst-case profit (Theorem 1a / 1b in the project's research proposal). The forecasters here are the upstream uncertainty source that the verifier consumes via either split-CP or online ACI.

Intended use

• Reproduce the forecaster-zoo leaderboard on DK1 imbalance prices.
• Feed quantile predictions into the project's apps/verifier for end-to-end bid acceptance experiments.

Out-of-scope use

• DO NOT use these weights for live trading. They were trained on publicly available ENTSO-E + Energinet data and have not been audited for any production trading system; the project's own verifier is what makes the bids safe for simulation, and that pipeline is not bundled here.

Architecture summary

ID	Family	Params	Heads
F0	AutoReg(24) (statsmodels)	≈25 (linear)	residual quantiles
F3	DeepAR-Lite (1-layer LSTM, 64 hidden)	≈50 k	Student-t
F7	patchTST (univariate)	≈2 M	quantile q∈(0.1, 0.5, 0.9)
F8	patchTST (multivariate: price, load, DA)	≈2 M	quantile q∈(0.1, 0.5, 0.9)

Training data

• Source 1 — ENTSO-E Transparency Platform v2 A85 imbalance prices, A44 day-ahead prices, A65 actual load.
• Source 2 — Energinet Open Data RegulatingBalancePowerdata for the hourly imbalance and mFRR up/down volumes.

Pre/post-EAM split:

• Train: 2020-01-01 → 2025-02-28 (≈175 k rows at 15-min resolution).
• Val: 2025-03-04 → 2025-04-30 (≈5.5 k rows; the EAM go-live is on the 2025-03-04 boundary, the sacred regime-shift line).
• Test: 2025-05-01 → 2026-04-30 (held out; evaluated only once per final config).

Pre-2025-03-04 the 15-min ENTSO-E A85 signal does not exist, so the target column is back-filled from the hourly Energinet imbalance signal for those rows. KE2 (sprint-day-9 kill experiment) verifies that this does not push trainer-relevant null fractions above 5 %.

Per-seed metrics (val window)

The headline numbers and per-quantile pinball losses are stored in the included metrics.json files at each <model>/seed-<n>/metrics.json path and aggregated at <model>/metrics.json. The full leaderboard, including mean ± std, lives in the project's notes/forecaster_leaderboard.md.

License

Apache 2.0 — LICENSE text inherited from the project root.

Citation

@misc{heimdall2026,
  title={Heimdall: A Verifier-Guarded LLM Society for Post-EAM Nordic Balancing Markets},
  author={Konrad, Phongsakon and Adam, Tim Lukas},
  year={2026},
  note={BSc thesis, SDU S{\o}nderborg; industrial partner Danfoss A/S. Preprint forthcoming.}
}

Notes

• The agent runtime (society of LLM-persona BRPs and the focal-orchestrator LLM tool stack) is not bundled with this checkpoint; that half of the project is owned by the co-author and is on a separate sprint track.
• This repository will be re-pushed as public (license: apache-2.0) at preprint time per the project's reproducibility commitment.