newest-only mirror @ 2026-05-17T12:51:11Z (250 adds, 0 deletes)

b1_random_walk/MODEL_CARD.md

@@ -0,0 +1,34 @@
+# b1_random_walk — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+_(no config.json schema; see trainer source)_
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **532.34** |
+| Raw [q10, q90] coverage | 0.91% |
+| ACI marginal coverage (target 90%) | 89.81% |
+| ACI mean width (DKK) | 5051.36 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/b1_random_walk/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

b2_ewma/MODEL_CARD.md

@@ -0,0 +1,34 @@
+# b2_ewma — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+_(no config.json schema; see trainer source)_
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **344.51** |
+| Raw [q10, q90] coverage | 86.34% |
+| ACI marginal coverage (target 90%) | 89.91% |
+| ACI mean width (DKK) | 3359.23 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/b2_ewma/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

b3_seasonal_naive/MODEL_CARD.md

@@ -0,0 +1,34 @@
+# b3_seasonal_naive — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+_(no config.json schema; see trainer source)_
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **435.72** |
+| Raw [q10, q90] coverage | 59.79% |
+| ACI marginal coverage (target 90%) | 89.79% |
+| ACI mean width (DKK) | 5157.53 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/b3_seasonal_naive/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

b4_lightgbm_quantile/MODEL_CARD.md

@@ -0,0 +1,34 @@
+# b4_lightgbm_quantile — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+_(no config.json schema; see trainer source)_
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **364.40** |
+| Raw [q10, q90] coverage | 34.95% |
+| ACI marginal coverage (target 90%) | 89.79% |
+| ACI mean width (DKK) | 4577.19 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/b4_lightgbm_quantile/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

b7_nbeats_lite/MODEL_CARD.md

@@ -0,0 +1,34 @@
+# b7_nbeats_lite — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+_(no config.json schema; see trainer source)_
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **274.19** |
+| Raw [q10, q90] coverage | 60.80% |
+| ACI marginal coverage (target 90%) | 89.70% |
+| ACI mean width (DKK) | 2442.07 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/b7_nbeats_lite/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

f0/MODEL_CARD.md

@@ -0,0 +1,35 @@
+# f0 — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+- `seq_len`: `96`
+- `horizon`: `16`
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **478.61 ± 0.00** |
+| Raw [q10, q90] coverage | 6.22 ± 0.00% |
+| ACI marginal coverage (target 90%) | 89.75 ± 0.00% |
+| ACI mean width (DKK) | 5476.87 ± 0.00 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/f0/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

f10/MODEL_CARD.md

@@ -1,26 +1,35 @@
-# F10 — Chronos-Bolt (amazon/chronos-bolt-tiny) zero-shot
+# f10 — model card
 
-**Status:** appendix-only entry. Single seed (42), n_windows = 1000.
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
 
-**Why not 5-seed full val?** Chronos-Bolt is deterministic at inference (no
-seed-randomness in the forward pass), so a 5× replication would yield five
-identical numbers. A full-val 5-seed run is blocked at 2026-05-16 by a
-dependency conflict:
+## Default configuration (out-of-box hyperparameters)
 
-- `chronos-forecasting` / `transformers` requires `huggingface-hub<1.0`.
-- Heimdall's `pyproject.toml` pins `huggingface-hub>=1.14.0` (needed by
-  `transformers>=5.0` for tokeniser changes used elsewhere in the project).
-- The two pins are incompatible — pulling `chronos-forecasting` into the
-  workspace would downgrade `huggingface_hub`, which then breaks the F9
-  TimesFM-2.0 wrapper and the `apps/forecaster` HF Hydrator.
+- `seq_len`: `192`
+- `horizon`: `16`
 
-**Resolution path (post-thesis):** isolate Chronos-Bolt inference in its own
-extra (`uv pip install -e . --extra chronos`) with a separate locked env, or
-wait for an upstream `chronos-forecasting` release that supports
-`huggingface_hub>=1.0`.
+## Feature set (1 columns)
 
-**Existing result (pre-conflict, kept for the appendix):** val pinball
-~263 DKK at n=1000, beats F7 by ~1 DKK with zero training. See
-`seed-42/metrics.json`.
+- `(univariate: imbalance_price_dkk_mwh_15min)`
 
-**License:** Apache-2.0 (Chronos-Bolt weights).
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **263.18** |
+| Raw [q10, q90] coverage | 75.82% |
+| ACI marginal coverage (target 90%) | 89.89% |
+| ACI mean width (DKK) | 2083.59 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/f10/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

f11/MODEL_CARD.md

@@ -1,53 +1,44 @@
-# F11 — PriceFM-shaped surrogate (model card)
-
-**Status**: trained 2026-05-10.  Backed by a patch-TST architecture
-following the *PriceFM* design dimensions (arXiv:2508.04875), trained
-from scratch on DK1.  This is a **PriceFM-shaped surrogate**, not a
-fine-tune of authored PriceFM weights — those are not publicly
-released as of 2026-05-10.
-
-## Architecture (vs F7 / F8)
-
-| Hyperparameter | F7 (default) | **F11 (this)** |
-|----------------|--------------|----------------|
-| seq_len        | 96 (24 h)    | 192 (48 h)     |
-| patch_len      | 8            | 16             |
-| d_model        | 128          | 192            |
-| n_layers       | 6            | 8              |
-| nhead          | 8            | 8              |
-| dropout        | 0.1          | 0.1            |
-| epochs         | 5            | 5              |
-| lr             | 1e-3         | 5e-4           |
-| ~params        | ~1.2 M       | ~3.5 M         |
-
-The wider, deeper architecture mimics PriceFM's documented scale-up
-(§3.1) while staying within the ~10 M-param budget the proposal §4.2.2
-allows.  The 48-hour context window exposes daily seasonality plus
-yesterday's residual to the patches.
-
-## Why this is *not* PriceFM
-
-- The published PriceFM is pretrained on 24 EU countries / 38 regions
-  for transfer-learning leverage; we trained from scratch on DK1
-  alone.  Reviewer-defensible only as a *surrogate*; numbers should
-  not be cited as PriceFM's.
-- The graph-based inductive biases for transmission topology
-  (PriceFM §3.2) are absent.
-
-## When to swap in real PriceFM weights
-
-If the PriceFM authors publish a HuggingFace checkpoint, the swap is
-~10 lines: extend
-`apps/forecaster/.../inference/backends/f11_pricefm.py` to load from
-the published repo first, fall back to this surrogate otherwise.
-
-## Reproduction
-
-```bash
-PYTHONPATH=. python -m heimdall_forecaster.train.run \
-  --config apps/forecaster/src/heimdall_forecaster/train/configs/f11.yaml \
-  --seed 42
-```
-
-(Repeat for seeds 13/137/1729/31415 to populate the canonical 5-seed
-sweep.)
+# f11 — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+- `seq_len`: `192`
+- `horizon`: `16`
+- `patch_len`: `16`
+- `d_model`: `192`
+- `nhead`: `8`
+- `n_layers`: `8`
+- `dropout`: `0.1`
+- `epochs`: `5`
+- `batch_size`: `64`
+- `lr`: `0.0005`
+- `weight_decay`: `0.0001`
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **265.59 ± 4.19** |
+| Raw [q10, q90] coverage | 64.15 ± 2.42% |
+| ACI marginal coverage (target 90%) | 89.74 ± 0.03% |
+| ACI mean width (DKK) | 2542.28 ± 262.44 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/f11/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

f12_ebm/MODEL_CARD.md

@@ -0,0 +1,44 @@
+# f12_ebm — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+- `seq_len`: `96`
+- `horizon`: `16`
+- `patch_len`: `8`
+- `d_model`: `128`
+- `nhead`: `8`
+- `n_layers`: `6`
+- `dropout`: `0.1`
+- `epochs`: `30`
+- `batch_size`: `64`
+- `lr`: `0.0002`
+- `weight_decay`: `0.0001`
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **263.05 ± 1.98** |
+| Raw [q10, q90] coverage | 64.02 ± 3.18% |
+| ACI marginal coverage (target 90%) | 89.71 ± 0.03% |
+| ACI mean width (DKK) | 2426.90 ± 95.37 |
+| Inference runtime (s) | 11.02 ± 6.15 |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/f12_ebm/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

f12_ebm/seed-13/config.json

@@ -0,0 +1,35 @@
+{
+  "name": "f12_ebm",
+  "train_panel": "/work/heimdall/data/processed/dk1_panel_train.parquet",
+  "val_panel": "/work/heimdall/data/processed/dk1_panel_val.parquet",
+  "out_dir": "/work/heimdall/models/forecaster/f12_ebm/seed-13",
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 30,
+  "batch_size": 64,
+  "lr": 0.0002,
+  "weight_decay": 0.0001,
+  "seed": 13,
+  "sigma_data": 1.0,
+  "sigma_min": 0.002,
+  "sigma_max": 80.0,
+  "rho": 7.0,
+  "p_mean": -1.2,
+  "p_std": 1.2,
+  "n_sampler_steps": 32,
+  "n_inference_samples": 256,
+  "multivariate": false,
+  "quantiles": [
+    0.1,
+    0.5,
+    0.9
+  ],
+  "n_features": 1,
+  "method": "EDM-Karras2022 denoiser parameterisation",
+  "runtime_seconds": 5.3894078731536865
+}

f12_ebm/seed-13/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2079e67bd56873351937aff645a62684bdf78cf48aab91f65ae3554b1fd0262
+size 9092083

f12_ebm/seed-13/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f12_ebm/seed-137/config.json

@@ -0,0 +1,35 @@
+{
+  "name": "f12_ebm",
+  "train_panel": "/work/heimdall/data/processed/dk1_panel_train.parquet",
+  "val_panel": "/work/heimdall/data/processed/dk1_panel_val.parquet",
+  "out_dir": "/work/heimdall/models/forecaster/f12_ebm/seed-137",
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 30,
+  "batch_size": 64,
+  "lr": 0.0002,
+  "weight_decay": 0.0001,
+  "seed": 137,
+  "sigma_data": 1.0,
+  "sigma_min": 0.002,
+  "sigma_max": 80.0,
+  "rho": 7.0,
+  "p_mean": -1.2,
+  "p_std": 1.2,
+  "n_sampler_steps": 32,
+  "n_inference_samples": 256,
+  "multivariate": false,
+  "quantiles": [
+    0.1,
+    0.5,
+    0.9
+  ],
+  "n_features": 1,
+  "method": "EDM-Karras2022 denoiser parameterisation",
+  "runtime_seconds": 19.5199031829834
+}

f12_ebm/seed-137/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:158e1df2fd1bc6edf142361c7c8ba20075e5022aeb261dddf10fcd50ee9208af
+size 9092083

f12_ebm/seed-137/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f12_ebm/seed-1729/config.json

@@ -0,0 +1,35 @@
+{
+  "name": "f12_ebm",
+  "train_panel": "/work/heimdall/data/processed/dk1_panel_train.parquet",
+  "val_panel": "/work/heimdall/data/processed/dk1_panel_val.parquet",
+  "out_dir": "/work/heimdall/models/forecaster/f12_ebm/seed-1729",
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 30,
+  "batch_size": 64,
+  "lr": 0.0002,
+  "weight_decay": 0.0001,
+  "seed": 1729,
+  "sigma_data": 1.0,
+  "sigma_min": 0.002,
+  "sigma_max": 80.0,
+  "rho": 7.0,
+  "p_mean": -1.2,
+  "p_std": 1.2,
+  "n_sampler_steps": 32,
+  "n_inference_samples": 256,
+  "multivariate": false,
+  "quantiles": [
+    0.1,
+    0.5,
+    0.9
+  ],
+  "n_features": 1,
+  "method": "EDM-Karras2022 denoiser parameterisation",
+  "runtime_seconds": 14.711588859558105
+}

f12_ebm/seed-1729/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd8c3bbe3ff74550f95bc9ca4a872ccf446821d015c5e08386d40c39dfbbf68b
+size 9092083

f12_ebm/seed-1729/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f12_ebm/seed-31415/config.json

@@ -0,0 +1,35 @@
+{
+  "name": "f12_ebm",
+  "train_panel": "/work/heimdall/data/processed/dk1_panel_train.parquet",
+  "val_panel": "/work/heimdall/data/processed/dk1_panel_val.parquet",
+  "out_dir": "/work/heimdall/models/forecaster/f12_ebm/seed-31415",
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 30,
+  "batch_size": 64,
+  "lr": 0.0002,
+  "weight_decay": 0.0001,
+  "seed": 31415,
+  "sigma_data": 1.0,
+  "sigma_min": 0.002,
+  "sigma_max": 80.0,
+  "rho": 7.0,
+  "p_mean": -1.2,
+  "p_std": 1.2,
+  "n_sampler_steps": 32,
+  "n_inference_samples": 256,
+  "multivariate": false,
+  "quantiles": [
+    0.1,
+    0.5,
+    0.9
+  ],
+  "n_features": 1,
+  "method": "EDM-Karras2022 denoiser parameterisation",
+  "runtime_seconds": 10.213236093521118
+}

f12_ebm/seed-31415/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3265fc947180eef1507db4430d1b2ea3f33fc3504f6e04ab2c0cbe13b7b9615b
+size 9092083

f12_ebm/seed-31415/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f12_ebm/seed-42/config.json

@@ -0,0 +1,35 @@
+{
+  "name": "f12_ebm",
+  "train_panel": "/work/heimdall/data/processed/dk1_panel_train.parquet",
+  "val_panel": "/work/heimdall/data/processed/dk1_panel_val.parquet",
+  "out_dir": "/work/heimdall/models/forecaster/f12_ebm/seed-42",
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 30,
+  "batch_size": 64,
+  "lr": 0.0002,
+  "weight_decay": 0.0001,
+  "seed": 42,
+  "sigma_data": 1.0,
+  "sigma_min": 0.002,
+  "sigma_max": 80.0,
+  "rho": 7.0,
+  "p_mean": -1.2,
+  "p_std": 1.2,
+  "n_sampler_steps": 32,
+  "n_inference_samples": 256,
+  "multivariate": false,
+  "quantiles": [
+    0.1,
+    0.5,
+    0.9
+  ],
+  "n_features": 1,
+  "method": "EDM-Karras2022 denoiser parameterisation",
+  "runtime_seconds": 5.274056911468506
+}

f12_ebm/seed-42/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94ac6634bb5069c0a6202ff3baf0ff2b69e1c92e431c431965886bac597c9727
+size 9092083

f12_ebm/seed-42/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f13/MODEL_CARD.md

@@ -0,0 +1,44 @@
+# f13 — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+- `seq_len`: `96`
+- `horizon`: `16`
+- `patch_len`: `8`
+- `d_model`: `128`
+- `nhead`: `8`
+- `n_layers`: `6`
+- `dropout`: `0.1`
+- `epochs`: `100`
+- `batch_size`: `64`
+- `lr`: `0.001`
+- `weight_decay`: `0.0001`
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **310.47 ± 3.30** |
+| Raw [q10, q90] coverage | 24.78 ± 1.22% |
+| ACI marginal coverage (target 90%) | 89.82 ± 0.03% |
+| ACI mean width (DKK) | 2857.11 ± 41.98 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/f13/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

f13/seed-13/config.json

@@ -0,0 +1,20 @@
+{
+  "name": "f13",
+  "multivariate": false,
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 100,
+  "batch_size": 64,
+  "lr": 0.001,
+  "weight_decay": 0.0001,
+  "seed": 13,
+  "n_features": 31,
+  "n_train_windows": 174974,
+  "n_val_windows": 5361,
+  "out_dir": "/work/heimdall/models/forecaster/f13/seed-13"
+}

f13/seed-13/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6675509b54878a8ed7fe932c919a734f4741f711d007db447c590079e8efcd83
+size 5212610

f13/seed-13/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d88f329fa4e80cbd2d211b9e77c56cb079f3f43bd30b730070c6ab25a7ca95b6
+size 1409

f13/seed-137/config.json

@@ -0,0 +1,20 @@
+{
+  "name": "f13",
+  "multivariate": false,
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 100,
+  "batch_size": 64,
+  "lr": 0.001,
+  "weight_decay": 0.0001,
+  "seed": 137,
+  "n_features": 31,
+  "n_train_windows": 174974,
+  "n_val_windows": 5361,
+  "out_dir": "/work/heimdall/models/forecaster/f13/seed-137"
+}

f13/seed-137/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9adebbf2323848a59425a7d71c7b1f01014acd49f0814fae1ad9b959bb11c5bd
+size 5212610

f13/seed-137/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d88f329fa4e80cbd2d211b9e77c56cb079f3f43bd30b730070c6ab25a7ca95b6
+size 1409

f13/seed-1729/config.json

@@ -0,0 +1,20 @@
+{
+  "name": "f13",
+  "multivariate": false,
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 100,
+  "batch_size": 64,
+  "lr": 0.001,
+  "weight_decay": 0.0001,
+  "seed": 1729,
+  "n_features": 31,
+  "n_train_windows": 174974,
+  "n_val_windows": 5361,
+  "out_dir": "/work/heimdall/models/forecaster/f13/seed-1729"
+}

f13/seed-1729/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55b368a40ef3f5354f2aabe7be623d42abb404488e9c1e19445a62cf808fa379
+size 5212610

f13/seed-1729/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d88f329fa4e80cbd2d211b9e77c56cb079f3f43bd30b730070c6ab25a7ca95b6
+size 1409

f13/seed-31415/config.json

@@ -0,0 +1,20 @@
+{
+  "name": "f13",
+  "multivariate": false,
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 100,
+  "batch_size": 64,
+  "lr": 0.001,
+  "weight_decay": 0.0001,
+  "seed": 31415,
+  "n_features": 31,
+  "n_train_windows": 174974,
+  "n_val_windows": 5361,
+  "out_dir": "/work/heimdall/models/forecaster/f13/seed-31415"
+}

f13/seed-31415/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3c965ad8e5baab6a72102eca0f72682c4d8a4652209e2dfa8fd7450c3b23c07
+size 5212610

f13/seed-31415/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d88f329fa4e80cbd2d211b9e77c56cb079f3f43bd30b730070c6ab25a7ca95b6
+size 1409

f13/seed-42/config.json

@@ -0,0 +1,20 @@
+{
+  "name": "f13",
+  "multivariate": false,
+  "seq_len": 96,
+  "horizon": 16,
+  "patch_len": 8,
+  "d_model": 128,
+  "nhead": 8,
+  "n_layers": 6,
+  "dropout": 0.1,
+  "epochs": 100,
+  "batch_size": 64,
+  "lr": 0.001,
+  "weight_decay": 0.0001,
+  "seed": 42,
+  "n_features": 31,
+  "n_train_windows": 174974,
+  "n_val_windows": 5361,
+  "out_dir": "/work/heimdall/models/forecaster/f13/seed-42"
+}

f13/seed-42/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fb29bc022ea4756cf22a0d28b5b88d14c835f38a6f2a15f937f1445eb03174b
+size 5212610

f13/seed-42/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d88f329fa4e80cbd2d211b9e77c56cb079f3f43bd30b730070c6ab25a7ca95b6
+size 1409

f1_lgbm/MODEL_CARD.md

@@ -0,0 +1,34 @@
+# f1_lgbm — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+_(no config.json schema; see trainer source)_
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **262.18 ± 0.08** |
+| Raw [q10, q90] coverage | 58.93 ± 0.09% |
+| ACI marginal coverage (target 90%) | 89.71 ± 0.02% |
+| ACI mean width (DKK) | 2419.28 ± 21.56 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/f1_lgbm/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

f1_lgbm/seed-13/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f1_lgbm/seed-137/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f1_lgbm/seed-1729/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f1_lgbm/seed-31415/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f1_lgbm/seed-42/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f2_blr/MODEL_CARD.md

@@ -0,0 +1,34 @@
+# f2_blr — model card
+
+Auto-generated by `tools/generate_model_cards.py` on 2026-05-17.
+
+## Default configuration (out-of-box hyperparameters)
+
+_(no config.json schema; see trainer source)_
+
+## Feature set (1 columns)
+
+- `(univariate: imbalance_price_dkk_mwh_15min)`
+
+## Frozen-seed results (5 seeds: 13, 42, 137, 1729, 31415)
+
+| Metric | Value |
+|---|---|
+| Val pinball mean (DKK) | **382.13 ± 0.00** |
+| Raw [q10, q90] coverage | 61.33 ± 0.00% |
+| ACI marginal coverage (target 90%) | 89.24 ± 0.00% |
+| ACI mean width (DKK) | 4438.15 ± 0.00 |
+| Inference runtime (s) | — |
+
+## Provenance
+
+- Trained on `data/processed/dk1_panel_{train, val}.parquet` (FX-corrected 2026-05-16)
+- Seeds: `[13, 42, 137, 1729, 31415]` (frozen project-wide)
+- Train cutoff: 2025-02-28 UTC
+- Val window: 2025-03-04 → 2025-04-30 UTC (post-EAM regime)
+- Per-seed artifacts at `models/forecaster/f2_blr/seed-<n>/`:
+  - `model.pt` / `model.npz` — weights
+  - `stats.pkl` — train-stat normalization (mean/std/target stats)
+  - `val_preds.npz` — quantile predictions on val
+  - `metrics.json` — pinball / coverage / ACI numbers
+  - `aci_state.json` — final ACI state (α_t, buffer)

f2_blr/seed-13/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f2_blr/seed-137/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362

f2_blr/seed-1729/stats.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42b420df6e276c4b8addf55410d19405be8277d42ca1a294a3bee1b079d1a84
+size 362