RacineCast-1-NC

RacineCast-1-NC is an agentic time-series forecasting system by Racine.ai: a fixed per-quantile weighted blend of public foundation-model forecasts. Zero-shot, no training, probabilistic outputs (quantiles 0.1–0.9).

#1 on GIFT-Eval on both leaderboard views (MASE_Rank and WQL_Rank) when added to the 2026-07-07 board, ahead of CastStar, independently verified (see Integrity). This is the strongest edition; one member is non-commercial, so this repository is CC-BY-NC-SA-4.0. For commercial use: RacineCast-1 (all-Apache, also #1 both views when submitted alone).

Recipe

0.55 · Toto-2.0-FnF + 0.16 · Toto-2.0-2.5B + 0.16 · TiRex-2
+ 0.045 · Chronos-2 + 0.045 · FlowState-r1.1 + 0.03 · PatchTST-FM-r1

Weighted average per quantile level ("vincentization"); the constants sum to 0.99 and are normalized to 1 at blend time (independently verified). Weights are fixed, pre-registered in git before any test evaluation, never fitted on test data.

Member	Weight	Source	License
Toto-2.0-FnF	0.55	Datadog/Toto-2.0-Family-and-Friends	Apache-2.0
Toto-2.0-2.5B	0.16	Datadog/Toto-2.0-2.5B	Apache-2.0
TiRex-2	0.16	NX-AI/TiRex-2-gifteval-pretrain	Apache-2.0
Chronos-2	0.045	amazon/chronos-2	Apache-2.0
FlowState-r1.1	0.045	ibm-research/flowstate	Apache-2.0
PatchTST-FM-r1	0.03	ibm-research/patchtst-fm-r1	CC-BY-NC-SA-4.0

PatchTST-FM-r1's weights are CC-BY-NC-SA-4.0; we conservatively license this whole edition the same way. Original model repositories retain their respective licenses and terms; transitive notes in PROVENANCE.md. Not legal advice.

Results

Official numbers will be published on the GIFT-Eval leaderboard once the submission is merged. Measurement tables, margins, and caveats: PROVENANCE.md.

Artifact layout

RacineCast-1-NC/
├── gift_eval/            # leaderboard entry: all_results.csv + config.json
├── reproduction/         # blend + evaluation code as used for the entry
├── audit/
│   ├── recompute_board.py    # recompute the leaderboard ranking yourself
│   └── all_candidates/       # all 28 pre-registered candidate blends, raw
├── assets/
└── PROVENANCE.md         # full provenance, disclosures, license notes

No model weights ship here by design: the neural weights live in the member repositories; this bundle contains the frozen recipe and everything needed to verify and reproduce the entry.

Reproduction

1. Install the official gift-eval harness and Salesforce/GiftEval data.
2. Member forecasts: Datadog FnF bundle + TiRex-2 via the official NX-AI/tirex-2 package.
3. Blend and score with reproduction/blend_eval.py; rank any day's board with audit/recompute_board.py.

Integrity

• Harness exactness: our FnF rebuild matches Datadog's official CSV (max deviation 5.1e-7, reproduced cross-platform).
• Recipes pre-registered before test evaluation; the published candidate was selected among 28 disclosed candidates (all raw CSVs in audit/all_candidates/, selection disclosed verbatim in PROVENANCE.md).
• Independently counter-audited end-to-end (board recomputation, pipeline rebuild, blend tensor recomposition): numbers exact. Summary in PROVENANCE.md.
• testdata_leakage: No: no training or weight fitting on test data.