Prospective resolved human forecasts
0Only blind submissions made before the comparison lanes were visible can enter.
Scientific forecast evaluation
Does human context improve the model—and does the market add more?
Two frozen arms answer that question on later event families: human + model, then the same system + a matched prediction-market sensor. Every input lane stays separately scored.
not eligible no prospective matched human outcomes
Already-public receipts excluded
7They cannot be backfilled as human estimates without anchoring and hindsight contamination.
Production hybrid probability
NoneNo weight or combined number is published before both held-out studies pass.
Current scientific result
7 current receipts were already public before the human lane opened and are therefore excluded. No prospective resolved human forecast is available; both hybrid arms, every learned weight, and the combined probability remain unavailable.
MacroGuru has preregistered the comparison, not demonstrated hybrid superiority. A negative result will remain public.
The two-arm test
Without prediction market
Blind analyst point probability + statistical-model point probability. This tests whether first-party context improves the model on its own.
With prediction market
The same two lanes + a pre-lock matched market point. Market data is leverage, but cannot hide whether MacroGuru adds independent value.
- Analysts see the atomic claim, resolver, horizon, and frozen evidence packet—not model, market, published, or peer probabilities.
- A point probability is required for Brier and log-loss scoring. An optional range describes epistemic uncertainty and is never converted to its midpoint.
- Revisions are append-only, time-locked, actor-bound, and linked through explicit supersession.
- Weights are non-negative, sum to one, and are fitted only on earlier resolved event families.
- Promotion requires two temporal holdouts, at least 30 held-out families overall, 10 per domain, calibration, paired uncertainty, and improvement over every available constituent and comparator.
Preregistered temporal holdouts
| Window | Held-out period | Earlier training families | Held-out families | Status |
|---|---|---|---|---|
| holdout-2026q4-2027q1 | 2026-11-01 → 2027-01-31 | 0 | 0 | Not eligible |
| holdout-2027q1-2027q2 | 2027-02-01 → 2027-04-30 | 0 | 0 | Not eligible |
Protocol receipt
15f5ed26a3e35228445fc35c89ccce6c8124333502ab1abfb7b4111ab91ca5cd
Evaluation receipt
f4f61935bf2baf1a7f1397e107ec1c0262ce7f203fccc9d2e49716f2984e50da