# This month: where the model was wrong

Month: 2026-08 · generated 2026-09-18

0 graded point(s) across 0 asset(s), from 2644 frozen ledger row(s).

Nothing has matured yet: the newest gradable reported year is 2025.

## Method

**What counts as a forecast.** A row in data/ledger/forecast-ledger.jsonl - the append-only record of the forward curve as it stood on a dated snapshot - with snapshotId on or before the last day of this month. Today's model is never consulted; that would be a backtest (lib/calibration.ts), not a track record.

**What counts as an actual.** A cited FY figure in the asset's `reported` record (lib/reference.ts), for a year at or before 2025. A fiscal year is reported in the year after it closes, so nothing later can have an actual yet.

**Unit of comparison.** One (asset, snapshot, year). Predicted is the SUM of that snapshot's program curves for that asset; the ledger is per indication program and `reported` is per asset, so a program row on its own has no comparable actual.

**Known bias.** The indication build-up covers eight markets and only the indications that carry a model; a reported FY figure is worldwide and covers the whole label. Both gaps push the build-up DOWN, so some part of any under-call here is definitional rather than error. See the SCOPE block in lib/forecastReconciliation.ts. Nothing is rescaled to close the gap.

**Determinism.** No model call, no network, no randomness. --month and --generated-at pin the only two clock reads.

**Excluded from this month's grading:**
- Rows frozen after this month: a scoreboard for a month cannot use a curve that did not exist yet.
- Non-FY reported rows (Q1/Q2/Q3/PART): real public facts, but not comparable to a full-year curve. A PART row is usually an acquisition stub covering part of a year.
- Reported years after 2025: not reported yet, so there is nothing to grade against.
- Outcome years at or before the snapshot year: knowable when the curve was frozen, so in-sample and excluded.
- (asset, year) pairs where any of that snapshot's program curves does not cover the year: a sum missing a contributor is a different forecast, not a smaller one.
- Reported values of zero or less: no usable denominator for a percentage error.

**Sources:** data/ledger/forecast-ledger.jsonl, lib/reference.ts (reported actuals, cited)

> Ormeval provides decision-support intelligence derived from public sources. It is not investment, medical, legal or regulatory advice, and states nothing about the value of any security. Forward figures are deterministic model estimates, not guidance.