42Analytics

Evidence / reviewed 20 October 2026

Every result
needs its context.

A forecast score means little without its baseline, population and test conditions. Here is the record behind FishFacts, including the limits.

How to read these figures. Each percentage is a reduction in mean absolute forecast error (MAE) relative to a causal rolling mean of five genuine observations (RM5). The baseline uses information available at the recorded forecast anchor. A horizon of H4 means four market observations ahead. Counts are resolved target–horizon outcomes, not independent days. Pooled scores combine forecast errors; species-balanced scores give each species equal weight. Different models and evaluation windows make these unsuitable as a market league table. They are not measured client savings.

Production-model historical evaluation

Peterhead

6.90%

lower pooled MAE vs causal RM5 · H4–H10

162,634 resolved outcomes · 5 January 2024–31 December 2025. Species-balanced improvement: 6.11%.

The separate H1–H3 evaluation records 6.43% lower pooled MAE (6.37% species-balanced), across 26,732 outcomes from 1 January–11 September 2026. These windows are kept separate.

Historical confirmation of a defined model and dataset. This does not establish the performance of every live forecast or every species.

Peterhead scorecard

Historical zero-shot challenger

Shetland

7.73%

lower pooled MAE vs causal RM5 · H1–H10

172,176 resolved outcomes · frozen historical replay. Species-balanced improvement: 7.03%.

Peterhead parameters transferred to Shetland. The September 2026 release audit verifies engine sample parity; the market adapter remains external to that release.

Historical confirmation, not a prospective operating record. The summary does not establish a single overall date window; consult the underlying replay record.

Cross-market evidence

Release-owned challenger

Vigo

3.24%

lower pooled MAE vs causal RM5 · H4–H10

621,748 resolved outcomes · 1 January 2020–31 December 2025. Species-balanced improvement: 3.94%.

A frozen confirmation set covering 72 eligible species and 103 target series. The challenger uses a market-specific configuration; its September 2026 scorecard records release ownership and parity.

Short-horizon replays and interval experiments are separate exploratory evidence. This result does not demonstrate procurement savings.

Vigo scorecard

Research only · opened-data diagnostic

Korea

7.46%

lower pooled MAE vs causal RM5 · H4–H10

1,543,504 resolved outcomes · already-opened 2024–2026 data, through 30 June 2026. Species-balanced improvement: 7.34%.

A research diagnostic using the Korean dataset’s recorded end-of-session anchor rules. The data had already been examined during development.

Development evidence only. Not prospective validation, a promoted production model or commercial proof. Interval calibration is not established by this result.

Research scope and evidence

Prediction ranges

Coverage is measured,
not assumed.

A Peterhead diagnostic covering 60,361 resolved outcomes from 7 January–25 September 2026 found that nominal 80% ranges covered 63.14% of outcomes (62.85% when balanced by date).

Those ranges were too narrow on that sample. A retrospective candidate improved coverage, but had no prospective results in the audit. We therefore make no general claim of calibrated price intervals across species, grades or markets.

Read the diagnostic and its scope ↗

Product development

What the work
demonstrates.

FishFacts demonstrates canonical data handling, versioned evidence, traceable forecast records and operational alerting. The engineering note explains how those pieces fit together.

FishFacts engineering note →

Strata is a research prototype for historical well exploration and drilling-duration estimates. It has not been used in a client engagement, and we claim no measured client outcomes.

Strata prototype and scope →

Let’s build something that lasts

What needs to
work better?

Bring the difficult pipeline, the cloud question, or the idea you want to put into production. Start with a 30-minute conversation with Pete.

Arrange a 30-minute call

Cruden Bay, Aberdeenshire · Working with teams across the UK and beyond.