How We Measure Accuracy

The Brain researches major economic events before they happen, then examines the actual results and market reactions afterwards. This page explains exactly how those assessments become a historical accuracy record.

How accuracy is calculated

Only claims explicitly stated in the approved pre-event research count. Each is assessed as correct, partially correct, incorrect, inconclusive or not assessable.

Accuracy = (correct + 0.5 × partially correct) ÷ (correct + incorrect + partially correct). A partially correct assessment therefore earns half credit.

Inconclusive and non-assessable outcomes

Inconclusive and not-assessable results are excluded from the accuracy figure, but they are always disclosed alongside it.

Market moves smaller than the noise threshold (EUR/USD 10 pips, Gold (XAU/USD) $5) are treated as inconclusive for direction.

Baseline accuracy and conditional scenario performance

Reported baseline accuracy uses reported consensus or baseline outcomes, not independent Brain forecasts.

Conditional scenario performance is shown separately. A scenario is not credited simply because one of several possibilities occurred: only a scenario whose condition was met is scored, and only on its stated market direction.

These categories are never combined into one headline percentage.

Minimum sample size

No accuracy figure or colour is shown until an event type has 20 assessed events. Until then, The Brain shows its progress towards that threshold.

Why each event counts only once

Claims derived from the same prediction share one prediction group and count once. Instruments and reaction windows are evidence for that one group, never extra observations.

Event-level accuracy counts each event once per category, so market measurements cannot inflate the record.

Original research and administrator approval

The Brain never rewrites its original research after an event. Claims are frozen and approved before the release.

Every result is checked by an administrator before it appears.

Recording results does not retrain the AI

Recording results does not mean the AI automatically learns or improves; any change to its methods requires further development and testing.

Historical performance limitations

Historical accuracy describes past assessments only. It is not the probability that the next event will be called correctly.

All figures are historical performance under scoring rules brain-score-v1, not predictions of future outcomes or trading guarantees.

← Back to The Brain