GBP/JPY market-structure demo — statistics

Window 2025-11-02 → 2026-10-30 (11.9 months) · all times UTC · capability demo, not a validated strategy

Rates with their uncertainty

MeasureAll (12 months)Discovery (first 70%)HOLDOUT (last 30%)
Independent alerts (4 of 5)743
Completion rate (alerts reaching step 5)4/7 = 57%
95% CI 25%–84%
2/4 = 50%
95% CI 15%–85%
2/3 = 67%
95% CI 21%–94%
Win rate (wins ÷ closed outcomes)1/4 = 25%
95% CI 5%–70%
1/2 = 50%
95% CI 9%–91%
0/2 = 0%
95% CI 0%–66%
Best move after entry (median R, middle half)1.2R (0.8–1.7)1.9R (1.5–2.4)0.8R (0.6–1.1)
Worst move after entry (median R, middle half)0.6R (0.2–1.6)0.6R (0.4–0.7)1.9R (1.0–2.8)
Failure modes (completed / lost / invalidated / expired)4 / 3 / 0 / 32 / 1 / 0 / 22 / 2 / 0 / 1
Every rate shows successes/n and a 95% Wilson interval. n is the number of INDEPENDENT setups (sequences on the same swing or zone within 24 H1 candles count once). With samples this small the intervals are wide - read them, not the point values.

How often the setup formed

LevelRaw sequencesIndependent setupsPer month
Watch37252.1
Alert1470.59
Complete1040.34
25 independent setups reached Watch, 7 reached Alert and 4 completed over 11.9 months (about 0.59 alerts a month).
Across 4 completed setups the median best move was 1.2R and the median worst move 0.6R (middle half: 0.8 to 1.7R best, 0.2 to 1.6R worst).

Trading sessions (discovery item preview)

Each session holds only 2 alerts (Asia), 2 alerts (London), 3 alerts (NewYork), 0 alerts (Off-hours); the intervals overlap, so no session difference is established.

Holdout

Discovery: 4 alerts, completion 50%, win 50%. Holdout: 3 alerts, completion 67%, win 0%. Samples this small cannot separate the two.

Control: is it better than random timing?

The setup does NOT clearly beat the random-timing control (its interval overlaps it). Control = same 2R-before-1R rule at random times matching hour-of-day and ATR(H1). Control trades overlap, so treat its interval as optimistic.

Discovery list (kept apart from Whits's own rules)

IDFindingHoldout resultStatus
RC-001Completion rate by trading sessionnot enough discovery-period data to form a claimfailed
RC-002Win rate by trading sessionnot enough discovery-period data to form a claimfailed
RC-003Deeper sweeps follow through further (median MFE in R)discovery period has too few completed setups to form a claimfailed
Findings are written to results/research_candidates.csv as “unvalidated”, tested only on the holdout, and never change an alert until Whits approves them in writing. Findings that fail stay on the list marked “failed”.