GBP/JPY market-structure recognition demo: method note
One setup · one pair (GBP/JPY) · four timeframes (D1, H4, H1, M15) · all times UTC
This is a capability demo, not a validated strategy. It shows what the system can recognise, what is still unconfirmed, and how the pattern behaved. It does not tell anyone to trade. Whits decides every trade himself.
The setup, in plain words (bearish; bullish is the exact mirror)
- 1 Level reached. Price trades within half an ATR(H4) of a confirmed major D1 or H4 swing high.
- 2 H4 sweep. One H4 candle pokes above that high and closes back below it.
- 3 H1 displacement. Within 12 H1 candles after the sweep, a fast drop of at least 1.5 x ATR(H1) in 3 candles or fewer that also closes below the last confirmed H1 swing low (a break of structure).
- 4 M15 change of character. The first M15 close below the last confirmed M15 higher low.
- 5 Return to zone. Price retraces into the supply zone (the last up-candle before the drop).
Steps must happen in order. Step 5 must follow within 48 H1 candles of step 3, or the sequence expires. A close above the sweep high before step 5 invalidates it. Alerts: Watch at 3 steps, Alert at 4 (the missing step is named), Complete at 5.
How things are confirmed
- Swings use an ATR zig-zag: a high or low only counts once price has moved 2.0 x ATR (D1, H4) or 1.5 x ATR (H1, M15) away from it. Every break is judged on the candle close, never the wick.
- Every label has two times: when it happened and when it could first be known. Nothing is drawn or used before the second time. A forming D1 or H4 candle is never used.
- Outcome rule (for statistics only): enter at the zone midpoint, stop at the sweep high + 0.25 x ATR(H1); a win is +2R before -1R within 72 H1 candles, otherwise a loss. A candle touching both counts as a loss.
Data and the checks that were run
- Source: dukascopy M1 candles, 2025-11-02 to 2026-10-30, 346,146 one-minute bars. All four timeframes are built from that one store (never mixed). Day boundary: New York 17:00 (to be confirmed with Whits).
- Checks: no duplicate or out-of-order timestamps; every gap longer than 5 minutes outside the weekend is listed (130 found); price sanity (high/low consistent, no zero or negative prices, no 3% one-minute jump); 254 D1 candles over 0.99 years = 257 per year (accepted 245-270).
- Failed checks: none. Second-source spot check of 3 random days: MANUAL (No second source supplied - compare these 3 D1 candles with HistData or your broker chart (NY 17:00 close).)
Proof that nothing used future data
The full labeller was run on the whole history. Then, for 200 random times, the raw data was cut at that moment and everything was rebuilt from scratch. Every label known by that moment had to match the full run exactly (2,747,077 label comparisons), and the alert log was re-checked the same way at 25 of those times. Result: PASSED, 0 differences.
What the numbers say (with their uncertainty)
- Independent setups: 25 reached Watch, 7 reached Alert, 4 completed, over 11.9 months. Sequences on the same swing or zone within 24 H1 candles count once.
- Completion rate of alerts: 4/7 (57%; 95% interval 25%–84%). Win rate of completed setups: 1/4 (25%; 95% interval 5%–70%).
- Holdout (last 30%, used once, nothing tuned on it): 3 alerts, completion 67%, win 0%.
- Control (same rule at random times, matched on hour and volatility): The setup does NOT clearly beat the random-timing control (its interval overlaps it).
Stated limits
- Small sample. A few dozen sequences at most, so every rate is illustrative and its interval is wide.
- Indicative prices. Dukascopy quotes are indicative, so spreads and fills differ from a broker's. Volume is tick count, not traded volume; the demo makes no volume claims.
- One setup, one pair, defaults. All thresholds are starting defaults in one config file. Some rules the brief leaves open were settled with documented defaults (sweep must follow step 1 within 5 days; a level older than 60 days is retired; a zone touched before the M15 change of character invalidates the sequence). Whits may well mark things differently.
- Discovery is separate. The one data-driven finding (trading sessions) sits in research_candidates.csv as “unvalidated”, was tested only on the holdout and changes nothing until Whits approves it in writing.