Dukascopy Tick Data · OOS 2022–2026 · IS 2017–2022 · commission included
After a large S&P 500 down day (below 0.6 standard deviations of its trailing 60-day returns), buy USD/JPY at 23:00 UTC and sell at 20:55 UTC the next day, earning one night of swap. Profitable in every in-sample year 2017-2021 (PF 2.09, +$7,200 on 150 trades); out of sample 2022-2025 it lost $32 on 172 trades (PF 1.00).
USD/JPY · C1
Does not hold up
$10k → $10.0k
-$32 net loss
Win Rate: 37.21% · PF: 0.997 · Max DD: 21.02%
USD/JPY · C2
Does not hold up
$10k → $9.7k
-$283 net loss
Win Rate: 48.26% · PF: 0.957 · Max DD: 12.61%
Strategy Rules
Reference series: the S&P CFD daily return r_t = USA500_CUSTOM close at 21:00 UTC today / close at 21:00 UTC on the previous trading day - 1 (read via the H1 bar that closes at 21:00 UTC; on days the index has no 21:00 bar, use the last H1 close before it). sigma_t = standard deviation of the previous 60 daily returns r_{t-60..t-1} (not including today).
Signal: r_t < -SigmaK * sigma_t (a large S&P DOWN day).
Entry: BUY USDJPY at market at the first tick at or after EntryHourUTC:00 (default 23:00 UTC) on the signal day. If the signal day is a Friday there are no FX bars at 23:00 (week ends 21:00 UTC): if TradeFridaySignals=true enter at the first tick at or after Monday 00:05 UTC (custom symbols TRADE from Monday 00:00 UTC; Sunday 21:00-24:00 is quote-only), else skip the signal.
One position at a time; a new signal while a position is open (consecutive selloff days) does not add - it re-arms the exit clock to the newer day's 20:55 UTC only if ExtendOnNewSignal=true (default false: keep the original exit).
NONE in the candidate. Pre-registered ablation (EnableShortLeg=true, IS only): SELL USDJPY at 23:00 UTC after r_t > +SigmaK*sigma_t, same exit. Screen: +6.9 pips gross vs 3.4 pips cost (spread 0.9 + commission 0.7 + short swap 1.8) = 2x, under the 3x bar, and -5 pips in 2021.
Time exit at 20:55 UTC on the day after entry (before the 21:00 UTC FX day end and, on Fridays, before the week end - the 21:00 UTC exit trap). One night held, swap earned: USDJPY long +10.064/lot/night (triple on Wednesday) - a positive cost line.
StopLossPips: hard stop; the measured effect has no stop, so the stop is a risk cap, not a signal - keep it wide (40/60/80).
No take-profit (the screen showed the rebound accrues through the whole next day: Asia +5 pips, London/NY +7 pips at 1.5 sigma).
Skip the signal if the USA500 60-day return history is incomplete (first 60 sessions of the test) or if today's r_t is not computable.
The signal symbol (USA500_CUSTOM) is read inside the tester via CopyClose on the other symbol; the EA must verify the bar exists and must never trade USA500.
TradeFridaySignals default changed to false AFTER the Phase 0.5 measurement (2026-09-12): Friday signals entered Monday 00:05 UTC measured -4 to -9 pips (n 14-40, all K) while Mon-Thu signals measured +12 to +17 pips. The parameter stays in the optimizer grid (both values run), so this is a default, not a selection.
Genetic optimization on the in-sample window with a net-of-costs robustness criterion; parameter-plateau selection (top candidates must have profitable neighborhoods, not just a best pass); out-of-sample window run once per frozen candidate on real ticks; verdicts gated on net profit factor, drawdown, and IS-to-OOS retention.
Key Takeaway
The rebound existed in the years it was measured on and ended with the regime the spec itself named as its killer, the 2022 yen collapse. Out of sample the strongest failed gate is profit factor (1.00 and 0.96 against 1.15); only 2024 was positive (+19.7%, the August carry-unwind rebound) and the 40-pip stop was hit on 101 of 172 trades once USD/JPY's daily range doubled, producing a 21% drawdown from a 4.3% in-sample one. The swap was earned as designed (+$583 out of sample) and was irrelevant. Real tick data, $3.50 per side commission, native swaps.
Disclaimer: Past performance is not indicative of future results. These backtest results are based on historical data with realistic commission assumptions. Real trading involves additional risks including execution delays, variable spreads, and emotional decision-making.