NightMeanRevert EA

The story in one line

A clean-room, 985-line MQL5 night-session mean-reversion EA — Bollinger-band rejection in low-vol, low-trend regimes, with genuinely well-built risk controls — loses money on real 2018–2026 Dukascopy M5 data: net −$181 (−3.62%), profit factor 0.655, Sharpe −0.48, win rate 29.6% across 71 trades. It is net-negative in every out-of-sample segment, and every single parameter perturbation tested is break-even at best. This is not a cost problem or a tuning problem — the mean-reversion edge is absent in the data.

This is the opposite failure mode of the basket-grid bots that over-traded into ruin: NightMeanRevert is so selective it produces only ~8 trades/year across 4 pairs, and the few it takes lose. The 4th archetype to fail — and the most disciplined of them.

Headline result

Metric	Value
Profit factor	0.655
Win rate	29.6%
Net P&L	−$181 (−3.62%)
Trades (8.5y, 4 pairs)	71 (8.4/yr)
Reverts to target	21% (15/71)
Stops out	66% (47/71)

The realised reward:risk is a healthy 1.56, but a 29.6% win rate needs to be ≥39.1% to break even. It isn’t. Price reaches the BB-center target only 21% of the time and hits the 1.40-ATR stop 66% of the time — the mean-reversion thesis does not hold within the 3-hour night window.

Per-segment performance

Segment	n	PF	Win%	Net $	Positive?
Dev 2018–2021	33	0.98	36.4	−4.01	❌
Validation 2022–2023	13	0.42	23.1	−60.61	❌
Locked OOS 2024–Jun 2026	25	0.47	24.0	−116.50	❌

Every segment is negative. The dev period is the “least bad” (a statistical zero on n=33), and performance degrades out-of-sample — the opposite of a robust edge. The README required ≥200 OOS trades; the locked OOS produced 25, an 8× miss, leaving the strategy statistically un-validatable.

No profitable island

A one-at-a-time parameter sweep finds no profitable neighbourhood. The best result anywhere in the grid is short_rsi_min=75 at PF 1.001 (dead flat, no net trades) and bands_period=15 at PF 0.984 — the least-bad perturbations are break-even at best, and they straddle a negative default rather than sitting on a robust plateau. Every other cell lands at PF 0.25–0.72. Loosening filters to get more trades (long_rsi_max=45 → n=181) makes it worse (−7.49%). DST/server-clock offsets (−1h / default / +1h) all lose (−$407 / −$181 / −$151). The “tunings” that nudge PF toward 1.0 do so by removing losing trades until nothing is left — converging to no-trade, not to a profitable system. This is the signature of no edge, not overfitting.

Verdict: REJECT (PF 0.655). Do not deploy. NightMeanRevert is the most disciplined of the failed retail EAs — the clean-room engineering is sound and the risk controls are genuinely well-built — but the signal those controls gate has no demonstrable edge on real data. A backtest that loses before any live frictions (real night spreads, swap, tick-fill slippage) are added will only lose more with them.