EURUSD 1H London-Session 2-Sigma Band Fade: short above / long below a 20-bar mean +/- 2 std, exit at the mean, after 12 bars or at a 1.5x ATR(20) intrabar stop (FX venue, long+short, 1.5x cap)

Hypothesis

Implementation

EURUSD.FX 1-hour fade of 20-bar 2-sigma bands during London hours. On a closed bar inside 07:00-16:00 Europe/London (the bar opens at 07:00 or later and closes at 16:00 or earlier; DST-aware via zoneinfo), a close above mean+2*pop-std opens a short and a close below mean-2*pop-std opens a long. Exits come from whichever happens first: a closed bar back at or through the current 20-bar mean, 12 bars after the entry fill (market orders, filled at the next bar's open on the FX session venue), or a reduce-only intrabar stop placed on the entry fill at fill -/+ 1.5*Wilder ATR(20) (submit_stop_order, crossed_policy=market). Sizing: notional = min(0.5% * equity / (1.5*ATR/close), 1.5 * equity), in whole EUR units, with a 1,000-unit minimum.

Faithful implementation of the settled rule. calculate_signal returns the continuous z-score (close-mean)/sd, and should_enter thresholds it at +/-band_k, which is the same as comparing the close with the bands. Session gating applies to entries only. A protective stop is placed from on_order_filled using the ATR of the signal bar; the base cancels it on a mean or time exit, and on_position_closed records the exit time so no entry happens on the bar an exit occurred. Risk route (a): with the 1.5x leverage cap binding (0.5%/stop% is usually above 1.5x), the loss at the stop is about 1.5 x 0.2-0.3% = 0.3-0.45% of equity, under 2%. All state is held in scalar attributes. The ATR is incremental, and the mean and std use only the last lookback bars (bounded). Execution substitutions follow the registry: next_open_fill for market orders and a platform-matched intrabar stop instead of an exchange-resting order. Risk noted: the expected average trade is close to the 0.10% FX analyst floor.

Verification Results

No code change needed. Note that the first-bar stop coverage depends on the platform's session-stop switch.

The protective stop is placed in on_order_filled while the entry fills at the next bar's open. Per src/backtesting/next_open.py, without SF_SESSION_STOP_FILLS an order sent at the open is released only after bar N+1's range has been matched, so in the entry bar itself the stop is tested only against that bar's close. This is platform behaviour, not a strategy defect. The stop protects from the second bar onward either way.

Optional: none required.

Wilder ATR updates only start once the base begins calling calculate_signal (after min_bars_required = 21 bars). The first true range has no previous close, so it falls back to high-low. The ATR seeds about 20 bars later than strictly necessary. This is harmless: the evaluation warmup is 60 bars and entries are gated on atr_value > 0.

Expected gross capture (~0.12-0.2% per trade by the hypothesis's own arithmetic) is close to the 0.10% FX analyst floor. The mean target is the moving 20-bar mean, which drifts toward price, so per-trade capture shrinks. About 1,000 sandbox trades in 735 days means roughly 1.4 trades per day. The analyst should check whether avg_trade_return_pct clears the floor after carry on trades held across the 17:00 New York roll.

Backtest Review

Sharpe

-0.66

Total return

-6.78%

Max drawdown

11.44%

Trades

500

Win rate

40.2%

Profit factor

0.90

The code does what the user's rule says: London-session gate (DST-aware), 20-bar population-std bands, entries against the move, exit at the mean, a 12-bar clock, a reduce-only 1.5x ATR stop, and risk-sized notional capped at 1.5x. Trades are balanced (201 long, 194 short) and entered on band breaches, as the hypothesis predicts.

The sample is reliable: 395 closed trades over the development period (to 2026-06-05), clean input audit, metrics_reliable true.

Risk control works: largest loss is $572 (about 0.57% of equity) and max drawdown is 10.2%, inside the pre-registered 15% limit.

Net expectancy is negative. Average trade return is -0.008% of notional against the 0.10% FX floor, profit factor 0.91, Sharpe -0.65, total return -5.7%, and 2025 and 2026 are both losing years.

The edge is not positive even before costs. Price PnL summed over the ledger is about -$2.9k, a mean gross of about -0.004% per trade. Commission ($4.0k) and overnight carry ($2.4k) add losses on top. Fees are not what kills this strategy: the fade itself has no reversion edge.

The QA concern is confirmed. The hypothesis expected about 0.12-0.2% gross capture per trade; the measured capture is around zero. Short fades do worst, with a 35% win rate, which fits band walks continuing rather than reverting.

Outcome Summary

Hourly EURUSD 2-sigma band breaches in the London session did not revert enough to give any gross edge, and short fades were worst (36% win rate), so a fade needs a regime or trend filter before costs are even relevant.

At backtest review the analyst voted abandon before optimization: the fade lost money even before costs (gross price PnL about -$2.9k, or roughly -0.004% per trade), so fees were not the cause. The record nonetheless shows a final stage of live_stage, with no optimization, analyst, risk-officer or abandon-reason record; the input does not explain this mismatch.

A user-submitted EURUSD 1-hour mean-reversion fade on the FX venue: during the London session it shorted closes above a 20-bar mean + 2 std and bought closes below mean - 2 std, exiting at the mean, after 12 bars, or at a 1.5x ATR(20) stop, with risk-sized notional capped at 1.5x leverage.

Over the full backtest (2024-10-30 to 2026-10-30) it lost 6.78% with a Sharpe of -0.66 across 500 closed trades: 40.2% win rate, profit factor 0.90, average trade -0.0088% of notional (against a 0.10% FX floor), and 11.4% max drawdown. Commission was $5,033 and overnight carry about $2,960; the signal diagnostic found no information (IC 0.004, t=0.1).