# MIROSCOPE VARIANT BACKTESTS — REPORT Date: 2026-09-25. Dataset: `backtest/data/eurusd-h1.csv` (10,605 H1 bars, 2025-01-01 to 2026-09-22 UTC). Engine: `variants/run_variants.py` (stdlib only, deterministic). Plan and mechanization choices: `variants/PLAN.md`. Accounting (all runs): $10,000 start, 1% risk per setup, R-multiple accounting, no spread/commission/slippage (same as the Phase 4 baseline, for comparability). ## TOP LINE (plain language) No variant beats the baseline convincingly. V3 (RSI divergence) has the best numbers (+0.635R expectancy, profit factor 5.13) but only 13 trades, which is too few to trust. V11 (moving-average origin) is marginally profitable on 23 trades (+0.068R, PF 1.12), also too few to trust. V4 is near breakeven. V2, V5, V9 lose. V1a, V1b, and V7 produce almost no trades at all: their filter stacks are so strict they barely trigger on 20 months of data. The M15 variants (6, 8, 10) were not tested and not faked: no M15 data exists here. ## RESULTS TABLE (ranked by expectancy, then profit factor) | Rank | Run | Variant | n | Win% | PF | Exp(R) | MaxDD% | Ret% | |------|-----|---------|---|------|----|--------|--------|------| | 1 | V3 | RSI-divergence 1-2-3 (13-17 UTC) | 13 | 84.6 | 5.13 | +0.635 | 1.0 | +8.25 | | 2 | V7 | Trader's Trick entry (07-17 UTC) | 3 | 66.7 | 1.55 | +0.183 | 1.0 | +0.55 | | 3 | V11 | MA-origin 1-2-3 (07-17 UTC) | 23 | 39.1 | 1.12 | +0.068 | 7.2 | +1.55 | | 4 | V4 | Neckline break + retest (07-17 UTC) | 24 | 45.8 | 0.94 | -0.029 | 4.2 | -0.69 | | 5 | V5 | Ross Hook continuation (07-17 UTC) | 16 | 37.5 | 0.89 | -0.068 | 4.9 | -1.09 | | 6 | V2 | 60-MA channel (07-17 UTC) | 24 | 50.0 | 0.79 | -0.088 | 4.6 | -2.10 | | — | — | **Phase 4 baseline** | 100 | 31.0 | 0.78 | -0.139 | 27.6 | -13.82 | | 7 | V9 | 50% pullback continuation (07-17 UTC) | 10 | 30.0 | 0.40 | -0.396 | 4.7 | -3.96 | | — | V1a | Classic 38.2-61.8% (07-11 UTC) | 1 | 100.0 | inf | +2.250 | 0.0 | +2.25 | | — | V1b | Classic 38.2-61.8% (13-17 UTC) | 1 | 0.0 | 0.00 | -1.000 | 1.0 | -1.00 | (V1a/V1b excluded from ranking: a single trade each is not a test.) ## PER-VARIANT NOTES - **V3** — 13 trades, 11 wins. Winners cluster in EURUSD's 2025 rally; the 18-bar live-trade expiry closing drift-profitable trades does heavy lifting (7 of 13 closed at expiry with profit). Mechanics verified (no lookahead; pivots confirmed 2 bars late; stops fill on later bars; accounting unit-tested). Verdict: interesting, NOT convincing. Would need out-of-sample data (more pairs, more years) before anyone acts on it. - **V7** — 3 trades. The correcting-bar-within-3-bars + 1.0R-to-P2 + slope + RSI + ATR stack almost never fires. Verdict: untestable as specified. - **V11** — 23 trades, marginally positive. Two full-target winners (+2.25R each: 0.5*1.5 + 0.25*2.5 + 0.25*3.5, math verified) carry the result; 12 of 23 took the full -1R stop. Verdict: marginally positive, not convincing at n=23. - **V4** — 24 trades, near breakeven (PF 0.94). The retest entry does not fix the baseline's problem. Verdict: no edge found. - **V5** — 16 trades, negative. The Ross Hook continuation as mechanized (first post-breakout pivot high, hook <= 5 bars, <= 50% retracement) loses mildly. Verdict: no edge found. - **V2** — 24 trades, 50% win rate but PF 0.79: wins too small. Better expectancy (-0.088R) and far smaller drawdown (4.6%) than baseline, but still negative. Verdict: not a convincing beat. - **V9** — 10 trades, PF 0.40, clearly negative. Buying the rebound after a 50% retracement, as mechanized, underperforms badly. Verdict: dead. - **V1a/V1b** — 1 trade each. The RSI-at-P1 filter (<=35 long / >=65 short) is the binding constraint: across 20 months it passes 1-15 times depending on swing size. Verdict: untestable as specified; the filter set describes a setup that hardly ever occurs on H1 EURUSD. ## ASSUMPTIONS (mechanization choices — see PLAN.md for the full list) - Fractal swings: N=2 bars each side; a pivot is only visible 2 bars after it forms (no lookahead). 1-2-3 = latest confirmed alternating triple (L,H,L) / (H,L,H). - "1 tick" stop entries implemented as trigger + 0.05*ATR buffer; fills at the stop price; if the fill bar also hits SL, the SL is assumed hit first (conservative). - Close-break entries (V2, V9) fill at the signal bar's close; SL/TP from the next bar. - Partials: 50/50 at two targets; 50/25/25 at three targets (brief's rule). - Session windows gate the entry-signal bar only; pendings may fill later and trade management continues outside the window. - V1 expiry 8 bars after P3; V2 6; V3 8; V4 3 after breakout; V5 5 after hook low; V7 8 after correcting bar (documented anti-stall); V9 10 after impulse; V11 6. Live-trade expiries per spec, closed at bar close. - V7: spec lists both 0.50R and 1.0R minimum entry-to-P2 distance; used the stricter 1.0R. V4 pre-breakout triple staleness 12 bars and V5 30 bars are documented anti-stall choices (spec is silent). Spread filters not modeled (no spread data; same as baseline). - One position per direction max. Max drawdown on closed-trade balance equity. ## UNTESTED (M15 — do not fake) - V6 (Ross Hook 25% retracement, M15), V8 (1-2-3 + Ross Hook congestion tie-breaker, M15), V10 (ATR-adaptive Ross Hook, M15): NOT tested. No M15 data exists in this environment. Nothing below was invented for them. ## COMPARABILITY CAVEAT The baseline ran a D1-bias/H1-trigger lane with a 5-setups/day cap and a -3% daily halt; the variants are pure-H1 with session windows and no daily cap/halt (per the brief's common simulator). Expectancy, profit factor, and return are comparable (same data, same $10k/1% accounting); trade counts and drawdowns are not apples-to-apples with the baseline's risk overlays. ## BOTTOM LINE FOR MICHAEL The research produced hypotheses, the backtest disposed of most of them. Two are worth keeping an eye on (V3's divergence filter, V11's MA-origin filter), but neither has enough trades to believe yet. Nothing here justifies paper trading. The honest next step, if he wants it: more data (more pairs, more years) to see whether V3/V11 survive out of sample. That is his call.