BreakoutGuard — The Parameter Dossier
The brochure promises five gates: a Bollinger squeeze, ATR compression, a breakout candle, a volume crowd, and MACD's blessing. The code tells a different story — two of those gates don't actually exist. And in a twist opposite to RevertX's, the exit we simulate by accident turns out to be better than the exit that was designed.
B.1The defendant
Design (docstring) consolidation (BB squeeze AND ATR compression) → breakout candle → volume confirms → MACD confirms → enter with the crowd
Roster config all 7 majors · H4 · pure defaults · ATR×1.5 stop, 1:2 target · magic 7703
Round-4 record sealed exam: PF 2.16 · USDJPY H4 · 11 trades
B.2The gates that aren't — code vs brochure
The census started with a paradox: on H4, the BB-squeeze condition (band width under 0.2% of price) is true on 0.0–0.2% of bars — for most pairs, never in five years — yet trades kept firing. The code explains it. Two discoveries:
| gate (as advertised) | what the code actually does |
|---|---|
| BB squeeze AND ATR compression | squeeze OR compression — and since the squeeze ~never triggers on H4, ATR compression (true ~6–11% of bars) does all the gatekeeping alone |
| volume must confirm | volume is not a filter at all — a volume spike only nudges the confidence number (×1.1 vs ×0.9); it never blocks a trade. We proved it: setting the volume threshold to 1.0 changes nothing, trade for trade |
So the real strategy is three conditions: ATR compressed · close outside the band · MACD agrees. This is the third code-vs-story mismatch in three dossiers (RSI's formula, RevertX's exit, now this) — and the recurring moral of the whole series: the brochure is not the bot. Read the code, then test the code.
B.3The eleven-trade throne
Same test as RevertX's: 200,000 no-edge coin-flippers, 11 trades each on a 1:2 bracket. 12.1% of them beat PF 2.16 — one in eight. RevertX's 14-trade throne could be faked by luck one time in seventeen; BreakoutGuard's is nearly twice as fakeable. Its baseline is the weakest evidence on the roster.
B.4The record, pair by pair
| pair | full history | sealed OOS | n (full) |
|---|---|---|---|
| USDJPY | PF 1.54 · +$181 | PF 2.07 | 46 |
| GBPUSD | PF 1.54 · +$121 | PF 2.78 | 23 |
| USDCHF | PF 1.68 · +$68 | PF 0.82 | 22 |
| NZDUSD | PF 0.82 · −$37 | PF 4.68 (n=6!) | 25 |
| EURUSD | PF 0.35 · −$165 | PF 0.16 | 30 |
| AUDUSD | PF 0.42 · −$166 | PF 0.48 | 26 |
| USDCAD | PF 0.57 · −$95 | PF 0.02 | 18 |
The same shape as every other dossier: an edge that lives in one or two places, surrounded by pairs that quietly bleed. USDJPY and GBPUSD agree in both halves; USDCHF's full-history shine fails its exam; EURUSD, AUDUSD and USDCAD lose everywhere and have always lost — they are on the roster entry only because the entry was written generously. Recommendation: narrow the roster pairs to USDJPY + GBPUSD (El's decision, before the campaign).
B.5The exit that accidentally works
Like RevertX, BreakoutGuard's entry is a one-bar event, so the engine's signal-flip exit closes nearly every trade after a single H4 bar (median hold: 4 hours; ~90% "signal" exits). And like RevertX, the design specifies its own exit — an ATR trailing stop, or "price closes back inside the band". So we simulated the designed exit too. The twist:
| pair | accidental 1-bar hold (as simulated) | designed trailing exit |
|---|---|---|
| USDJPY | PF 1.54 · +$181 | PF 1.22 · +$80 |
| GBPUSD | PF 1.54 · +$121 | PF 1.29 · +$70 |
| USDCHF | PF 1.68 · +$68 | PF 1.34 · +$40 |
| losers (EUR/AUD/CAD/NZD) | still lose | still lose |
The designed exit is worse on every winning pair. On H4, the breakout's profit is concentrated in the burst bar itself; waiting for price to fall back inside the band means donating the burst back. RevertX's cage was strangling it — BreakoutGuard's cage was protecting it. Same investigation, opposite verdicts, one shared lesson: an exit story is a hypothesis like any other. Test it; never assume it helps. The roster bracket stays.
B.6The tuning court
Could looser gates find more treasure? Pooled across all seven pairs, tuning half → exam half:
| variant | IS (pooled) | OOS (pooled) |
|---|---|---|
| defaults | PF 0.86 (n=129) | PF 1.17 (n=61) |
| squeeze ×2 looser | PF 0.85 | PF 1.04 |
| squeeze ×4 looser | PF 0.96 (n=947) | PF 0.98 |
| compression 0.9 | PF 0.97 | PF 1.06 |
| everything loose | PF 0.96 (n=1241) | PF 0.98 |
Every loosening buys trades and pays for them with edge — at the fully-loose corner the strategy degrades into "trade every band poke", a proven loser. The tight gates aren't hiding a bigger harvest; they are the only thing separating BreakoutGuard from noise. Defaults keep the crown, for the third dossier running.
B.7Dossier verdict
Parameters: unchanged — every tested loosening loses its edge out of sample. Exit: keep the roster bracket — the designed trailing exit measurably underperforms the accidental one-bar hold. Roster entry: narrow to USDJPY + GBPUSD (recommendation; EUR/AUD/CAD lose in every period examined). Docket: the squeeze-OR-compression and volume-doesn't-filter mismatches deserve either a code fix or a docstring fix — decide which story is the real strategy, then make the code and the words agree. Standing caveat: an 11-trade baseline that luck fakes one time in eight — the forward campaign is where BreakoutGuard's real trial begins.
Next: Dossier 4 consolidates Sentinel's parameter evidence (already earned across its four-part deep dive) — and then the docket goes to El for decisions.