← Academy home
Proving Ground · Round 4 of 4

Tuning the leaders — and letting the exam judge.

The four Round-3 leaders went through plateau-seeking parameter search: tuned on the first 70% of history, then graded on the sealed final 30% they never saw. Defaults sat the same exam. Whoever scores better OOS wins — no exceptions, no excuses.

The headline — and the lesson

In-sample, tuning "improved" everything. Out-of-sample, it won only twice. RevertX's and BreakoutGuard's defaults beat or matched their tuned versions on the exam — Lesson 9's tailored suit, caught in the act by walk-forward discipline. The two real wins: RSI on GBPUSD (gentler 35/65 zones: −7.7% → +2.3% OOS) and a small consistent Sentinel tweak (lower dead zone, 1:3 targets).

This is the system working, not failing: every rejected parameter set here is a live-money loss that never happened.

RSI

EURUSD · H4 · KEEP DEFAULTS
PF in-samplePF sealed examreturn OOSmax DD OOStrades OOS
tuned1.000.88-1.3%2.5%82
defaults1.071.03+0.9%3.9%247

tuned params: rsi_period=20, zone=[25, 75], atr_mult=1.5, rr=2.0
Tuning looked fine in-sample and failed the exam (0.88 vs default 1.03). Textbook mild overfit — defaults stay.

GBPUSD · H4 · ADOPT TUNED
PF in-samplePF sealed examreturn OOSmax DD OOStrades OOS
tuned1.161.06+2.3%3.9%239
defaults1.100.80-7.7%11.6%226

tuned params: rsi_period=20, zone=[35, 65], atr_mult=1.5, rr=2.0
Gentler zones (35/65) + period 20 flipped GBPUSD from −7.7% to +2.3% OOS with a third of the drawdown. A real, exam-passing improvement.

RevertX

GBPUSD · H1 · KEEP DEFAULTS
PF in-samplePF sealed examreturn OOSmax DD OOStrades OOS
tuned1.821.48+0.8%0.5%37
defaults1.442.34+0.6%0.2%14

tuned params: adx_range_threshold=30.0, rsi_oversold=30.0, atr_mult=1.5, rr=3.0
Defaults scored OOS PF 2.34 — the tuned version was worse. RevertX's four locks are already well-placed. Caveat: only 14 OOS trades; the wobble band is wide.

EURUSD · H1 · KEEP DEFAULTS
PF in-samplePF sealed examreturn OOSmax DD OOStrades OOS
tuned1.781.19+0.1%0.2%12
defaults0.752.13+0.3%0.2%8

tuned params: adx_range_threshold=25.0, rsi_oversold=25.0, atr_mult=1.0, rr=2.0
Same story, tiny sample (8 OOS trades). Defaults win; judge again after forward testing accumulates trades.

Sentinel

USDJPY · H1 · ADOPT TWEAK
PF in-samplePF sealed examreturn OOSmax DD OOStrades OOS
tuned1.211.06+2.2%4.3%582
defaults1.101.04+1.5%4.2%592

tuned params: wae_dead_zone_pip=20.0, atr_mult=1.5, rr=3.0
Lower Waddah dead zone (20) + wider 1:3 targets nudged OOS PF 1.04→1.06 and return +1.5%→+2.2% across 582 trades — small but consistent, and it agreed across both windows.

USDJPY · H4 · H1 ONLY
PF in-samplePF sealed examreturn OOSmax DD OOStrades OOS
tuned1.290.91-3.3%8.1%237
defaults1.190.90-3.4%6.3%248

tuned params: wae_dead_zone_pip=20.0, atr_mult=2.0, rr=3.0
Both tuned and default lose OOS on H4 (~0.90). Sentinel's USDJPY edge lives on H1 in recent conditions — restrict it there.

BreakoutGuard

USDJPY · H4 · KEEP DEFAULTS
PF in-samplePF sealed examreturn OOSmax DD OOStrades OOS
tuned2.251.98+1.2%1.0%14
defaults1.292.16+1.2%0.8%11

tuned params: squeeze_threshold=0.003, volume_ma_mult=1.2, atr_mult=2.0, rr=3.0
Defaults OOS PF 2.16; tuning matched but didn't beat. With 11–14 trades per window, changing anything on this evidence would be needle-chasing.

GBPUSD · H4 · KEEP DEFAULTS
PF in-samplePF sealed examreturn OOSmax DD OOStrades OOS
tuned1.072.77+1.1%0.5%10
defaults1.072.77+1.1%0.5%10

tuned params: squeeze_threshold=0.0015, volume_ma_mult=1.0, atr_mult=1.5, rr=2.0
Tuner literally converged to near-defaults — identical OOS result. The strategy is what it is: rare, high-quality trades.

The roster going forward — what graduates to forward testing

RevertX — defaults, all majors H1 (few, high-quality fades; needs forward trades to firm up the small sample)

Sentinel — USDJPY H1 only, with the tweak (dead zone 20, 1:3 targets)

BreakoutGuard — defaults, H4 majors (rare trades, strong quality)

RSI — GBPUSD H4 with tuned zones (35/65, period 20); other pairs benched

Benched pending redesign: MA Crossover (blanket use), AdaptiveEdge, Combined, Grid, both scalpers.