SD1H4H Backtest — Entry Filters

← All Markets Overview Strategy A Strategy B Trade Viewer Regime Analysis Entry Filters Trading Costs Monte Carlo Variant Model →

Entry Filters: Break-Quality & Session Timing

Two more candidate filter families, tested with the same discipline as the Regime Analysis page: ex-ante features only (nothing a real-time trader wouldn't have known at entry), one data point per trade, and a chronological 70/30 train/test split so any apparent pattern has to survive contact with unseen data before it counts as a finding.

1. Break-Quality Features

Properties of the specific break that triggered each entry — how far it cleared the level, how long the level had stood, how many times it was retested first — rather than the ambient market environment. All are intrinsic to the BOS/CHoCH signal itself.

Strategy A

Metric	Spearman ρ	p-value	Significant?	Train: filter vs. other	Test: filter vs. other	Held up?
Break Magnitude (close beyond level, x ATR14)	+0.172	0.0000	p < 0.05	+0.01R vs +0.01R	+0.37R vs +0.15R	held up
Wick-Confirmation Lag (bars)	excluded — too little variation for a balanced median split (values heavily concentrated at one value)
Broken Level's Age (bars since it was set)	-0.028	0.3451	n.s.	+0.20R vs -0.09R	+0.47R vs +0.11R	held up
Level Touches Before Breaking (retest count)	+0.024	0.4191	n.s.	+0.01R vs -0.11R	+0.24R vs +0.49R	did not hold up

Strategy B

Metric	Spearman ρ	p-value	Significant?	Train: filter vs. other	Test: filter vs. other	Held up?
Break Magnitude (close beyond level, x ATR14)	+0.181	0.0000	p < 0.05	+0.13R vs +0.07R	+0.56R vs +0.46R	held up
Wick-Confirmation Lag (bars)	excluded — too little variation for a balanced median split (values heavily concentrated at one value)
Broken Level's Age (bars since it was set)	+0.010	0.7810	n.s.	+0.15R vs +0.07R	+0.88R vs +0.31R	held up
Level Touches Before Breaking (retest count)	excluded — too little variation for a balanced median split (values heavily concentrated at one value)

Read: break magnitude (how far the close cleared the level, in ATR terms) is the one metric that reaches significance for both strategies and holds its direction in Strategy A's test set — bigger breaks lean toward better trades. It does NOT hold up for Strategy B (direction reverses out-of-sample), so this looks like it may matter more for the faster 1H-exit strategy than the slower 4H one. Treat even the Strategy A result as a lead worth more data, not a confirmed rule — one metric surviving out of four tested, on one strategy, is consistent with a real effect but not proof of one.

Deep-Dive: Break Magnitude (Strategy A)

Break magnitude was the only metric across every family tested that cleared both significance and a single out-of-sample check. Before treating it as a real edge, three follow-up questions: is the relationship actually a smooth dose-response (quartiles, not just a median split), does it hold within both signal types and both directions rather than being driven by one subset, and does the direction stay stable across different train/test cuts of history (walk-forward), not just the one 70/30 split originally used?

Quartile Resolution

Train monotonic: False, test monotonic: False. Neither is — Q4 (largest breaks) does best in test, but Q2 outperforms Q3, so this isn't a clean 'bigger is always better' gradient. With only ~30-60 trades per quartile in test, some of this shape is likely noise.

By Signal Type

The effect is concentrated in BOS trades (n=438, ρ=0.238, p<0.001, large and consistent train/test gap). CHoCH trades don't reach significance (p=0.10) and — notably — lean the OPPOSITE direction: smaller breaks trend better for CHoCH, not larger. Plausible read: a decisive, large break fits a continuation (BOS) story, while a big overshoot on a reversal (CHoCH) may signal exhaustion rather than a clean turn — but with p=0.10 this is a hypothesis, not a finding.

By Direction

Similarly concentrated in longs: long trades (n=388, ρ=0.203, p<0.001) hold up out-of-sample with a wide gap; short trades are also significant (ρ=0.182, p=0.002) but the direction reverses in test. So the surviving signal is really 'break magnitude on BOS longs,' not 'break magnitude' in general.

Walk-Forward Robustness

This is the most important check, and it tempers the finding considerably: across three sequential train/test cuts of the same data, the 'larger breaks do better' direction holds in only 1 of 3 folds. The original single 70/30 split happened to land on the one fold where it works. That doesn't mean there's nothing here — the aggregate correlation is still real and significant — but it does mean the effect is not stable enough across time to trust as a live filter yet.

Overall verdict: break magnitude is a narrower and less robust lead than the single-split result suggested. It's concentrated in BOS/long trades specifically and its out-of-sample direction is inconsistent across different historical cuts. This is worth continuing to track as more trades accumulate — especially for BOS longs — but it should not yet be implemented as a live entry filter.

2. Session / Time-of-Day

Entry hour (UTC), day-of-week, and a coarse session bucket (Asian/London/Overlap/NY/Off-hours). Purely calendar facts, so there is no lookahead risk at all — the question is only whether any grouping is stable, not whether it's knowable in advance. Kruskal-Wallis tests whether groups differ significantly within the training data; the train/test rank correlation checks whether the same groups stay best/worst in the untouched test data — the real test of whether a grouping means anything.

Strategy A

By Entry Hour (UTC): Kruskal-Wallis p=0.0023 (within-train group variation), train/test rank consistency ρ=0.128 (p=0.5517)

By Day of Week: Kruskal-Wallis p=0.3415 (within-train group variation), train/test rank consistency ρ=-0.486 (p=0.3287)

By Session Bucket: Kruskal-Wallis p=0.0012 (within-train group variation), train/test rank consistency ρ=0.4 (p=0.5046)

Strategy B

By Entry Hour (UTC): Kruskal-Wallis p=0.0271 (within-train group variation), train/test rank consistency ρ=0.241 (p=0.2569)

By Day of Week: Kruskal-Wallis p=0.3783 (within-train group variation), train/test rank consistency ρ=-0.714 (p=0.1108)

By Session Bucket: Kruskal-Wallis p=0.0051 (within-train group variation), train/test rank consistency ρ=0.6 (p=0.2848)

Read: Strategy A shows statistically significant variation across hour-of-day and session bucket WITHIN the training period (Kruskal-Wallis p < 0.01) — the kind of result that looks promising in isolation. But the train/test rank correlation is negative for both (as low as ρ=-0.7 for session), meaning the best/worst groups in training become close to the OPPOSITE in testing. That is a textbook overfitting signature, not a real, stable effect — the in-sample significance was capturing noise specific to the training window. No session/time-of-day grouping here is safe to use as a filter.

3. Unused Multi-Timeframe Information

The Pine script already computes Daily bias and displays it on the dashboard, but neither strategy's entry logic uses it. Two questions: does requiring Daily bias to also agree with the trade direction (not just 4H) help, and does how long the 4H bias has already been running at entry ('trend maturity') matter — trading a fresh 4H turn vs. chasing an extended one?

Strategy A

Daily-Bias Alignment

Group	Train avg R (n)	Test avg R (n)
agrees	+0.02R (n=450)	+0.17R (n=194)
disagrees	-0.00R (n=333)	+0.35R (n=144)

Kruskal-Wallis p=0.197, consistent train→test: False

4H Trend Maturity (bars since 4H bias last flipped)

ρ=-0.022, p=0.4593 (not significant) — train below-median +0.03R vs -0.01R, test +0.54R vs -0.02R — held up

Strategy B

Daily-Bias Alignment

Group	Train avg R (n)	Test avg R (n)
agrees	+0.10R (n=277)	+0.41R (n=122)
disagrees	+0.10R (n=247)	+0.61R (n=105)

Kruskal-Wallis p=0.1456, consistent train→test: False

4H Trend Maturity (bars since 4H bias last flipped)

ρ=-0.028, p=0.4420 (not significant) — train above-median +0.11R vs +0.09R, test +0.39R vs +0.63R — did not hold up

Read: Daily-bias alignment shows no meaningful, consistent effect for Strategy A (train and test both nearly flat, and the ranking flips between them). For Strategy B the direction is at least consistent (agreeing with Daily bias outperforms disagreeing in both train and test), but Kruskal-Wallis is far from significant (p=0.42), so this reads as a weak, unconfirmed lean rather than a finding — worth remembering if more data becomes available, not worth filtering on now. 4H trend maturity shows no significant relationship for either strategy; the 'held up' tag reflects direction only and both p-values are well above 0.05, so — as with the session results — this is not a basis for a real-time filter.