VCP Screener — port, study & verdict

Verdict

The one-line summary

Buying the pattern does not pay, and ranking the market with the pattern does not pay either — once you give the test enough history. The first window's +1.30pp excess was a favourable period, not an edge.

Two windows, same code — 50 dates vs 86 dates

Only two lines are stable across windows, and both are negatives: the pivot median (−2.22%) and the pivot win rate (41.3%). Everything that looked like an edge moved.

The methodological lesson — worth more than the result

On the 50-date window I ran a split-half test. The best rule held in both halves (+4.78pp early / +4.64pp late) and beat a size-matched control. It looked robust.

It was not. Those two halves were adjacent five-month spans of the same favourable regime. Adding 36 earlier dates — spanning a different market phase — took the same rule to −5.55pp early / +2.41pp late and the overall excess to −1.52pp.

A split-half test inside a single window cannot detect a window effect. It measures stability within a regime. The only protection is a window long enough to contain more than one.

Sample

stride 5 sessions · horizon 20 sessions · cache spans 2024-01-25 → 2026-10-15 · 0 live requests. Window 2 exists because the cache was extended backwards specifically to test whether window 1 was a regime artefact — it was.

Look-ahead controls

• Selection uses bars ≤ end only; the forward return starts at end. No future bar touches the decision.
• The base rate uses the identical entry, exit, horizon, universe and dates — so every number quoted as an excess is comparable.
• The study refuses to start before max(252, LOOKBACK+1) sessions, so a 12-month RS and a 252-day high/low always exist.
• An as-of date counts only when its full forward window is in the cache (last = n_sess − horizon).

Order-type semantics — the part that decided the study

A stop can never fill better than its trigger; a limit below the market can. Confusing them is what produced the fake first result (tab 6).

Design choices

• RS benchmark = cross-sectional percentile inside the scanned universe (~1,540 NSE names), not an index — an S&P-relative rating would put a third of NSE above 90.
• Extension penalty retained (up to −35, hard cut above +40% vs the 20DMA) — corroborated independently by discover.py and rule 18.
• The trend gate is a gate: a valid VCP below its 50/200 DMA is not a candidate. Tab 5 tests whether the pattern adds anything given that gate.

A · enter at the as-of close

Per-trade weights broad days heavily; per-date deploys equally every day. Both degrade on the wider window.

B · base rate — every eligible name, identical rule

The universe itself was roughly flat over both windows — so the VCP picks were not fighting a falling tape. They simply lost to it.

C · paired excess — picks minus the universe, per date

This is the single most important row in the report. The 70%-of-dates claim does not reproduce; on the wider window the picks are a coin flip against simply owning the universe equally.

D · pivot policy — the textbook Minervini entry (BUY-STOP)

The median and the win rate are identical to one decimal across a 36% larger sample. That is the most reproducible finding in the whole study, and it says the entry loses money.

E · dip policy — resting BUY LIMIT below the as-of close (rule 17/18 structure)

Unlike the day-1 breakout study (rule 18, where dip −5% was the best policy), deeper dips get no better here and the stop-out rate rises to ~40%. The fixed stop is 7.5% below the fill — see tab 6.

Composite score — inverted on both windows

Higher composite score ⇒ worse outcome, consistently. A warning about the 25/25/20/15/15 weighting, not a tradable inverse (n is tiny above 80).

RS percentile — the bucket that looked promising, then decayed

On 50 dates, RS 95+ was the only bucket with a positive median and ≥60% winners — which is what launched the control in tab 5. On 86 dates it decays but does not invert: still the best of the upper buckets, but +2.58% → +1.44% and 60.0% → 54.2% winners, while the lowest bucket (RS <70, n=21) is also positive. The bucket's magnitude is not stable, so do not size on it. (Note the two big buckets barely move: 85–95 stays ~−0.35% → −0.75%.)

Other buckets — window 1 (50 dates)

The one durable signal, and it holds on both windows: No VCP / Damaged at −2.47% / 27.6% win (50d) and −2.55% / 27.8% win (86d) — the screen is genuinely good at identifying what not to touch. Everything else in this tab moves between windows. "Textbook VCP" is n=1 in the rated bucket (n=5/7 in the pattern bucket): noise.

Breadth — a market timer in disguise, and it does not reproduce

What is stable: the pick count correlates negatively with the forward market return in both windows (−0.585 on 50 dates → −0.444 on 86) — few tight bases form just before a rally. That is a market-timing signal, not a selection one, and it is worth knowing: a crowded screening day is a poor day to buy.
What is not: the same count's correlation with the market-relative excess is weakly positive in both windows (+0.049 → +0.185) yet the bucket table flips sign between them (50 dates: thin days best; 86 dates: thin days worst at −2.05pp, broad days best at +1.41pp). A +0.05…+0.19 correlation and a sign-flipping bucket table are both consistent with noise. Do not build a rule on either.

Why this test exists

RS percentile is a 12-month momentum rank. So "the top VCP names by RS" is also "the strongest-momentum tight bases there are" — and ranking the whole universe by RS might pay just as well. If it does, the pattern is an extra step and we would be adopting momentum with more code.

Every sleeve is scored on the identical dates and forward window, and reported as its excess over the whole eligible universe on that date.

50d win rates: 52.0% / 72.0% / 66.0% / 71.4%.  86d win rates: 38.4% / 29.1% / 33.7% / 49.4%. Every sleeve loses to the universe on the wider window.

How to read it

• The ordering is stable; the sign is not. VCP ranks above the size-matched Stage-2 draw, which beats Stage-2, which beats raw RS — in both windows. So the trend gate plus the pattern genuinely sort better than momentum alone.
• But the whole momentum family underperformed the equal-weight universe over 2025-01 → 2026-09. Every sleeve is negative; the VCP sleeve is merely the least negative.
• "VCP adds +1.87pp over the size-matched control" is true and useless here — it is the difference between two losing sleeves (−1.52pp vs −3.39pp).
• Confound that cannot be resolved with this data: the study cannot separate "the VCP pattern has no edge" from "momentum had no edge over this 1.8 years on NSE". Both describe the same numbers. Separating them needs a momentum-matched benchmark, not just an equal-weight universe.
• The consistent ordering (VCP > Stage-2 > universe-ranked) is worth keeping as a hypothesis — it is the only thing here that reproduced across windows.

Where the earlier "+4.71pp, stable across halves" came from

It was real in that window. The 50-date window sat entirely inside the stronger phase of the sample; its two halves were adjacent spans of the same regime, so a split-half test had nothing to compare against. The moment the window reached back into Jan–Sep 2025, the same rule went to −5.55pp early / +2.41pp late.

This is the third time in this project that a window has decided the answer. The lesson is recorded in memory as a standing rule.

The two bugs that made the first study lie

The first run reported +2.83% mean, 50.7% win — an apparently excellent strategy. It was fiction, and both errors were order-type semantics.

331 of 641 signals were affected by bug 2 alone. Both are regression-tested (test_vcp.py, 70 checks green).

Earlier port errors, fixed and tested

Decision

• NOT ADOPTED  VCP as an entry signal — any policy, any stop. Do not add it to the strategist's entry pool and do not size on it.
• NOT ADOPTED  VCP + RS as a ranking sleeve. The 50-date result did not reproduce on 86 dates.
• USE AS-IS  the screener as a monitoring surface — --screen N, local cache only, no live requests, no orders. Its genuinely reliable output is the negative filter: No VCP / Damaged names are the worst cohort on both windows (−2.47% / 27.6% win on 50 dates, −2.55% / 27.8% on 86).
• OPEN  the ordering hypothesis (VCP > Stage-2 > raw RS) reproduced across both windows and is the one thing worth a further test — against a momentum-matched benchmark, not an equal-weight universe.

The durable lessons are not about VCP. First: an entry-policy study must state which order type it models — a stop cannot fill better than its trigger and a limit below the market can; mixing them manufactured a 2.8% edge from nothing. Second: a split-half test inside one window proves nothing about the next window. Extend the window before believing the result.