trading / confluences / b_dryup

b_dryup — volume dry-up (the anti-c3 hypothesis) UNDERPOWERED

Corpus lane — EN-37 (inversion) · One arm too rare to refute anything.

What it asks

The opposite of a volume spike: does the breakout work BETTER when the decision bar is quiet — 0.70× or less of the 20-bar average?

Exact definition · verbatim docstring from backtests/conf_lab/bits.py (sha1 5473adbbf694)

EN-37 volume dry-up (the ANTI-c3 hypothesis). 1 iff the decision 5m
bar's volume is <= 0.70 x the mean volume of the 20 preceding 5m bars on
the CONTINUOUS series (the same series and lookback c3 / `b_volhard` use).

Fewer than 21 closed 5m bars, or a non-finite / non-positive mean -> 0.

The 0.70 multiple is OURS, not sourced: the backlog card gives no number.
Declared as such in SCREEN-2026-08-07-DESIGN.md §A.1 and in
`SCREEN_PROVENANCE`; it was fixed there before any value was computed.

Fixed parameters & provenance

Every parameter below was fixed in the design document before any candidate value was computed. No sweep, no tuning: a bit that needs tuning to pass has failed.

Screen analytics — screen-2026-08-07

Conditional association on mean solo R, stratified on the count of the OTHER live bits, against a per-bit prevalence-matched null (50,000 i.i.d. Bernoulli draws at this bit's own prevalence, scored on its own strata). Population: 2,882 candidates, 2,289 filled, MES, 2019-06-04 → 2026-07-31.

Per era · reported separately, never pooled into a verdict — each cell is a smaller sample than the headline

Funnel & fill conditioning · a bit that mostly predicts whether a limit FILLS is not the same discovery as a bit that predicts whether a trade WINS

Novelty vs the six live bits · high agreement means the bit is mostly restating a gate that is already in the stack

By stratum — does it work where the method actually trades? · the live A/B take rule is count >= 4, so the count>=4 rows are the only ones that could change a live trade

Status — UNDERPOWERED

UNDERPOWERED, and that word was chosen before the number was seen. The bit fires on 2.5% of filled rows (61 of 2,289), so one arm holds well under 5% of the data and the screen declared in advance that a null here means 'we could not tell', NOT 'refuted'. The measured pooled delta is +0.103 R with matched p=0.562 and a bootstrap CI spanning [−0.207, +0.440] — a range wide enough to contain almost any hypothesis. Its 0.70 multiple is also OURS, not sourced: the backlog card gives no number, and the design doc says so out loud (§A.1) while §A.2 records why the same latitude was refused to EN-12.

Retest path

It needs POWER, not a rerun: many more rows in the rare arm, which means a longer forward window or a wider population, and a pre-declared minimum detectable effect. Re-running it on these 2,289 rows would produce the same shrug and spend another multiplicity slot.

Standing caveats

• Fill model optimistic: touch fills, zero net slippage, entry-bar SL/TP not checked inside the fill minute.
• Not a holdout — 2019-2026 was selected on, tuned on and reported on before this lab existed. A good number here means 'worth pre-registering', nothing more.
• MES only; the P1 edge is MES-centric and recent-regime.
• Masters miss four Sunday-evening blocks (May-Jun 2026); data ends 2026-07-31.
• Mean solo R is an average per candidate. The solo ledger is not a book and is never summed (PROTOCOL.md §0).
• Per-era cells are reported separately and are NOT pooled into a verdict; each is a smaller sample than the headline.
• Nothing on this page is TV-verified. A backtest number becomes actionable only after an independent TradingView Pine port reproduces entries AND per-trade PnL on a Deep Backtest export.
• Screening is not certification. Promotion requires all seven PROTOCOL gates, ending in a forward paired evaluation and a Pine referee.

← all confluences