Fade family validated · continuation null Scalping research program Bybit perps (383 liquid) 60d · 5m bars · UTC

Scalping Research —
Evidence-Selected Setups

The research program behind the tradeable scalps: a hypothesis registry (clusters A–L), an event-study harness over 383 liquid Bybit perps, and an evidence-tiered 32-strategy brief. This is where the fades were proven and the popular rules were falsified.

Fades = the edge · 80% Edge Rule and ITSM falsified on this universe
+0.649 ATR
E8 VWAP-fade 1h forward
2.67M events · strongest signal in the program
+0.411 ATR
E4 IVA break-in fade
Failed-expansion reversion · 51k events
NULL
80% Edge Rule
Traverse only 26% vs claimed ~80%
0.838–0.852
L5 crash-warning AUC
Risk gate, all 3 OOS modes

Verdict

The program's core conclusion: the reversion family is the consistent winner on crypto perps — VWAP-deviation fade (+0.649 ATR), IVA break-in fade (+0.411), CVD-divergence fade (+0.117), VA-rejection fade (+0.040) — while the breakout/continuation family is weak-to-moderate and the two most popular practitioner rules fail on this universe: the 80% Edge Rule (26% traverse vs ~80% claimed) and ITSM (first-30m/last-30m correlation −0.017).

J1 sweep-reclaim was validated wide: high-vol up-sweeps mean-revert −0.303 ATR across 383 symbols (−0.402 on the screened subset), cost-viable only at ATR≥1%. The L5 crash-warning model (AUC 0.838–0.852, top-decile lift 4.2–5.0×) works as a risk gate, not an entry.

Delta dislocation (E9) from the orderflow video is decisively inverted on this universe — buying or selling the dislocation both lose (all 6 walk-forward folds strongly negative), and the signal predicts continuation, not reversal.

What the program produced

Hypothesis registry

Clusters A–L covering value-area reversion (A/B), CVD divergence (C), liquidation sweeps (J), funding microstructure (L), ITSM (G), execution anchors (K). Each with pre-registered gates.

Event-study harness

One memory-safe harness over 383 liquid perps (≥$1M daily turnover), 60d 5m bars, UTC days. Mean forward returns in ATR(15) units vs a random-entry baseline (−0.03 drift).

Strategy brief

32-strategy evidence-tiered menu (T1 peer-reviewed / T2 practitioner / T3 anecdotal), ranking what to pursue: 80% Edge Rule, ITSM, short-term reversal, OFI/CVD, sweep-reclaim.

Priority stack

J1 sweep-reclaim (validated) → E8 VWAP-fade (new strongest) → E4 IVA break-in → E3 IVA confirmed breakout → E5 CVD-divergence fade, with λ/Amihud gating and VPIN/funding as regime filters.

Headline event-study ranking (mean 1h forward return, ATR units)

RankStrategynMean 1h (ATR)Verdict
1E8 / S16 VWAP-deviation fade2,673,802+0.649Strong — verify cost/slippage; the revert-to-mean workhorse
2E4 / S4b IVA30 break-in fade51,316+0.411Strong, consistent (44.7%→higher on wide)
3E3 / S4a IVA30 confirmed breakout8,928+0.263Positive continuation (needs acceptance + volZ)
4E5 / S6 CVD divergence fade174,062+0.117Positive reversion, directionally correct
5E2 / S2 VA-rejection fade3,739,326+0.040Weak-positive (huge n, small edge)
6E7 / S11 extreme reversal158,133+0.047Fee-thin (1-bar horizon)
7E1 / S1 80% Edge6,615traverse 26%, fwd −0.074NULL — 80% rule NOT supported

What did NOT work

ITSM (E6/S9)

corr(first30m, last30m) = −0.017 (n=21,471) — no intraday session momentum at the UTC anchor, confirming the G4 null on wide.

Volume clock (E13/S31)

corr(past, fwd) = +0.004 (n=1.2M) — volume-clock adds nothing over time clock at this horizon.

Horizon map (E15/S10)

Hour-lag autocorr ≈ 0 (lag1 −0.039 → lag6 −0.002) — no persistent intraday momentum structure; slight 1-hour reversal only.

Overnight→intraday (E12/S29)

corr(first60m, rest-of-day) = −0.066 — very weak, no reliable reversal.

Delta dislocation (E9)

Decisively inverted — 6/6 WF folds negative (−0.31 to −0.48 ATR), net PF 0.71. Predicts continuation, not the claimed reversal.

Gauges (context only)

Kyle λ median 1.7e−7 · Amihud 1.8e−6 · VPIN median 0.52 (contested metric; filter-only). Hourly seasonality real but noisy: best 15:00 UTC +0.239 ATR, worst 18:00 −0.258.

The validated edge + risk gate

FindingResultRead
J1 sweep-reclaim (wide)−0.303 ATR @1h (383 syms)−0.402 screened · −0.256 unscreened — edge is in the signal, not the screen
ATR stratification−0.431% @ ATR≥1%Net-positive after ~0.3% costs only at ATR≥1%; −0.094% below 0.5% (cost-negative)
L5 crash-warningAUC 0.838–0.852Top-decile lift 4.2–5.0×; realized vol weight ~0.8; use as risk gate, not entry
Practical J1 rule≈ +0.14%/tradeFade high-vol up-sweeps (volZ≥1 through prior-day high, close back inside) on ATR≥1% symbols

Run it yourself

# strategy engine over the wide store, then site build (see scalping-strategies)
cd /home/jose/workspace/scalping-research
python3 backtest/strategy_engine.py --store scalping-data/store_wide --out-dir scalping-data \
  --screened-list scalping-data/screened_list.txt
python3 backtest/build_strategy_analysis.py --store scalping-data/store_wide --out-dir scalping-data
# hypothesis registry & brief
cat hypotheses.md && cat strategy_brief.md