Eight Experiments on What Actually Works Around VWAP — and What Only Looked Like It Did

Dataset: CME Micro Nasdaq-100 continuous futures, January 2024 through August 2026 (~2.5 years; 930,136 1-minute bars / 277,293 5-minute bars). Parity anchor: a real TradingView export, 392 trades, $18,314.50 net PnL on the out-of-sample window. Eight separate experiments, run in VectorBT Pro.

Experiment 1 — ATR-relative sizing vs. the static-points baseline

Sweeping the v4.0.0 per-session ATR-multiplier inputs in-sample (2024-01-01 → 2026-05-10, 900 combos/session) and validating the top performer out-of-sample (2026-05-10 → 2026-08-22, holdout).

London’s IS window flips from -$1,705 to +$1,626.50 under ATR sizing, and full-history max drawdown nearly halves (9.93% → 5.51%). Asia and NY did not show the same benefit — both perform better under their existing fixed-point brackets (45/85 Asia, 35/70 NY), likely because their intraday ranges already conform reasonably well to static levels. Recommendation: switch London to ATR-relative sizing, leave Asia and NY on points.

Experiment 2 — Confluence filter tuning (Efficiency Ratio & MACD Velocity)

Sweeping Kaufman’s Efficiency Ratio and MACD Histogram Velocity in-sample to see if tighter signal gating eliminates false entries without over-filtering real ones.

Optimal settings found: London ER period 9 / threshold 0.15 / fade-drop 0.15 (with MomVel active); NY ER period 4 / threshold 0.05 / fade-drop 0.15.

The tuned fade-veto settings cut 82 low-probability trades and lifted full-history net PnL from $16,609.50 to $20,089.00, with Sharpe improving from 1.744 to 2.253.

Experiment 3 — MNQ ATR(14) characterization

Correlation between MNQ price level and raw-point ATR: r = 0.2317 (Pearson), 0.2854 (Spearman) — meaningfully weaker than the multi-decade cross-instrument correlation found in the companion volatility-sizing study (which used yearly-aggregated data across 26 years; this measures bar-level correlation within one 2.5-year window). ATR does not trend mechanically with price level over this window — sustained bull runs (16k → 30k) often coincided with compressed ATR, not expanded.

Volatility regime quartiles across 825 trading days:

ATR expands 4.36x between the quietest and loudest quartiles. The 1st-percentile bar ATR is 4.05 points — a 10-point floor on any ATR-relative bracket is doing real work, preventing degenerate sub-10-point stops during quiet sessions.

Experiment 4 — Exploratory VWAP entry mechanics (161,439 bars, frictionless)

An initial, theoretical forward-excursion study across several VWAP-based entry archetypes — band fades, momentum continuation, liquidity sweep reclaims, break-and-retest. Band fades looked like the standout:

This experiment measured pure bar-level MFE/MAE with no execution model at all — see Experiment 5 for why that matters.

Experiment 5 — Realistic execution reality-check on Experiment 4

Re-running the standout Band-Fade and Liquidity Sweep findings under actual execution assumptions: next-bar-open fill, 1-tick slippage ($0.25/pt entry+exit), $0.62/order commission, conservative same-bar stop-before-target tie-breaking, and 5 discrete 6-month in-sample blocks plus an independent OOS holdout.

The band-fade edge does not survive realistic execution. The Experiment 4 numbers were inflated by evaluating pure theoretical MFE/MAE with no fill delay, no slippage, no commission, and no single-position lock. Under real conditions, the -2.0σ fade flips from a promising +$918 in-sample result to a losing -$1,369 out-of-sample, and the +2.0σ upper-band fade is consistently losing across every block tested. Verdict: do not build a standalone band-fade module into the live strategy. This is the single most important finding in this lab — a theoretical edge that evaporates the moment realistic friction is applied is worse than no edge at all, because it would have shipped.

Experiment 6 — Asia session, 4-month temporal chunks

Testing why the shipped Asia config drew down in 2024, by measuring what % of Asia trades actually reached a given target across 8 sequential 4-month windows:

In early 2024, only 7.6% of Asia trades ever reached 40 points — with a shipped 45-58pt TP, over 92% of trades stalled around +15-25pts and reversed into the -85pt SL. By mid-2026, 60.8% of trades reached 40pts. The 45+ point Asia TP was structurally overfit to a high-volatility period that didn’t hold across the full window.

A 1,560-combination grid sweep (TP 10-50pts, SL 25-100pts, VWAP slope 0.5-2.5pts, ER filter on/off/threshold, MACD velocity) found configurations profitable in all 8 of 8 chunks — notably a 30pt TP / 50pt SL / 1.5pt slope / ER on (0.15) / MACD velocity on setup: 401 trades, +$6,311.26, 71.82% win rate, PF 1.62, profitable in every single chunk, worst-chunk PnL still positive at +$77.34.

Experiment 7 — VWAP slope momentum expansion & drift-retest mechanics

The mental model: VWAP’s slope is a lagging, volume-weighted reaction to price already printed. Entering short purely because price already fell hard and VWAP is sloping down means selling into an extended zone where institutional liquidity tends to absorb the move — the “chasing expansion” trap. The edge instead sits on the first minor pullback back toward the sloping VWAP (0.2-0.8x ATR away), entered in the direction of the confirmed volume trend.

Session breakdown of where this edge actually lives (13,604 bars, active down-slopes on NQ):

Notably, this reads opposite to what a same-day internal chart review suggested (Asia read as weak/continuation-only, NY as the strong session) — this dataset shows Asia actually holding the best edge ratio of the three, with NY the weakest without added filtering. Worth resolving directly against Tokyo Drift’s own VWAP Expansion mode results before trusting either read in isolation.

Empirical MFE reach rates once a drift-retest signal fires: 0.5x ATR reached 89.7% of the time, 1.0x ATR 80.6%, 1.5x ATR 73.5%, 2.0x ATR 65.8%, 3.0x ATR 49.0%. Bar-by-bar excursion peaks around bar 12 (60 minutes, 2.30x ATR, 53.9% of trades in profit) before mean-reversion risk climbs.

Experiment 8 — First 5-minute VWAP cross (“trend inception”) scalping

Does entering immediately on the very first 5-minute candle closing across VWAP work, without waiting for a pullback? Raw, unfiltered (N=1,245 signals):

Requiring the crossing candle to show real “expansion energy” (volume >1.2x its 20-bar average, or Kaufman ER >0.30) meaningfully improves results — PF rises from 1.17 to 1.31-1.33, EV from +2.4 to +4.2-4.5 pts.

Session breakdown: London is the best standalone inception-cross session (46.3% WR, PF 1.38 at 40/25) — “clean expansion follow-through.” NY is solid when backed by a volume surge. Asia shows “high VWAP ping-pong” and underperforms without waiting for pullback confirmation (39.3% WR, PF 1.03) — consistent with Asia needing the retest-based entry rather than the raw inception cross.

The underlying microstructure reality: across all 1,245 fresh crosses, 83.0% pulled back to touch VWAP again within the first 5 bars (25 minutes) — only 17.0% were “runaway” moves that never looked back. This is why a staggered two-order model (a small inception probe plus a limit order at the retest) outperforms either approach alone.

Consolidated recommendations

1. Do not build a standalone band-fade module — Experiment 5 showed it fails realistic execution.
2. London: switch to ATR-relative sizing (2.0x/1.0x drift-pullback bracket, or the swept 4.0x/5.0x); Asia and NY: keep fixed points, but revisit Asia’s TP given Experiment 6’s finding that 45+pt targets were tuned to an unrepresentative high-volatility stretch.
3. A staggered entry model — half-size on the inception cross (gated by volume/ER), half-size as a limit order at the VWAP retest (0.3x ATR) — captures both the 17% of runaway moves and the 83% that pull back first.
4. NY needs the most filtering of the three sessions across nearly every experiment in this lab — gate it with Efficiency Ratio and MACD Velocity rather than trading it on raw VWAP conditions alone.