OB midpoint reversion — the 10:30–11:30 trade on range days

A live strategy layered on the 09:30–10:30 opening-range box. Four signals (MidShort / MidLong / FlipShort / FlipLong), order-of-formation FSM, and a four-part backtest (faithful replay + stop sweep + entry-cutoff sweep + breakeven sweep) across 1,370 sessions. Ship config: MidLong + FlipShort only, MN_INACTIVE only, cutoff 11:30, OB-extreme stop, TOWARD-VWAP gate on MidLong. +19.4 pts/trade across ~154 trades/year.

The setup

After the 09:30–10:30 opening-range box finishes, price usually travels toward one extreme before retracing. The reversion trade: enter at the box midpoint, target the opposite extreme. Direction depends on which extreme formed first.

• High formed first, then low → price came down through mid once → on the next touch of mid from below, short targeting range low
• Low formed first, then high → price went up through mid once → on the next touch of mid from above, long targeting range high

The MidpointReversion.cs strategy was already live with this logic and a dashboard claiming 70–83% win rates per signal. The full-sample OOS replay disagreed — sharply — and a parameter sweep produced a much cleaner config.

The FSM

Inside the 09:30–10:30 window, track four state variables bar-by-bar (all thresholds recomputed live each bar — this matters for flip detection):

At 10:30 the FSM produces:

• isHighFirst = rangeHighTime < rangeLowTime
• flipLong = isHighFirst AND obHighApproached AND obHighApproachedTime > rangeLowTime
• flipShort = isLowFirst AND obLowApproached AND obLowApproachedTime > rangeHighTime

Flip logic = “price came back toward the first-formed extreme after the second extreme printed” — the original one-sided travel was broken, so trade the other way.

Four signals fall out:

Part 1 — faithful replay (backtest_midpoint_reversion_v2.py)

1,370 MNQ sessions, straddle entry (any bar with Low ≤ mid ≤ High), entry window 10:30–14:00, stop = opposite OB extreme, target = opposite OB extreme + 10 pts extension (the v1 sweep winner).

MidShort is structurally break-even. FlipLong is marginal. The in-strategy dashboard’s 70–83% claims don’t survive the full sample. Two signals carry the edge: MidLong and FlipShort.

Cross-and-close vs straddle

The live strategy required cross-and-close (Low ≤ mid AND Close > mid for longs). I A/B’d it against straddle (any bar touching mid):

Straddle wins. Cross-and-close waits for confirmation and eats 2.65 pts of edge per trade.

VWAP filter direction

The live strategy applied STRATEGY_PASS — long requires Close > VWAP, short requires Close < VWAP (trade with VWAP). My prior work on FILL signals used TOWARD_VWAP — long from below VWAP (mean-reversion toward VWAP). Tested head-to-head:

Standard mean-reversion signals want TOWARD-VWAP. Flip signals are momentum-like (price already broke the originating extreme) — they want WITH-VWAP. The live strategy’s VWAP filter was backwards for MidLong and correct for FlipShort.

MN HOLD cross-filter

Applied the MN HOLD classifier from mn-hold-side:

Symmetric opposite of the ORB × MN HOLD result: on trend days, mean-reversion at the OR midpoint gets run over — regardless of which direction you take it. MN_INACTIVE is the only safe universe.

The winning stack (Part 1 exit)

Part 2 — stop / cutoff / breakeven sweep (backtest_midpoint_reversion_v3_def.py)

Took the two winning cohorts (plus three comparison cohorts) and ran three parameter sweeps back-to-back.

D — fixed-tick stop vs OB-extreme stop

Tested stops of OB / 40 / 30 / 25 / 20 / 15 / 10 pts.

Tighter stops destroy the edge on every cohort. Average wins are +76 pts; average losses on the OB stop are −65 pts. The math only closes on the wide stop because the wins have to clear the losses by a full win’s worth, not a tight stop’s worth. Tight stops get hit before the reversion plays out.

E — entry cutoff sweep

Tested cutoffs 11:00 / 11:30 / 12:00 / 12:30 / 13:00 / 13:30 / 14:00. Winners in bold:

Pattern is universal: earlier is better, 11:30 is the sweet spot. The live strategy ran to 14:00 and caught a lot of fading post-lunch setups.

F — breakeven move

Tested moving stop to entry after price moved ≥ 50 pts in favour.

BE@40 is a rounding-error improvement on MidLong (+0.67/tr on 86 trades). On FlipShort and every other cohort the BE hurts materially. BE turns wins into flat exits faster than it saves losers, and the win-size distribution (+70–80 pt average) dwarfs the stop savings. Conclusion: don’t use breakeven.

The ship config

Combined expected performance on the live stack (MidLong + FlipShort under V3 defaults):

• n ≈ 154 trades across 1,370 sessions (~11% strike rate)
• ~52% win rate
• ~+19.4 pts/trade expectancy
• ~+2,978 total pts

Implementation

• MidpointReversion.cs — upgraded with seven new V3 parameters defaulting to the winning config:
  • MnHoldSkipActive (reads hud_data.txt, skips entries on active MN HOLD sessions)
  • EntryCutoffHHMM (default 1130)
  • UseStraddleEntry (default true)
  • InvertMidLongVwap (default true — MidLong requires Close < VWAP)
  • EnableMidLong / EnableMidShort / EnableFlipShort / EnableFlipLong (individual kill switches; MidShort + FlipLong default OFF)
  • HudDataPath for the MN HOLD reader
  • Legacy behavior is still reachable via the toggles, so live A/B is possible without rebuilding.
• PredictionModel.cs / PredictionModelWeb.cs — added OB MID and OB MID MAGNET to the MN HOLD confluence dictionary with {−5 aligned, −20/−18 opposed, 0 PRIME}. Mean-reversion signals get a light aligned penalty and a heavy opposed penalty because the backtest showed MN-active sessions hurt mid-reversion regardless of direction agreement.

Related

• Underlying line model: OB RANGE (model 10)
• Trend-day cross-filter: MN HOLD-SIDE (model 26) and the full research writeup
• Inverse case (why ORB direct breakouts want MN HOLD aligned): ORB break & retest targets

History

• 2026-04-24 — Full backtest + strategy upgrade shipped. backtest_midpoint_reversion_v2.py (signal + VWAP + entry-pattern sweep) and backtest_midpoint_reversion_v3_def.py (stop / cutoff / breakeven sweep) under C:\SMC\scripts\. Results archived at C:\SMC\backtest_results\midpoint_reversion_v2.txt and midpoint_reversion_v3_def.txt. MidpointReversion.cs patched; confluence table updated in both indicators; this writeup published.