momentum-system/docs/research/orderflow_notes.md

Microstructure Shelf → ORB Entry-Conversion: Synthesis Note

The bot's proven #1 problem is entry conversion: triggers fire but only ~35-48% of breakouts reach +1R before -1R, and eight selection-side levers have been refuted — the population is fine, what happens AT and AFTER the breakout is the bottleneck. This shelf (Hasbrouck, Bouchaud et al., Cartea et al., Cont-Kukanov-Stoikov, Dalton, Bellafiore, Neill, Cetin-Danilova) is unusually well-matched to that problem, because market microstructure is precisely the study of which price moves stick (permanent, information-driven impact) and which revert (transitory, liquidity-driven impact). The academic and practitioner books converge, from opposite directions, on the same handful of observable signatures — and critically, most are computable from 1-min bars and 15s quote snapshots without tick data. Everything below is either a conditioning gate on already-fired entries or an exit-config variant; nothing re-opens the refuted selection axis.

Convergent mechanisms

1. Failed breakouts are transitory impact; the default fate of a one-shot volume burst is reversion. (Hasbrouck ch.8/9/15; Bouchaud ch.12.4; Cartea ch.4/10.4) Price moves decompose into a permanent (information) component and a transitory (liquidity) component that mean-reverts once the initiating flow stops. Bouchaud is bluntest: after a metaorder completes, impact decays abruptly — reversion is the empirically normal outcome. A ~50-65% breakout failure rate is not a bug in the setup; it is the base rate for one-shot bursts. Continuation requires sustained flow.

2. Return into the broken range is the structural invalidation signal, and it arrives before the stop. (Dalton — balance-area re-entry; Bellafiore — fake-breakout recross at AIG $25/AMZN $86.60; Neill — false moves; Hasbrouck — the transitory signature is the round-trip) Four independent traditions say the same thing: once price trades back through the level it broke, the participants who drove the break are gone and the trade thesis is dead — waiting for the fixed stop is paying extra for the same information.

3. Continuation is fed by persistent same-side flow, not by the size of the trigger print. (Bouchaud ch.10 — long-memory order flow from order splitting/herding; Cartea ch.4/12 — same-side MO clustering, imbalance regimes with 0.97-0.98 persistence; Hasbrouck ch.9 — momentum-feedback impulse responses that keep rising; Neill — volume growing on the advance) The discriminator between real and fake breakouts is whether one-directional participation was ongoing into the trigger, versus a single isolated spike.

4. Spread dynamics around the trigger carry real-time information — but the books disagree on the sign. (Cartea ch.2, Hasbrouck ch.5: widening = Glosten-Milgrom adverse-selection pricing = informed flow = bullish for follow-through. Bouchaud ch.22, Cont-Kukanov-Stoikov (β~1/D), Bellafiore: widening = liquidity pulling back / thin book / chop-zone fragility = bearish.) Five sources agree the spread's relative behavior at trigger is informative; the direction is an empirical question and must be pre-registered two-sided.

5. Speed and shape of the initial rejection through the level grades conviction. (Dalton — Open-Drive vs Open-Auction taxonomy, tails "made by time" carry no conviction; Neill — gradual advance on steady volume = durable buying, vertical spike = short-covering; Bellafiore — held-bid absorption precedes real breaks) A decisive drive that doesn't look back differs measurably, at the moment it happens, from a limp tag of the level — this is same-bar shape, not delayed confirmation.

6. Heavy volume with no price progress = exhaustion, at any timescale. (Neill — "churning"; Dalton — time-beyond-level without acceptance; Bellafiore — chop zones) A joint volume-high + price-stall signature — explicitly distinct from the refuted volume-decline fade.

7. The open is a structurally thin, high-impact regime; every quote-based feature must be normalized to the name's own trailing baseline. (Bouchaud ch.4.2 intraday U-shape; CKS — depth 2x lower, impact 5x higher 9:30-10:00; Bellafiore — time-of-day regimes) Not a lever itself — a normalization requirement that guards every hypothesis below against re-deriving the refuted liquidity floor.

Ranked hypotheses

H1 — Trigger-erasure / recross fast invalidation (exit_config)

Convergence: 5 books (Dalton, Bellafiore, Neill, Hasbrouck, Bouchaud). Strongest on the shelf.

Statement: Fires where a post-entry 1-min bar closes back below the ORB trigger price within the first 5 minutes are exhibiting the transitory-impact / balance-re-entry / false-move signature and are disproportionately headed to -1R; exiting on the first such close (rather than riding to the stop) reduces average realized loss without materially truncating winners.

Computable features (1-min bars only): trigger_price from entry-gate receipts; per post-fill bar, erased = close < trigger_price; first_erasure_bar_index within a 5-bar window; secondary variant recross_count = number of bars whose low trades back through trigger. Trigger must be defined from the OR-high, independent of the stop price (no tautology).

Draft pre-registration:

population: {min_entry_gap_pct: 0.10, min_day_volume: 300000, price_min: 2.0, price_max: 20.0}
intervention: {kind: exit_config, rule: "exit at market on first 1-min bar CLOSE < trigger_price within 5 bars post-fill; else standing exit stack", sweep: {window_bars: [3, 5], condition: [close_below, low_through]}}
bars: {expR_delta_min: 0.05, min_n: 300}
fidelity: {atr_frac: 0.25, spread_mult: 3, adverse_ticks: 1}

Expected direction: Erasure fires convert to +1R at a materially lower rate; the early-exit variant lifts fire-weighted expR by shrinking the loser tail.

Kill criteria: expR delta < +0.05 across the sweep; OR lost-winner rate > 15% of eventual +1R trades exited by the rule (erasure-then-recovery fakeouts eating the gain); OR the rule's benefit vanishes when BE@+5% ratchet is active (redundant with existing protection).

Confound guards: Family resemblance to refuted volume-fade exits — but this is price-structure, not volume, and must independently clear the same bar. Measure lost-winner rate explicitly, not just loss reduction. Use consolidated bar closes, not quote mids, to avoid bid-ask bounce on thin names. Report separately by OR-width tercile so wide-range chop doesn't drive the pooled result.

H2 — Pre-trigger order-flow persistence gate (keep-or-zero)

Convergence: 4 books (Bouchaud, Cartea, Neill, Hasbrouck).

Statement: Fires preceded by persistent same-direction bar-level flow (volume-weighted sign agreement over the run-up window) are the signature of an ongoing metaorder/herd and convert better than fires whose pre-trigger flow was choppy/balanced — the trigger is a mid-stream observation of sustained flow, not a one-off spike.

Computable features (1-min bars only): Over the last 15 closed pre-trigger bars: sign_i = sign(close_i − open_i); imbalance_persistence = |Σ sign_i · vol_i| / Σ vol_i; secondary agree_frac = fraction of bars with sign matching breakout direction. No look-ahead into the trigger bar.

Draft pre-registration:

population: {min_entry_gap_pct: 0.10, min_day_volume: 300000, price_min: 2.0, price_max: 20.0}
intervention: {kind: keep_or_zero_gate, rule: "zero fires in the bottom imbalance_persistence tercile", stratify_by: [gap_pct_bucket, rvol_bucket]}
bars: {expR_delta_min: 0.05, min_n: 300}
fidelity: {atr_frac: 0.25, spread_mult: 3, adverse_ticks: 1}

Expected direction: Top tercile beats bottom tercile on +1R-before--1R rate; zeroing the bottom tercile lifts kept-population fire-weighted expR ≥ +0.05.

Kill criteria: Monotonicity fails across terciles; OR the spread between top and bottom tercile collapses inside gap%/RVOL strata (i.e., it was a refuted selection lever in disguise); OR fewer than min_n fires per tercile.

Confound guards: Bar-sign is a crude tick-rule proxy — accept the noise, demand the effect survive stratification by gap% and RVOL (the refuted axes) before believing it. Check redundancy against the surviving runner-grade before shipping as an independent lever.

H3 — Spread-expansion-ratio at trigger (keep-or-zero, two-sided)

Convergence: 5 books cite the mechanism (Cartea, Hasbrouck, Bouchaud, CKS, Bellafiore) — but they split on direction, so this is registered two-sided.

Statement: The spread's behavior at trigger, normalized to the name's own trailing baseline, discriminates converting from failing breakouts. Adverse-selection theory (Cartea/Hasbrouck) predicts widening = informed flow = keep; fragility theory (Bouchaud/CKS/Bellafiore) predicts widening = liquidity vacuum = zero. The test determines which regime dominates $2-20 gappers.

Computable features (15s quote snapshots): spread_baseline = median((ask−bid)/mid) over the 10 min pre-trigger; spread_at_trigger from the last snapshot at/before trigger; expansion_ratio = spread_at_trigger / spread_baseline; secondary post-fill variant post_ratio(1min) = spread at t0+1min / baseline.

Draft pre-registration:

population: {min_entry_gap_pct: 0.10, min_day_volume: 300000, price_min: 2.0, price_max: 20.0}
intervention: {kind: keep_or_zero_gate, rule: "two-sided: compare top vs bottom expansion_ratio tercile; zero the losing tercile only if pre-registered margin met", normalize: per_name_baseline, stratify_by: [price_bucket, rvol_bucket, minutes_since_open]}
bars: {expR_delta_min: 0.05, min_n: 300}
fidelity: {atr_frac: 0.25, spread_mult: 3, adverse_ticks: 1}

Expected direction: Two-sided by design; a tercile spread in conversion rate that is stable across the backtest window and survives stratification, in either direction.

Kill criteria: No stable tercile ordering across time halves of the window; OR the effect disappears within price/RVOL strata (= liquidity floors relitigated → discard permanently); OR >20% of fires lack a usable pre-trigger snapshot baseline.

Confound guards: Ratio-to-own-baseline only, never absolute bps (absolute spread = refuted liquidity floor). Stratify by minutes-since-open (Bouchaud/CKS U-shape mechanically widens everyone's open spread). Test for independent value vs H1 — post-fill spread blowout and price erasure are correlated symptoms of the same reversal.

H4 — Volume ramp vs isolated-spike trigger (keep-or-zero)

Convergence: 3 books (Bouchaud, Neill, Hasbrouck).

Statement: Fires where volume built progressively across the pre-trigger bars (a ramp — consistent with an unfinished metaorder still executing) convert better than fires where nearly all volume arrived in the single trigger bar with no build-up (a completed one-shot burst whose impact, per Bouchaud, reverts once done; per Neill, the vertical short-covering spike that "dies at the top").

Computable features (1-min bars only): concentration_ratio = trigger_bar_volume / Σ(volume of preceding 5 bars); ramp_slope = sign of linear volume trend over the last 5 pre-trigger bars. Strictly pre-trigger — no post-entry volume conditioning (keeps it categorically distinct from the refuted volume-fade exit).

Draft pre-registration:

population: {min_entry_gap_pct: 0.10, min_day_volume: 300000, price_min: 2.0, price_max: 20.0}
intervention: {kind: keep_or_zero_gate, rule: "zero fires in the top concentration_ratio tercile (isolated-spike triggers)", stratify_by: [rvol_bucket, gap_pct_bucket]}
bars: {expR_delta_min: 0.05, min_n: 300}
fidelity: {atr_frac: 0.25, spread_mult: 3, adverse_ticks: 1}

Expected direction: Isolated-spike (high concentration) fires convert worse; ramp fires convert better; zeroing spikes lifts kept expR.

Kill criteria: No monotone relation across concentration terciles; OR effect collapses within RVOL strata; OR feature is >0.7 correlated with H2's imbalance_persistence on the fire population (redundant — keep only the stronger).

Confound guards: Must survive RVOL stratification (day-level RVOL is a refuted axis; this is intrabar structure). Check against how the gap formed premarket (instant-gap vs premarket-ramp names) to avoid re-deriving runner-grade. Single odd-lot block prints on thin names can fake a spike — winsorize trigger-bar volume at a per-name percentile.

H5 — Post-entry churn/stall exhaustion tightening (exit_config)

Convergence: 3 books (Neill, Dalton, Bellafiore).

Statement: Positions showing, within the first 3 minutes post-fill, either (a) no new 1-min bar high above the entry bar's high (Dalton: no range extension = weak other-timeframe control), or (b) volume spiking above baseline while price makes no net progress (Neill: churning), are disproportionately failures; tightening the stop to breakeven-minus-buffer on that signal cuts loser size without materially cutting winners.

Computable features (1-min bars only): entry_bar_high; first_new_high_bar_index within 3 bars; joint churn flag on bars t0+1, t0+2: vol_ratio = bar_vol / trailing_15bar_mean > 2.0 AND |close − entry_price|/entry_price < 0.5 × trailing_bar_range_mean.

Draft pre-registration:

population: {min_entry_gap_pct: 0.10, min_day_volume: 300000, price_min: 2.0, price_max: 20.0}
intervention: {kind: exit_config, rule: "on stall (no new high in N bars) OR churn flag, tighten stop to entry − 0.25×ATR; sweep N in [2,3,4]", baseline: current_stop_stack}
bars: {expR_delta_min: 0.05, min_n: 300}
fidelity: {atr_frac: 0.25, spread_mult: 3, adverse_ticks: 1}

Expected direction: Average realized loss on eventual losers shrinks; eventual-winner capture rate stays within 10% of baseline; net fire-weighted expR ≥ +0.05.

Kill criteria: Winner-capture drops >10% at every N in the sweep (whipsaw cost exceeds loss savings); OR the churn leg alone reproduces the refuted volume-fade result (volume component fires without the price-stall component adding anything — then drop the volume leg and keep only price-stall, or kill entirely); OR gains are fully redundant with H1 erasure exits on the same trades.

Confound guards: This is a joint high-volume+stall or no-new-high signature, never volume-decline (the refuted fade). Run H1 first; only fires H1 doesn't already catch count toward H5's incremental value. Sensitivity-sweep N — real winners on small caps routinely chop 2-3 bars before extending.

Rejected during synthesis

• Gap-still-open-at-trigger gate (Dalton) — single source; heavily confounded with time-since-open and gap% (refuted axis); clean but keep on the bench behind the five above.
• Round-number proximity gate (Cartea) — single source; evidence from a $450 large-cap at penny resolution, transfer to $2-20 names unvalidated; mechanically confounded with price level.
• Time-since-open depth-regime gate (CKS) — single source; collides with the dead 10:30-cutoff history and with early-trigger/gap% correlation; refuted-adjacent selection flavor.
• Pre-trigger quote-size imbalance (Cartea/CKS) — hard dependency on bid_size/ask_size existing in the snapshot logs; unverified = dead on arrival; the paper's OFI variable explicitly needs queue data we don't have.
• Volume-node thickness at trigger (Bellafiore) — single source; volume-at-price concentration is RVOL in new clothes without careful shape-normalization; weakest power.
• Overhead same-day resistance distance (Neill) — single source, and the book itself flags naive levels as unreliable "until confirmed by the action of the stock."
• Bid-erosion early exit (Bellafiore) — single source; 15s sampling too coarse for "held bid"; largely restates mark-to-market drawdown; folded conceptually into H1/H3.
• Impulse decay half-life adaptive stops (Hasbrouck) — 5 noisy data points per fire can't support a fitted decay rate; strictly a refinement of H1 if H1 shows signal.
• Trigger-bar close-location drive gate (Dalton) / trigger-bar volume conviction (Neill) — too close to the refuted close-above-trigger confirmation family; their live content (shape/participation quality) is already carried by H2 and H4 with better mechanism backing.
• Everything from Cetin-Danilova — existence-proof mathematics over unobservable signed order flow; zero computable features.

Data upgrades that would unlock more

1. Verify whether bid_size/ask_size are in the 15s quote snapshots — if yes, the CKS order-flow-imbalance family (the highest-R² predictor on this shelf) becomes partially computable; costs one grep.
2. Tighten quote snapshot cadence to ~5s for watched names in the 5 minutes around triggers — halves the aliasing that weakens H3.
3. Log NBBO at the exact trigger/fill instant in the entry receipts (not nearest-15s) — removes the largest measurement error in every spread hypothesis.
4. Per-minute uptick/downtick trade counts (available from SIP aggregates on some feeds) would upgrade H2's crude bar-sign proxy to a real tick-rule imbalance.
5. Tick-by-tick trades remain the true unlock for trade-sign imbalance and held-bid detection — everything above is the best available proxy without them.