Webull TRAIL Lab — Can a Vision-Read Regime Pick a Defined-Risk Option?NULL-HYPOTHESIS STUDY

Webull TRAIL Lab · Phase 1 (paper mode) · 2026-06-17 · ~9 min read · N = 0 closed option-trades at publication · pre-registered protocol

Abstract. The Webull TRAIL Lab is the options-rail successor to the TRAIL Lab (Trailing Reality-Adjusted Inference) — the lab that fed a GPT-4o read of a raw chart frame into a paired-leg paper bet and measured whether the paper edge survived friction. It tests one pre-registered null hypothesis, stated in full here and never softened downstream: H₀ — a vision-extracted regime does not let a defined-risk options selector beat break-even; the "vision-read direction → defined-risk vertical edge" claim is false. TRAIL Phase 1.5 measured a 36% directional hit rate from the vision model on a small live sample — below random — and reported it as the finding the lab exists to produce. This lab asks the obvious follow-on: when that same vision read selects a defined-risk spread instead of a symmetric straddle, does the capped downside plus payoff-shaping rescue an edge, or does the weak directional read simply propagate? We reject H₀ only on the §3 evidence; a failure to reject is the lab's expected and fully successful outcome. No real order is placed this phase; every leg is a dry_run=True broker echo.

Contents

Lineage — a 36% directional read, reported honestly
Method — classifier, selector, calibration, ergodic evaluation
Null hypothesis & stop criterion
What a failure to reject tells us
Phase 2 — must-do / must-not
References

1. Lineage — a 36% directional read, reported honestly

This lab descends from the TRAIL Lab (trail_lab_service.py + vision_extractor.py), documented in the TRAIL methodology whitepaper and the Phase 1.5 study. TRAIL's founding mission is that the engine infers edge from the screen capture: a multimodal model (GPT-4o) reads the rightmost candles and returns {direction, confidence, last_close}, and the engine opens a paired call/put straddle sized by an asymmetric ¼-Kelly tilt proportional to that confidence, then grades each leg against a re-captured close under a pristine/realistic double-entry friction book.

Phase 1.5's headline result is the one this lab must keep front and centre, quoted from the study:

"The vision model did not produce a directional edge on this run. On the 11 ticks where the asymmetric tilt was non-zero … it called the realised 2-minute sign correctly 4 times — a 36.4% hit rate against a 50/50 null. … A friction-honest harness that returns 'your gut was wrong' is doing its job." — trail_lab_phase1_5_study.html §6

Phase 1.5 also surfaced three mechanisms that bear directly on the options pivot: (1) the straddle insured against direction errors — a wrong directional read could still book a profit because the structure was hedged, which means the binary straddle was already doing payoff-shaping by accident; (2) the model showed a systematic short bias (10 of 11 directional ticks bet short), a calibration defect, not a directional one; (3) vision latency and resolution drift are real, unbudgeted systematic errors. These map onto the pivot's headline findings #1 (broken calibration in our direction) and #3 (b ≈ 1 is a lie). The Webull TRAIL Lab replaces the always-both-legs straddle with an explicit, selected defined-risk structure, turning the accidental payoff-shaping of Phase 1.5 into the deliberate object of study (finding #5).

2. Method

2.1 The regime classifier — which forecast engine feeds it

The classifier is vision_extractor.extract_chart_state() wrapping GPT-4o (OpenAI 2024), used unmodified as the forecast layer. It returns {direction, confidence, last_close, support, resistance} with confidence clamped to [0.5, 0.95]. The TRAIL lab maps this to the §2.2 regime block: direction from the vision read, strength from the clamped confidence, and a vol_regime hint from the support/resistance band width relative to recent range. As in Phase 1.5, a missing confidence defaults to 0.5 ("default to no edge, never default to no tick") — which in options terms means a neutral, near-symmetric defined-risk structure rather than a skipped observation.

2.2 The strategy selector — the candidate set

The selector is restricted, for this lab, to the defined-risk directional subset of the Periodic Table (WEBULL_PIVOT §3) — the spreads that turn a directional read into a capped-risk payoff:

Vision read	Primary candidate	Alternates
Long, confidence high, IV-rank mid/low	`bull_call_spread`	`long_call`, `bull_put_spread`
Short, confidence high, IV-rank mid/low	`bear_put_spread`	`long_put`, `bear_call_spread`
Direction held, confidence ≈ 0.5 (neutral default)	`long_iron_butterfly`	`long_call_calendar_spread`
Vision returned `ok=false` (unreadable frame)	declined — no structure, counted stale	—

The iron butterfly in the neutral default is the deliberate heir to Phase 1.5's symmetric straddle: when the model has no conviction, the lab takes a defined-risk, range-bound structure rather than a directional one — no edge claimed, no edge missed, the friction clock still runs. The ok=false path preserves TRAIL's honesty rule: a misread is a declined measurement, never a fabricated number, and the leg falls through to a counted, zero-P/L stale state exactly as a dead OHLC feed did.

2.3 The calibration pipeline — reuse, do not reinvent

Closed trades persist on the existing paper_trade_service.py row (no parallel DB) with the §2.4 columns. The Phase 1.5 short-bias and overconfidence findings make the calibration columns load-bearing here: webull_strategy_id + the per-bucket Wilson lower bound let us see whether the vision read's stated confidence maps to realized win rate on each spread type. /api/cv/paper/calibration?strategy_id= buckets bull_call_spread apart from bear_put_spread so the short bias (if it persists) shows up as one bucket carrying the sample. Wilson 95% lower bound per bucket (Wilson 1927).

2.4 The ergodic Monte-Carlo evaluation (F7)

Break-even is p* = 1/(1+b) off the vertical geometry; sizing is fractional Kelly off p_lower (F7 README — never p_hat), the direct successor to TRAIL's asymmetric ¼-Kelly tilt. The additive ergodic Monte Carlo (Peters 2019) reports time-average growth, the same instrument that turned Phase 1.5's $100 → $73 bankroll line into a finding rather than a verdict. A positive ensemble EV with ≤ 0 growth rate is a fail (finding #2).

Inherited threats to validity. Phase 1.5 flagged vision-model nondeterminism, resolution-time drift, and OTC-vs-real-market differences. On the Webull rail the resolution-drift problem largely disappears — option marks come from the broker chain at close, not a re-captured screenshot — but vision nondeterminism at entry remains. The selector's confidence-to-structure mapping must therefore be robust to a ±jitter in the extracted confidence; this is why a near-0.5 read routes to a symmetric defined-risk structure rather than a sharply directional one.

3. Null hypothesis & stop criterion

H₀ (pre-registered): a vision-extracted regime does not let the defined-risk selector beat break-even — across every funded webull_strategy_id bucket, the Wilson lower bound p_lower stays at or below the geometry break-even p* once friction is applied.

We reject H₀ only if, over ≥ 100 closed paper option-trades in at least one webull_strategy_id bucket, the Wilson lower bound exceeds break-even:

p_lower(bucket) > p* = 1 / (1 + b), b = max_gain_usd / max_loss_usd (bucket median)

and the stated vision confidence is calibrated to within 5 points of that p_lower for the bucket (the Phase 1.5 overconfidence clause). A bucket that clears p* only because one direction happened to win during a trending sample — while the confidence remains miscalibrated and the short-bias persists — does not reject H₀; it reproduces the Phase 1.5 pathology in a new wrapper. The per-strategy bucketing is what distinguishes "vision predicted direction" from "the tape trended our way."

4. What a failure to reject tells us

Per the lab's stance — set explicitly by the Phase 1.5 study — a null-not-rejected result is a successful run. A sustained failure to clear p* teaches:

Payoff-shaping does not rescue a 36% read. If a defined-risk spread still cannot beat break-even on a vision read that hits 36% directionally, we have learned that the bottleneck is perception, not structure — no amount of clever leg-building compensates for a direction call worse than a coin flip. That redirects Phase 2 effort to the extractor (prompt, calibration, self-consistency), not the selector.
The short bias is the defect, not the direction. If one directional bucket carries the whole sample and underperforms, the Phase 1.5 diagnosis stands: 0.90 confidence on a model that hits 36% is the actual defect. Damp the tilt, calibrate the confidence — do not flip the sign.
Defined-risk capping held. A clean null with bankroll intact confirms the defined_risk_only guard plus the iron-butterfly neutral default kept a confident-and-wrong vision read (the ticks 5–8 pathology of Phase 1.5) from compounding into ruin.

The invalidating failure (vs. a valid null) is any dry_run=false broker response, any undefined-risk leg booked while defined_risk_only=true, or — TRAIL-specific — a fabricated last_close from a frame the extractor could not actually read. The first two are caught by the smoke-test sign-off; the third is contained by the ok=false → stale path.

5. Phase 2 — must-do / must-not

Must-do:

Reliability diagram per strategy bucket — the direct options heir to Phase 1.5's calibration recommendation: of all entries the vision model called 0.90, what fraction of bull_call_spread trades won? Below-diagonal → damp the confidence-to-tilt mapping.
Per-strategy theta budget — verticals and butterflies carry decay the binary straddle did not; track theta_paid_usd vs realized R per bucket.
Adversarial-frame jitter test — feed the extractor the same frame N times, measure direction-flip rate, and confirm the near-0.5 → symmetric-structure routing absorbs the jitter (Phase 1.5 §10 carried forward).
Binary-payout variant — if the lab eventually routes Pocket-Option-style binaries, add the +0.8/−1.0 contract; this is the exact contract Phase 1.5 said the original "80% hit rate" intuition should be benchmarked against.

Must-not:

Tune the selector or the prompt on n < 100. "There is no signal here, only setup" — Phase 1.5, verbatim.
Deploy capital — the lab remains a measurement instrument; capital follows a friction-honest sustained p_lower > p* across many sessions and symbols.
Flip defined_risk_only=false while p_lower < p* — a confident-and-wrong vision read on an undefined-risk leg is precisely the failure mode the defined-risk default exists to contain.

References

Cox, J. C., Ross, S. A., & Rubinstein, M. (1979). Option Pricing: A Simplified Approach. Journal of Financial Economics, 7(3), 229–263. (Binomial valuation of the verticals/butterflies the selector emits; alternative to Black & Scholes 1973.)
Black, F., & Scholes, M. (1973). The Pricing of Options and Corporate Liabilities. Journal of Political Economy, 81(3), 637–654.
Kelly, J. L. (1956). A New Interpretation of Information Rate. Bell System Technical Journal, 35(4), 917–926. (Payoff ratio b, break-even p* = 1/(1+b); successor to TRAIL's asymmetric ¼-Kelly tilt.)
Wilson, E. B. (1927). Probable Inference, the Law of Succession, and Statistical Inference. Journal of the American Statistical Association, 22(158), 209–212.
Peters, O. (2019). The Ergodicity Problem in Economics. Nature Physics, 15, 1216–1221. (Time-average vs ensemble growth — the F7 objective that turned Phase 1.5's bankroll line into a finding.)
OpenAI (2024). GPT-4o System Card / Hello GPT-4o. The multimodal vision backbone behind the TRAIL extractor. link
Taleb, N. N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House. (Fragility of edges estimated from thin samples; optional.)
Lineage: TRAIL Lab — Trailing Reality-Adjusted Inference (methodology) and the Phase 1.5 Study (36% directional hit rate, vision mode).

Webull TRAIL Lab · Phase 1 (paper mode) · 2026-06-17 · simulation only — no capital is ever at risk. Calibration reused from paper_trade_service.py (no parallel DB). Companion instruments: Edge, Exhaustion, Divergence, Hybrid.