Webull Edge Lab — Does Stated AI Confidence Survive the Options Rail?NULL-HYPOTHESIS STUDY

Webull Edge Lab · Phase 1 (paper mode) · 2026-06-17 · ~8 min read · N = 0 closed option-trades at publication · pre-registered protocol

1. Lineage — the binary edge that died of bad calibration

This lab descends from the binary Edge Lab (edge_lab_service.py + edge_lab.db): the dashboard that logged a binary observation, graded it after a fixed forecast window, and bucketed the results by the AI's stated confidence band [50–60, 60–70, …, 90–100]. That lab never shipped a standalone whitepaper because its finding was a number, not a method — the number that the Monte-Carlo F7 lab later turned into the project's first headline conclusion. The closest methodological ancestor in writing is the TRAIL Lab methodology whitepaper, which fixed the Wilson-lower-bound discipline this lab inherits, and the calibration-honesty section of the hybrid_v1 whitepaper.

The binary Edge Lab's verdict is headline finding #1 of the pivot brief, quoted verbatim:

"Confidence calibration is broken in our direction. The AI's stated confidence consistently overstates p_hat by 5–15 pts across the upper buckets (70–90). Wilson lower-bound p_lower collapses below break-even once realistic friction (spread + slippage + missed fills) is applied." — WEBULL_PIVOT_STUB_PROMPT.md §1.4, finding 1

Two further findings frame why a new lab is warranted rather than a simple verdict of "the model is overconfident, stop." Finding #3 — "b ≈ 1 is a lie our binary UI was telling. When you derive real b from |target−entry| / |entry−stop|, most setups are 0.6–0.9 — meaning we needed p ≥ 0.55 just to be Kelly-neutral" — says the binary contract's payoff geometry was working against us before calibration even entered the picture. Finding #5 — the project's adopted conclusion — says the response is not to keep predicting the coin flip but to "use the same predictive engines to select an options structure whose payoff is shaped to match the predicted regime."

The Webull Edge Lab is the most literal test of that pivot. It changes one thing versus its binary ancestor: the confidence number no longer sizes a one-touch binary bet — it picks a defined-risk vertical spread with an explicit, geometry-derived payoff ratio b, and sizes it off p_lower. The question is brutally simple: does shaping the payoff repair the calibration gap, or does the gap travel with the model regardless of the instrument we wrap around it?

2. Method

2.1 The regime classifier — which forecast engine feeds it

The Edge Lab's classifier is the thinnest of the five labs by design: it is the raw Candlevision / ai_intelligence_engine.py verdict — a {direction, confidence} pair extracted from the chart frame by vision_analyst.py / vision_extractor.py, with no exhaustion, divergence, or ensemble fold layered on top. That deliberate thinness is the control: the Exhaustion, Divergence, TRAIL and Hybrid labs each add structure on top of this same primitive, so the Edge Lab measures what the bare stated confidence is worth on the options rail before anyone improves it. The classifier emits the §2.2 regime block: {direction, strength, iv_rank, vol_regime, exhaustion:null, divergence:null}.

2.2 The strategy selector — the candidate set

The selector (options_strategy_mapper.select_strategy(regime, ctx)) is restricted, for this lab, to the directional defined-risk subset of the Periodic Table (WEBULL_PIVOT §3) — the elements whose payoff is a clean monotone function of being right about direction:

With defined_risk_only=true (the binding default, WEBULL_PIVOT §6.2) the selector never returns a naked or undefined-risk leg. Every candidate carries the §2.2 output: primary, alternates, rationale, legs, max_loss, max_gain, breakeven. The vertical's payoff ratio is the real b — b = max_gain / max_loss — read straight off the leg geometry, which is exactly the honesty finding #3 demanded.

2.3 The calibration pipeline — reuse, do not reinvent

Every closed paper option-trade is persisted on the existing paper_trade_service.py trade row (no parallel DB — WEBULL_PIVOT §6.3), extended with the workboard §2.4 columns: webull_strategy_id, iv_rank_entry, iv_rank_exit, theta_paid_usd, vega_at_entry, defined_risk, max_loss_usd, max_gain_usd, legs_json, broker_client_order_id. The /api/cv/paper/calibration endpoint gains a ?strategy_id= filter so the evaluation buckets per options structure instead of averaging an edge in one element against a drag in another — the whole point being to discover which element of the periodic table, if any, carries a calibrated edge. The Wilson 95% lower bound (Wilson 1927) is computed per bucket exactly as in the binary lab:

2.4 The ergodic Monte-Carlo evaluation (F7)

Sizing and growth are evaluated by the F7 machinery (Wilson lower-bound Kelly, ergodic Monte Carlo). The break-even win rate for a defined-risk spread with payoff ratio b is p* = 1/(1+b); the growth-optimal stake is fractional Kelly off p_lower, never p_hat ("Kelly off a thin sample is how accounts blow up" — F7 README). The simulation is additive (bankroll += payoff), reporting the time-average growth rate Peters (2019) distinguishes from ensemble EV — the same evaluation that rendered the flat-to-negative verdict on the binary phase. A positive ensemble EV with a ≤ 0 time-average growth rate is, per finding #2, still a fail.

3. Null hypothesis & stop criterion

We reject H₀ only if, over a sample of ≥ 100 closed paper option-trades in at least one webull_strategy_id bucket, the Wilson lower-bound p_lower for that bucket exceeds the geometry-derived break-even threshold:

and the bucket's stated-confidence midpoint sits within 5 points of that same p_lower (the calibration-gap clause). Both conditions must hold simultaneously: an edge that exists only because the AI got lucky in one band, while its confidence remains miscalibrated, does not reject H₀ — it just means the geometry, not the prediction, paid. The ?strategy_id= bucketing is what lets us tell those two apart.

4. What a failure to reject tells us

Per the lab's stance — inherited directly from the TRAIL Phase 1.5 study, where a 36% directional hit rate was reported as a finding, not buried — a null-not-rejected result is a successful run. If after hundreds of closed trades no bucket clears p* , the lab has produced a real, expensive-to-fake piece of knowledge: wrapping a defined-risk spread around a miscalibrated confidence number does not, by itself, manufacture an edge. That is worth knowing precisely because the pivot's central hope is that payoff-shaping rescues the prediction layer. Three concrete things a sustained failure teaches:

• The gap is in the model, not the contract. If the same 5–15pt overstatement appears across verticals as it did across binaries, the defect is the perception/confidence layer (vision_analyst.py), and no amount of structure selection will fix it. That redirects effort to calibration of the engine, not the selector.
• Defined-risk capping worked. A failure to reject with the bankroll intact (no ruin path) confirms the defined_risk_only guard and floor sizing did their job — we learned the model has no edge without paying to learn it. That is the entire reason the lab exists.
• It isolates the per-structure question for the other four labs. The Edge Lab is the control; if it can't reject H₀ but the Exhaustion or Divergence lab can, the edge lives in the regime structure those labs add, not in raw confidence.

5. Phase 2 — must-do / must-not

Must-do:

• Reliability diagram per strategy bucket. Of all entries the AI called 0.80, what fraction of bull_call_spread trades actually won? Below-diagonal means dampen tilt amplitude, exactly as TRAIL Phase 1.5 §10 prescribed for the binary case.
• Per-strategy theta budget. A vertical bleeds theta differently than the binary contract had no concept of; track theta_paid_usd against realized R per bucket so a "win rate" isn't quietly funded by carry.
• Binary-payout variant. If the lab eventually routes Pocket-Option-style binaries, add an asset class paying +0.8×stake on a correct sign and −1.0×stake on a wrong one — the actual contract the original "80%" intuition came from — and re-derive the break-even there.

Must-not:

• Tune the selector on n < 100. There is no signal in a thin bucket, only setup. The binary phase's blow-up risk was Kelly off a thin sample; do not repeat it on options.
• Deploy capital. The lab is a measurement instrument. Real money follows only a friction-honest, sustained p_lower > p* across many sessions and structures.
• Flip defined_risk_only=false while any funded bucket has p_lower < p*. Undefined-risk legs on a miscalibrated edge is the precise failure the pivot was built to escape.

References

1. 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 defined-risk structures the selector emits; alternative to Black & Scholes 1973.)
2. Black, F., & Scholes, M. (1973). The Pricing of Options and Corporate Liabilities. Journal of Political Economy, 81(3), 637–654.
3. Kelly, J. L. (1956). A New Interpretation of Information Rate. Bell System Technical Journal, 35(4), 917–926. (Payoff ratio b and growth-optimal sizing; break-even p* = 1/(1+b).)
4. Wilson, E. B. (1927). Probable Inference, the Law of Succession, and Statistical Inference. Journal of the American Statistical Association, 22(158), 209–212. (Per-bucket lower bound used for sizing.)
5. Peters, O. (2019). The Ergodicity Problem in Economics. Nature Physics, 15, 1216–1221. (Time-average vs ensemble growth; the F7 evaluation objective.)
6. Taleb, N. N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House. (Fragility of edges estimated from thin samples; optional.)
7. Lineage: TRAIL Lab — Trailing Reality-Adjusted Inference (methodology) and the calibration-honesty section of Binary Exhaustion Lab — hybrid_v1; binary Edge Lab findings: WEBULL_PIVOT_STUB_PROMPT.md §1.4.

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