Every trading night at 23:00 ET, the GammaRips pipeline scans roughly 3,500 optionable US stocks, filters to the 100 most liquid names (3M+ session shares, 25+ listed strikes), and selects one out-of-the-money call per bullish underlying.
For months, we asked a blunt question in Catalyst Lab: Can a supervised machine-learning model (XGBoost + Gemini 3072-d news embeddings) beat a simple baseline when ranking overnight options flow out-of-sample?
To answer it without look-ahead bias, we ran a strict purge-and-embargo walk-forward evaluation across N = 5,575 enriched option signals (74 scan dates, 2026-04-10 to 2026-08-07, with the post-August holdout sealed) joined to realized minute-bar option outcomes in BigQuery (enriched_option_outcomes).
Here is what we killed, what survived, and how those findings now govern the Combo A dataset served over the GammaRips MCP Server.
1. What We Killed: Black-Box Composite Scores & News Embeddings (M1 and M2)
We tested three models using an expanding walk-forward split (minimum 20 training days, 5-day purge embargo to prevent 3-day excursion overlap, and 1,000x day-block bootstrap confidence intervals):
B0(Baseline):overnight_scorealoneM1(Tabular XGBoost): 60+ point-in-time quantitative, flow, technical, and regime featuresM2(Tabular + Text XGBoost):M1plus 16-component PCA over 3,072-dimensionalgemini-embedding-001grounded catalyst embeddings
| Cohort | Label | B0 AUC | M1 AUC | M2 AUC | M2 - M1 Text Lift (95% Day-Bootstrap CI) |
|---|---|---|---|---|---|
All Dev (N = 5,575) | y_touch | 0.4939 | 0.4765 | 0.4711 | -0.0054 ([-0.0148, +0.0028]) |
All Dev (N = 5,575) | y_close | 0.4818 | 0.4544 | 0.4601 | +0.0057 ([-0.0070, +0.0166]) |
Bullish Only (N = 4,045) | y_touch | 0.5124 | 0.4947 | 0.4994 | +0.0047 ([-0.0077, +0.0184]) |
Bullish Only (N = 4,045) | y_close | 0.4926 | 0.4927 | 0.5103 | +0.0176 ([-0.0020, +0.0346]) |
Verdict — Killed: Neither tabular XGBoost (M1) nor dense news embeddings (M2) produced a statistically significant out-of-sample AUC lift over 0.50 across all 74 scan dates (95% CI spans zero on every split). Shipping a black-box ML "pick" or composite probability score over noisy overnight options flow is curve-fitting.
2. What Survived: Realized-to-Implied Volatility (rv_iv_ratio)
When we decomposed feature gain across every walk-forward fold on the bullish cohort (N = 4,045), one continuous primitive ranked #1 on both y_touch and y_close:
rv_iv_ratio = (ATR_14 / Underlying_Price * sqrt(252)) / Implied_Volatility
| Rank | Feature (y_touch) | Gain Importance | Feature (y_close) | Gain Importance |
|---|---|---|---|---|
| 1 | rv_iv_ratio | 0.0390 | rv_iv_ratio | 0.0307 |
| 2 | reversal_probability | 0.0314 | expected_liquidity_MISSING | 0.0287 |
| 3 | industry_OTHER | 0.0295 | industry_MISSING | 0.0276 |
| 4 | high_52w_rel_price | 0.0292 | log_recommended_iv | 0.0268 |
| 5 | sma_200_rel_price | 0.0275 | vix3m_at_enrich | 0.0267 |
Why does rv_iv_ratio >= 0.85 matter? Unusual options activity often coincides with already-bid implied volatility. When an underlying's 14-day ATR annualized realized volatility exceeds or nearly matches the option's implied volatility (rv_iv_ratio >= 0.85), you are buying gamma that the options market has not yet overpriced relative to the stock's actual daily range.
3. Contract Microstructure & Convexity: The Combo A Specification
Next, we audited how contract-level microstructure gates affect entry fill speed (<= 5 minutes from the 10:00 ET decision mark) and 3-day Maximum Favorable Excursion (MFE) across all 4,045 bullish contracts:
| Filter Rule | Contracts (N) | Avg / Day | Fast Fill <= 5m (%) | 3d MFE >= +40% (%) | 3d MFE >= +80% (%) | Median 3d MFE (%) |
|---|---|---|---|---|---|---|
| 0. Unfiltered Bullish Pool | 4,045 | 54.7 | 30.3% | 27.6% | 13.7% | +21.9% |
1. Mid >= $1.00 | 3,393 | 45.9 | 29.1% | 26.2% | 12.9% | +20.8% |
2. OI >= 500 | 1,751 | 23.7 | 50.3% | 38.3% | 18.7% | +27.0% |
3. OI >= 1,000 | 1,254 | 16.9 | 54.5% | 39.6% | 18.5% | +28.7% |
4. Delta in [0.25, 0.45] | 2,857 | 38.6 | 33.8% | 31.1% | 15.3% | +23.7% |
5. RV/IV Ratio >= 0.85 | 3,852 | 52.1 | 30.8% | 28.1% | 14.0% | +22.0% |
6. Combo A (Mid >= $1.00, OI >= 500, Delta 0.22–0.48, DTE 10–45, RV/IV >= 0.85) | 996 | 13.5 | 50.1% | 35.4% | 16.7% | +25.2% |
7. Combo B (Mid >= $1.50, OI >= 1,000, Delta 0.25–0.45, DTE 14–45, RV/IV >= 0.90) | 485 | 6.6 | 56.5% | 35.3% | 15.3% | +25.4% |
Four physical rules explain the jump from the raw pool to Combo A:
- Resting Book Depth (
OI >= 500, log-saturated to5,000): Requiring at least 500 contracts of resting open interest lifts 5-minute fast execution fills from30.3%to50.1%(56.5%atOI >= 1,000). - Escaping Sub-Dollar Tick Friction (
Mid >= $1.00): Below$1.00, exchange minimum tick increments ($0.05on non-penny classes) impose a 5–10% structural bid/ask tax on entry and exit. - The Gamma/Theta Sweet Spot (
Delta 0.22–0.48,DTE 10–45): Below0.22delta, options require tail-event moves to overcome implied volatility crush; above0.48delta, capital efficiency drops toward linear stock exposure. PairingDelta 0.22–0.48withDTE 10–45lifts 3-dayMFE >= +40%from27.6%to35.4%. - Underpriced Volatility (
rv_iv_ratio >= 0.85): Ensures the underlying's realized ATR range supports the option's implied volatility mark.
4. How to Query the Combo A Dataset in Your AI Agent
We rebuilt the GammaRips overnight scanner and MCP Server (https://mcp.gammarips.com) around these exact Combo A floors. There is no pick endpoint: your AI agent reads the dataset, verifies live 10-minute intraday liquidity snapshots (get_liquidity) and earnings windows (get_signal(view="earnings")), and builds the trade plan with you from historical 3-day excursion distributions (query_outcomes).
Connect the Free 3-Contract Preview in 10 seconds (no card, no API key):
claude mcp add --transport http gammarips https://mcp.gammarips.com/mcp
Or unlock the Full Daily Combo A Dataset (all 35–50 daily contracts, 10-minute intraday liquidity snapshots, 3-day MFE/MAE outcome surfaces, and minute-bar contract replay) on GammaRips Pro ($29/mo founding price, 30-day free trial).
Educational market data and historical walk-forward research only. Not investment advice. Options trading involves substantial risk of loss.