Verdict
No detectable skill at this sample size
Best option-feature variant gained +0.057% per session over the better baseline (t = 0.58, p = 0.26); reality-check p = 0.37 across 8 variants.
- Development sessions 2025-07-10 to 2026-05-25 (214 sessions); the last 100 sessions (from 2026-05-26) were never scored.
- Power statement fixed beforehand: the smallest detectable mean difference is about 0.14% per session, which is large; a smaller real effect cannot be excluded.
- Option features did not help LightGBM (worse than its price-only twin); ridge with option features was marginally best but indistinguishable from noise.
- The option history (464 sessions, no bid/ask) is too short for a decisive test; a longer history would be needed to answer this properly.
- Implied-vol features had implausible values near expiry; ridge winsorised them (Amendment 1, written before any model was fitted).
- Cost hurdle under the standard Dhan model (lot 65, a typical 2-leg debit spread, 4 fills): about ₹135 charges plus ₹166 slippage at 0.5% per fill = ₹301 per round trip (₹632 at 1.5%). The best variant's mean edge of +0.037% a session on a delta of about 0.25 is roughly ₹143 gross per trade, so it would lose about ₹158 per trade after costs. The skill gate failed before costs; costs make it worse.
Promotion gates
No gates published for this strategy.
Explain the strategy
We asked whether clues from the options market (how expensive options are, how lopsided, how many bets on falls versus rises) could help guess tomorrow's Nifty direction. In about 214 test days, they did not help in a way we could detect.
Report
All 8 pre-registered variants (skill only, before trading costs)
| Variant (model · features · rule) | Test sessions | Better baseline | Strategy mean per session | Strategy minus baseline | t-stat | Bootstrap p |
|---|---|---|---|---|---|---|
| ridge · A1_price+option · sign | 214 | momentum | +0.037% | +0.057% | 0.58 | 0.257 |
| ridge · A0_price · confident | 214 | momentum | +0.020% | +0.040% | 0.45 | 0.27 |
| ridge · A1_price+option · confident | 214 | momentum | +0.021% | +0.042% | 0.45 | 0.277 |
| lightgbm · A0_price · sign | 214 | momentum | +0.005% | +0.025% | 0.26 | 0.308 |
| ridge · A0_price · sign | 214 | momentum | -0.009% | +0.011% | 0.12 | 0.416 |
| lightgbm · A0_price · confident | 214 | momentum | -0.020% | +0.001% | 0.01 | 0.413 |
| lightgbm · A1_price+option · confident | 214 | momentum | -0.056% | -0.036% | -0.39 | 0.603 |
| lightgbm · A1_price+option · sign | 214 | momentum | -0.096% | -0.075% | -0.77 | 0.699 |
Rules and parameters
Rules
- Sessions: 464 Nightjar option sessions, 2024-12-03 to 2026-10-16; decision at 15:14, next-session 15:14 label.
- Option features: ATM IV, put and call skew, put/call open-interest ratio, straddle over spot, IV minus VIX, IV over realised vol, share of option volume in the last hour, sessions to expiry, day-over-day changes (missing across expiry). Price features: Gate 1 features at the previous close plus today's move to 15:14.
- Models: ridge logistic (winsorised, imputed) and small LightGBM; with and without option features; rules sign and confident; 8 variants.
- Walk-forward: first fit on sessions 0 to 149, retrain every 20 sessions, purge 1 plus embargo 2; development test sessions 150 to 363.
- Pass rule: best option variant beats the better of long and momentum with bootstrap p < 0.05 and reality-check p < 0.10, beats its price-only twin, mean above zero.
Parameters
- pre-registration
- docs/PREREG_STUDY_A.md (+ Amendment 1)
- variants
- 8
- test sessions
- 214
- detectable effect
- about 0.14% per session
- sealed
- last 100 sessions, unopened
Instrument
Nifty spot direction, features from weekly option minute data (2024-12 to 2026-10) · capital ₹2,000,000
Known limitations
- Only 464 sessions of option data exist and they have no bid/ask quotes.
- No trading costs applied at this gate (skill only).
- IV values near expiry are noisy.
Logs and raw data
Downloads appear once a data window is loaded.
Files are plain CSV so results can be re-computed independently. Trade dates are IST sessions; returns are net of modelled costs and slippage.