Added after Satya's 19:55 IST (27-Oct) question: “on expiries opposite to trend from around 20 premium there is 2x or 3x jumps”. Design pre-registered at 20:05 IST in PREREG.md (EXPIRY-CHEAP-COUNTER) before any number was computed. Real 1m bars of the expiring contract:

Spot 15:15–15:29 is treated as missing. Settlement uses the official close. Option buyer only.

Verdict. The jumps are real, but they are not a counter-trend effect and they are not tradable.

How often: after the first counter-trend touch of 15–25, the option later printed ≥2× on {fn.hit2x:.0f}% of signals (Nifty) and {fs.hit2x:.0f}% (Sensex). It printed ≥3× on {fn.hit3x:.0f}% / {fs.hit3x:.0f}% and ≥5× on {fn.hit5x:.0f}% / {fs.hit5x:.0f}%. The median peak came {fn.med_min_to_peak:.0f} / {fs.med_min_to_peak:.0f} minutes after the signal.
Not counter-trend specific: a ~20 option picked at a random minute does about the same (2×: {fr('NIFTY','random minute 15–25 (30 reps)').hit2x:.0f}% / {fr('SENSEX','random minute 15–25 (30 reps)').hit2x:.0f}%), and so does the with-trend option ({fr('NIFTY','with-trend 15–25').hit2x:.0f}% / {fr('SENSEX','with-trend 15–25').hit2x:.0f}%).
Most of them die: {fn.near_zero_settle:.0f}% (Nifty) and {fs.near_zero_settle:.0f}% (Sensex) settle at ≈0.
As a buyer trade: all 6 pre-registered test rows (2 indices × 3 exits) lose money in 2026 ({EP[EP.split=='test']['mean'].min():+.1f} to {EP[EP.split=='test']['mean'].max():+.1f} pts per trade), and all 6 train rows lose too. 0 of 6 count. {npos} of {len(st)} sensitivity test rows are positive (bands, Sensex scaled band, DECAY, time windows), and none of them is significant. Picking a regime on train does not rescue it.

EC1. Definitions (pre-registered)

EC2. Frequency facts

“Reached N×” = a later 1m high ≥ N × the signal-minute close, at any time up to 15:29. Measured from the fill (next open +slippage), ≥2× drops slightly. The median signal minute is 09:36, because a 15–25 option on the counter side usually already exists right after the trend check starts. Sensex points are about 3.3× Nifty points, so a raw 15–25 Sensex option is much further OTM; the “scaled band” row fixes that.

Share of expiry days where a counter-trend first signal (either side) reached 2× / 3× / 5×, by year:

EC3. Tradable test (primary, pre-registered; BH over the 6 test rows)

Net option points per unit (Sensex rows are in Sensex points; divide by ≈3.3 for Nifty-equivalent). A row counts only if the test mean > 0, p < 0.05, q < 0.10, it beats random (≥90th percentile) and the with-trend control, the train mean > 0, and the mean without the top 3 is > 0.

Why “2×–3× happens 20–40% of the time” still loses money: the median trade loses most of the premium. Exit (a) wins only 1 trade in 4 to 6, and most of the rest expire near zero, so the 2× and 3× winners don't pay for them. Hold-to-settlement has rare huge winners (best +{EP[EP.exit=='hold to settlement'].best.max():.0f} pts) but a win rate of 3–14%, and it is negative without the top 3. The p < 0.05 rows are significantly negative.

EC4. Sensitivities (test 2026)

EC5. Regime chosen on train only, then tested on 2026

The best train buckets were mostly still negative on train. The two that were positive on train both failed in 2026: Nifty DMA-trend up, hold to settlement (+8.2 → −7.8), and Sensex ER20 mid, trail (+0.0 → −9.1). The only positive test bucket (Sensex hold, +9.3, n=19) had a train mean of −4.6, wins on 1 of 19 trades, and is −23.5 without its top 3.

EC6. Fill realism

Every signal and fill bar had traded volume (median about 3.0M Nifty and 0.37M Sensex units at the fill minute), so the prints are real trades, not stale quotes. Bid/ask spreads are unknown (no history). On a 20-premium option a 0.5–1 pt spread is 2.5–5% per side, so real results would be somewhat worse than shown.

Expiry day: do cheap (~20) counter-trend options jump 2×–3×?