Nightjar: overnight short strangle on Nifty and Sensex weekly options, inspired by a Tradetron back-test

Version 1.0. Closed 2026-10-21. Owner: Prashant Maheshwari. Status: no robust edge after costs; passed the one-time Nifty sealed check but failed the pre-registered Sensex replication, so no paper run was started.

1. The report that inspired it

A Tradetron back-test report (generated 2026-10-14) of "Sensex OTM1 Strangle Unhedged STBT": positional (sell today, buy tomorrow), weekly expiry, 6 entry "sets", 2024-09-09 to 2026-09-08, 497 sessions, ₹7.5 lakh capital. Headline: gross P&L ₹8.53 lakh (+113.7%, gross CAGR 46%), win rate 65% of days, gross Sharpe 3.36, max drawdown 14.2%. What the report itself shows underneath the headline: * It is gross. Costs in its own cost lab are ₹4.96 lakh (58% of gross) on 5,676 order fills; net P&L ₹3.56 lakh, net Sharpe 1.38, at an assumed 0.5% slippage. * Repairs destroyed the edge. Exit attribution: the planned exits earned +₹44.86 lakh gross over 1,535 round trips; the "repaired (re-strike)" exits lost ₹36.33 lakh over 898. * The call leg did everything. Call leg ₹7.61 lakh, put leg ₹0.23 lakh: a period of rising markets. * Volatility regime: losses when India VIX was elevated or high (25 days, -₹55k); day-of-week edge on Monday to Wednesday and none on Thursday or Friday. * Its own risk flags: skew -1.59, excess kurtosis 21, worst day -10.1%, nine losing trades in a row, and real index-option fills about 1.5% worse than the minute open (3.5% in the first 15 minutes), measured on 21,011 fills. * Entry and exit prices are the best of each minute; the strategy was run unhedged with a 1-minute price check.

2. What was tested here

The same idea on Nifty 50 (the data on hand: 464 sessions, 2024-12-03 to 2026-10-16, Dhan 1-minute ATM-relative weekly options, strikes ATM-2 to ATM+2) and, as a replication, on Sensex (fetched 2026-10-21). One short strangle (one lot each side) per night: sell at an entry minute, buy back at a chosen minute of the next session. Variants searched: 8 entry times (09:30 to 15:15), 8 exits (next-session 09:20 to 15:15, or same day 15:15), 15 strike structures (call ATM/OTM1/OTM2 with put ATM/OTM1/OTM2, and each leg alone), 5 stop levels (none to +200% of premium), 3 profit targets, 3 defined-risk structures (iron fly and condor with long wings), and 29 session filters (VIX level, weekday, sessions to expiry, implied-over-realised volatility, calm or falling VIX): about 287,000 configurations.

Fixed strikes across the overnight gap. Dhan's rolling series are ATM-relative, so the strike sold at 15:15 is not in the series the next morning about a quarter to a half of the time. Those prices come from a Black-76 model (forward from put-call parity at the ATM pair, ATM implied volatility, skew fitted on observed strikes). On strikes that ARE observed the model's median error is about 0% to 2% with a mean absolute error of 10% to 12%. About 40% to 60% of overnight exits use the model price, and they are the big-move nights, so the model matters.

Costs. Slippage 1.5% of the premium per fill, 3.5% in a session's first 15 minutes (the report's own measurement of live fills), brokerage ₹20 an order, exchange 0.03503% of turnover, SEBI 0.0001%, GST 18% on those, stamp 0.003% on buys, STT on the premium sold (0.10% to 31 March 2026, 0.15% from 1 April 2026, per the exchange's published table), lot 75 to 29 December 2025 and 65 after. "2x" and "3x" multiply every cost and the slippage.

Periods (fixed before any result). Development = first 50% of sessions (to 2025-11-xx), validation = next 25%, sealed = last 25% (from about 2026-04). One caveat: a diagnostic on a baseline configuration printed its sealed-period figures before the finalist was chosen; the finalist's sealed result has not been looked at.

3. Results on Nifty

The plain version has no stable edge. Sell the OTM1 strangle at 15:15 and buy it back next morning: +₹364 a night (t 1.9, 1x costs) in development, -₹572 in validation; at 2x costs -₹155 and -₹1,163. One night, 2 February 2026, lost ₹32,933 per lot pair.
Searched space. At 2x costs 38% of all configurations are positive in development but only 5.8% are positive in both development and validation (1.1% at 3x). The 40 best by development lose ₹1,634 a night on average in validation, and none of them is positive.
Development profits were market direction. Put-only configurations carry the development profits (rising market) and lose in validation (falling market); call-only configurations show the reverse. A sign-flip reality check shows the best development t-statistic (8.7) is far above what luck would produce in that period (95th percentile of the best of 226,800: 4.1), which is consistent with a persistent regime, not an edge that carries over.
Defined-risk wings fail on cost. Iron fly and condor structures lose ₹2,000 to ₹2,600 a night on average at 2x costs, in both periods.
Volatility timing fails. Selling only when implied volatility is more than 1.2 times the 10-day realised volatility is positive in development almost everywhere and negative in validation almost everywhere (2% of strangle configurations positive in both).
Late entries before expiry are the one cluster that survives. Strangles opened at 15:00 or 15:15 within two sessions of expiry and closed at next-day midday are positive in both periods even at 3x costs (18 of 7,938 strangle configurations). It is a small cluster chosen using both periods.

4. The cluster in detail (development and validation only)

Finalist = best of the strangle configurations by the lower of its development and validation results at 2x costs: sessions to expiry 1 or 2, sell at 15:15 one ATM call and one put two strikes out of the money (100 points), buy both back at 12:00 the next session, no stop and no target. 146 nights: development 99 (mean +₹1,080 at 2x, +₹663 at 3x, win rate 64%), validation 47 (+₹1,114 and +₹619, win rate 77%). Worst night -₹41,185 per lot pair on 2 February 2026 (about 37 average winning nights); 61% of its exits need the model price. If model-priced exits were 10% worse for the seller, validation at 2x falls to +₹462 and at 3x to -₹44; at 20% worse it is -₹191 at 2x.

5. Honest reading so far

The report's profit is mostly a gross, one-regime, one-leg result. After realistic costs and a validation period that included falling markets and a budget-weekend gap, the plain overnight strangle does not survive; the exception is a narrow pre-expiry cluster whose average gain (about ₹1,100 a night per lot pair, about 0.7% of the ₹1.5 lakh margin) is smaller than a single bad night and smaller than the uncertainty in the model prices.

6. What a paper run would be (only if the checks pass)

Notional paper trading of the finalist at 1 to 2 lots with the same costs, no real orders. Because paper trades use observed quotes, the paper run also measures the real fill price for the strike that drifts outside the data window, which this back-test cannot.

7. Pre-registered checks (written 2026-10-21 before they are run)

Sealed check (Nifty, run once). The finalist in section 4, exactly as specified, on the sealed period (the last 25% of sessions). Pass if the mean per night at 2x costs is above zero AND at 3x costs is above zero AND at 2x costs with model-priced exits 10% worse for the seller is above zero.
Sensex replication (run once, no tuning). The same structure translated to Sensex (sessions to expiry 1 or 2, sell at 15:15 the ATM call and the put two strikes out of the money (200 points), buy back at 12:00 the next session) over the whole Sensex sample. Pass if the mean per night at 2x costs is above zero in each half of the sample.
Tail rule. The worst night in the whole sample, multiplied by the lots to be traded, must be at most 5% of ₹15 lakh (₹75,000), which limits the paper run to 1 lot at the Feb-2026 loss. Paper trading starts only if 1, 2 and 3 all pass; otherwise the idea is closed and documented as a negative result.

8. Results of the pre-registered checks (run 2026-10-21)

Check 1, Nifty sealed period (2026-04-30 to 2026-10-16, 50 nights): PASS. Mean per night per lot pair +₹871 at 2x costs (74% of nights positive), +₹461 at 3x, +₹448 at 2x with model-priced exits 10% worse for the seller (at 20% worse: +₹25). Worst night -₹12,132. Over all 196 qualifying nights (development, validation and sealed): +₹1,035 a night at 2x costs, total +₹2.03 lakh for one lot pair, maximum drawdown ₹52,368 (3.5% of ₹15 lakh), worst night -₹41,185. A regression of the nightly result on the overnight market move gives a slope of -₹3,327 per 1% (the structure is short a call, so it loses when the market rises: nights with a rise of more than 0.5% averaged -₹4,465) and an intercept of +₹984 a night (t 2.9); the average overnight market move was about zero in each period (+0.06%, -0.17%, -0.02%), so the result is not a drift effect. Check 2, Sensex replication (214 nights, 2024-09 to 2026-10, no tuning): FAIL. Mean per night at 2x costs: +₹368 in the first half (106 nights), -₹331 in the second half (108 nights), +₹15 overall; at 3x costs -₹87 and -₹783; with model-priced exits 10% worse, negative in both halves at 2x. (The Sensex lot size is an assumption of 20; it scales the rupee amounts but not the signs.) The rule was "positive in each half at 2x costs". Check 3, tail rule: PASS for one lot only. Worst night -₹41,185 per lot pair is 2.7% of ₹15 lakh; two lots would be -₹82,370, above the 5% limit. Decision under the section 7 rules: no paper run. Checks 1 and 3 passed, check 2 failed.

9. Gross versus net, and what the report's headline leaves out

Selling the overnight strangle does earn a gross premium. Per night per lot pair, before any cost or slippage: Nifty plain OTM1 strangle ₹637 to ₹965, Sensex ₹726 to ₹949; the pre-expiry finalist ₹1,904 on Nifty and ₹922 on Sensex. After costs at the 1.5% slippage the report itself measured on live fills: Nifty plain strangle +₹107 to +₹441 at 1x and -₹422 to -₹84 at 2x; Sensex +₹185 to +₹412 at 1x and -₹355 to -₹124 at 2x. The report's +46% gross CAGR therefore does not survive its own execution evidence. Its own net figure (+₹3.56 lakh on ₹7.5 lakh over two years) assumes 0.5% slippage, a third of what it measured live.

10. What was learned

The overnight strangle is a real but thin premium: roughly the size of the costs. The direction of a net result on a given stretch is mostly decided by the volatility regime and the market's drift, not by tuning: development winners (put-selling in a rising market) lost in validation.
Of 287,000 configurations on each underlying, 5.7% (Nifty) and 0.2% (Sensex) were positive in both development and validation at 2x costs; the 40 best by development lost money in validation on both.
Defined-risk wings, stops, targets, volatility timing and weekday or VIX filters did not rescue it. Repair or re-strike exits, which the report uses, destroyed most of its profit in the report's own attribution.
One narrow cluster (sell within two sessions of expiry at 15:15, buy back next midday) survived development, validation and the Nifty sealed period, but not the Sensex replication, and about 60% of its exits rely on a modelled price. The honest label is "not established".

11. If this is revisited

A paper run that records real contract prices would remove the largest uncertainty (the model price for strikes outside the data window). paper_ovn.py is written for that and tested only in part (a token problem blocked the live test); it is not scheduled. It needs a valid Dhan token every evening, and it catches up on missed evenings.
More data: the Sensex first chunk (2024-09-30 to 2024-10-27, calls at the money) failed to download and is missing; a rerun is a one-line command once the token is valid.
Needed for any live use: real bid and ask quotes, margin from the broker, and a tail rule that is larger than the worst night in a short sample.

Research and education only. Hypothetical and back-tested results have inherent limitations and do not guarantee future results. Not investment advice; the publisher is not presented as a SEBI-registered investment adviser or research analyst.