FnO Ban Selling Pressure

Does entry into the NSE F&O ban create selling pressure in the cash market that an outside trader could supply liquidity to — net of realistic costs, and under conditions designed to make a false positive hard to produce?

A ban turns a stock's derivatives market one-way: holders may close positions but no one may open new ones. Bans usually follow a crowded build-up of leveraged positions, so a ban entry can force a wave of exits on its first session. If those exits spill into the cash market, the stock should face temporary selling pressure. The ban list for each session is published the evening before, so every event is known before the open and is not chosen by the researcher.

Events. A ban entry is a stock on the NSE ban list for session D but not for the previous session. Entries are counted only when the ban files for both sessions exist, so a missing file cannot create a false entry. An event is kept if the stock is an F&O member on D, its first trade is within 20% of the previous close (a guard against split and bonus days), and entry and exit prices exist.

Sample. 434 development events (2020–22), 469 validation events (2023–24), 91 events in the 2025 holdout and 33 in January–October 2026 — 1,027 in all.

Sources. The NSE F&O ban-list archive (2020–2026); split- and bonus-adjusted one-minute bars for about 264 F&O stocks (2020–2025); one-minute bars for the 2026 event stocks; an equal-weight F&O panel and NIFTY 50 for market adjustment; NSE daily open-interest and MWPL files for utilisation and crowding; and NSE F&O membership by month.

Coverage limits. The 2026 price data cover only about 30 stocks, and the 33 events of 2026 fall on 30 dates in 10 stocks, so 2026 results are indicative.

The design fixes the trade, the periods, the costs and the pass/fail gates before any validation or holdout data is seen, and commits each document to version control with a timestamp before the corresponding results.

The trade. On each ban-entry day, short the stock at the open of the 09:16 one-minute bar and cover at the open of the 15:15 bar (the 15:14 close from 3 August 2026), on ₹5 lakh notional. Beta-adjusted returns replace the stock's session return with its return minus β times the equal-weight F&O market's session return, with β estimated from the 120 prior sessions.

Costs. Brokerage, securities transaction tax, exchange charges, the SEBI fee, GST and stamp duty — about 4.5 bps in all — plus a spread charge of 2.5 times a per-stock-day half-spread estimate (the average of the Corwin–Schultz and Abdi–Ranaldo estimators). Mean round-trip cost is 20.8 bps in 2020–22, 11.0 bps in 2023–24 and 8.6 bps in 2025.

Periods and gates. Development 2020–22, validation 2023–24, a 2025 holdout opened once after validation passed, and January–October 2026 as further evidence. A period passes with a positive mean and a 95% interval above zero; validation also had to survive a red team: dropping the 10 best events, a positive result in each year, beating a placebo, and staying positive under higher costs. Because development overlaps an earlier descriptive study that suggested the effect, only validation, the 2025 holdout and 2026 count as confirmation.

Inference. Primary inference uses two-way clustering by day and by stock, with day-only, stock-only and a wild-cluster bootstrap as checks. The paper shows why it does not use the "pigeonhole" two-way bootstrap: when clustering is weak it inflates the standard deviation by about √3.

Look-ahead checks. A truncation test deletes all data after the decision time for sampled events and recomputes every input; it found no mismatches.

Before 2025 the effect is large and consistent. Net of costs, with two-way clustered 95% intervals:

• Development 2020–22 (434 events): +54.9 bps raw [+14.8, +94.9]; +30.6 bps beta-adjusted [−0.8, +62.0], p = 0.056.
• Validation 2023–24 (469 events): +52.4 bps raw [+22.9, +81.9]; +40.0 bps beta-adjusted [+16.7, +63.3], p = 0.001.
• Pooled 2020–24 (903 events): +53.6 bps raw [+29.5, +77.6]; +35.6 bps beta-adjusted [+16.9, +54.2], p < 0.001.

In 2025 and 2026 it is gone. The 2025 holdout earns +2.7 bps (day-clustered CI −50.4 to +61.6); January–October 2026 earns +10.9 bps, in line with its placebo.

Where the return comes from. In development about 21 bps of the stock's underperformance occurs in the opening gap, before a 09:16 entry is possible; the tradable part is the session decline of −52 bps relative to the market. The short captures the session, not the gap.

Placebo. The same short on randomly drawn non-ban sessions of the same stocks in validation earns −2.7 bps, against +50.5 bps on ban-entry days.

Breadth. Before 2025 the profit was broad, with medians above the means and win rates near two thirds. In 2025 the 10 best trades earned 19 times the year's total; the other 81 lost money in aggregate.

• Inference choice. Day-only, stock-only and two-way clustering agree closely, showing clustering is weak in this sample. The pigeonhole bootstrap's intervals are about 1.7 times wider — the inflation the paper predicts — which is why an earlier version of the analysis judged development and validation marginal.
• Red team (validation, run before the holdout). Dropping the 10 best events leaves +33.8 bps; the effect is positive in each year (2023: +45 bps; 2024: +54 bps) and remains positive under higher costs.
• Crowding does not explain the size. Three pre-registered tercile splits — 5-day run-up, MWPL utilisation on the day before, and 5-day change in utilisation — show no significant difference after Holm's correction, and two point estimates have the wrong sign. Every tercile shows a session decline of 28 to 92 bps.
• Threshold design. A regression-discontinuity comparison around the 95% utilisation threshold finds no detectable jump, with wide intervals.
• Ban exit. A long trade on the first session after a ban ends earned about −1 bp and was stopped at development.

The 2025 holdout was opened once, after validation passed, and its ban lists were downloaded only then. It earns +2.7 bps (CI −50.4 to +61.6), below the 95th percentile of its placebo distribution; dropping its 10 best events gives −52 bps. Beta-adjusted, January–May 2025 was already +2.8 bps — months before SEBI's revised MWPL rules took effect on 1 October 2025. January–October 2026 (+10.9 bps raw) is in line with its placebo.

Ban entries also fell in steps. A Poisson model with BIC-selected breaks finds four regimes; in entries per month the main steps were November 2024 (about 25 to about 9) and October 2025 (to about 3). The November 2024 fall began before SEBI's index-derivative measures of 20 November 2024, which applied to index derivatives only. Whether the fall came from stocks that stayed in the F&O segment or from stocks that left it depends on the window used.

At the 2025–26 rate of about 3.4 entries a month, a forward test with 80% power to detect +50 bps would take roughly seven years.

The evidence fits a simple account: before 2025, ban entry forced one-sided unwinding that the cash market absorbed at a price, and outside liquidity providers were paid about 30–50 bps for supplying it within the session. Three features support reading this as price pressure rather than information: the event is mechanical and known before the open, the placebo on the same stocks earns nothing, and the decline does not scale with how crowded the stock was.

Why the effect ended is less clear. Plausible explanations include fewer and smaller leveraged positions in ban-prone stocks, changes in the F&O universe, and faster anticipation of entries by other traders once utilisation data became widely tracked. The paper cannot separate these with the data at hand.

Two lessons extend beyond this setting: inference choices matter — a plausible-looking two-way bootstrap nearly led to calling a reliable effect marginal — and rule-driven edges can be real, validated and still temporary. A pre-registered holdout is what distinguishes an edge that ended from one that never existed.

• Discovery overlap. Development overlaps the earlier descriptive study that suggested the effect, so only validation, 2025 and 2026 count as confirmation; development alone is borderline under the primary inference.
• Uneven returns. Development leaned on 2021, and validation quarters ranged from −41 to +114 bps. Event returns have a standard deviation near 300 bps.
• Costs are estimated. Spreads come from one-minute bar estimators, not quotes; the conservative 2.5× half-spread charge partly protects against underestimation.
• Coverage. The one-minute panel covers about 264 F&O stocks; events outside it are not priced, and 2026 covers only about 30 stocks.
• Mechanism unidentified. Who unwinds positions at ban entry is not observed, and the pre-registered crowding measures do not explain the size of the decline.
• Capacity. Results assume ₹5 lakh per trade; at larger size, market impact would reduce returns.

Entry into the F&O ban produced economically large, statistically reliable selling pressure in Indian single stocks from 2020 to 2024: an intraday cash short earned about +50 bps per trade after costs, +31 to +40 bps after beta adjustment, in pre-registered development and validation periods, while a placebo on the same stocks earned nothing. The pressure was broad across levels of crowding. It then disappeared: the 2025 holdout and 2026 show no effect, the decline predates SEBI's revised MWPL framework, and ban entries became too rare to study further at useful power.

This is research, not investment advice.