Overnight-Premium in Indian Stocks

“Buy Today, Sell Tomorrow” is a widely repeated piece of Indian market folklore: buy near the close, hold overnight, sell at the open. The paper asks three questions in order. Is the underlying belief about overnight returns true in this market? If it is, does it survive the scrutiny a skeptical referee would apply? And does any increasingly sophisticated way of collecting it survive realistic 2026 Indian trading costs and — just as importantly — correct real-time implementability?

The primary contribution is the first: establishing that the close-to-open premium exists in this panel and decomposing how much of it is market beta and how much is cross-sectional ranking skill. The execution analysis is secondary and cautionary.

Universe. 189 NSE-listed, F&O-eligible stocks chosen because F&O eligibility identifies a liquid, currently investable universe. One-minute OHLCV data from 1 January 2020 to 31 December 2025 was obtained for 185; one recent listing was dropped for insufficient history. The working panel is 184 stocks and 274,263 stock-days.

Returns. Overnight = ln(open(t)/close(t−1)); intraday = ln(close(t)/open(t)), using the continuous session's first and last one-minute bars rather than auction prices, which one-minute data does not carry.

Corporate actions. Days with an absolute overnight return above 8% are cross-checked against NSE's corporate-actions record; 29 of 603 flags were confirmed actions and excluded.

Periods. In-sample 2020–2023 (182,634 stock-days), out-of-sample 2024–2025 (91,629), and a separate 2026 extension (1 January – 11 September 2026; 31,832 stock-days, 173 trading days) pulled after the model, hyperparameters and execution definitions were frozen.

Inference. The primary standard is two-way (stock × date) cluster-robust inference, because 184 stocks moving together on one date are not 184 independent observations. A naive per-observation HAC t-statistic is reported only as a supplementary figure. A calendar-time block bootstrap (2,000 resamples, 10-day blocks) and per-stock Wilcoxon tests are used as checks.

Market adjustment. A unit-beta adjustment subtracts NIFTY 50's own overnight return; a more precise version uses each stock's rolling 120-day beta from strictly trailing data. Ranking skill is tested by residualising the persistence signal against beta, liquidity and realised volatility jointly.

Selection rule. Stocks are ranked by a simple 60-day trailing mean of their own overnight return. A seven-feature LightGBM model and a ridge regression were tested against it on identical purged, embargoed walk-forward folds.

Costs. Fees are built from Zerodha's published 2026 rate card, separately for individual and non-individual accounts; market impact uses a square-root model calibrated two ways; bid-ask spread is estimated three ways.

Ex-ante versus ex-post. The paper separates a ranking formed before the fact (a predictive test) from any statistic that conditions on whether a stock is also selected the next day (descriptive only). Only the former is described as forward validation.

The premium is real. Pooled overnight return +17.97 bps/day; two-way clustered t = 9.48. In-sample +19.60 bps/day, out-of-sample +14.73; the intraday leg is negative in both (−11.75 and −10.96). Block-bootstrap 95% CI [12.95, 22.16] bps.

It is pervasive. All 184 stocks have a positive mean overnight return (2.50 to 36.56 bps/day, median 18.23), and Wilcoxon tests are significant in 182 of 184. It is positive and significant in every year from 2020 to 2025.

Beta versus ranking skill. A unit-beta adjustment reduces the premium to 6.93 bps/day and a rolling-beta adjustment to 9.13 — roughly a third to a half is market-beta exposure. Ranking skill survives much better: the persistence ranking's edge over random selection goes from 9.24 to 8.09 bps/day after neutralising beta, liquidity and volatility, an 87% retention.

Persistence beats machine learning. Mean information coefficient 0.1026 for persistence against 0.0821 for LightGBM on identical folds; persistence wins 4 of 5 folds. A ridge regression is slightly better than persistence; nonlinearity does not help.

Harvesting fails. Buying every eligible stock every night: −63.2% CAGR out-of-sample. Forced nightly liquidation of the top decile at ₹20,00,000: −0.5% CAGR (retail) and −9.5% (non-individual), with 0 of 30 cost-sensitivity combinations both net-positive and significant. The corrected Hold-Through discipline: −18.1% (retail) and −22.7% (non-individual) on the discovery window and −25.1% on 2026.

• Alternative constructions. Value-weighting gives 20.20 bps, winsorising 18.97 bps and excluding the least liquid quartile 17.58 bps — none moves the estimate by more than about 2 bps.
• Point-in-time correction. 61 stocks joined the F&O segment mid-sample. Using only genuinely eligible stock-days moves the headline from 17.97 to 16.95 bps/day (two-way clustered t = 10.05).
• Corporate-action window. Varying the matching window from 0 to 5 days moves the panel mean by at most 0.12 bps.
• Regimes. In the February–June 2020 COVID window the overnight mean is indistinguishable from zero. The large up-day/down-day asymmetry is essentially all market beta.
• Costs and scale. No market-impact assumption or spread estimator rescues the corrected Hold-Through result; every cost layer only deepens the loss.
• Internal audit. A five-track, author-conducted audit checked for look-ahead timing, compounding, statistical formulas, data integrity and spread formulas. It found the Hold-Through bug, a latent merge bug that never affected a reported number, and a small beta-estimator inconsistency, since corrected.

The paper is explicit about which results count as forward validation.

The one forward-validated result. A same-day top-minus-bottom-decile spread, ranked only on information available before the day's open, replicates on genuine 2026 data the model never saw: 11.13 bps/day (HAC t = 4.17), against 14.90 bps/day (t = 9.78) in the original window — a later-sample replication under a frozen specification.

A false positive, caught and corrected. A “Hold-Through” discipline first appeared to return +75.0% CAGR on the discovery window and +41.2% on 2026. A deliberately adversarial, referee-style review of its information timing found it decided whether to hold a position through the session using the next day's selection — which is not knowable until that day's close. Corrected, it is negative at every capital level, cost scenario and window tested.

Not a forward test. A bottom-decile intraday short overlay shows a positive point estimate (+2.4% CAGR discovery, +12.2% 2026) but is not distinguishable from zero (HAC t = 0.40 and 0.83; bootstrap intervals span zero). Its scripts were written after the 2026 data existed, so its 2026 figure is in-sample relative to when the code was written.

The folklore is right about where the return lives: essentially all of this panel's long-run positive return is overnight-sourced, and the overnight return is real by every dependence-robust standard applied. It is wrong that the return is straightforwardly collectable. The paper's most promising harvesting method failed a more basic test — correct real-time implementability — before transaction costs became the deciding factor at all.

The methodological lesson is the paper's second contribution: a stress-test battery passed on the wrong mechanic proves nothing about whether that mechanic can be traded in real time. Catching that requires adversarial scrutiny of exactly what information is available at every decision point.

This is research, not investment advice.

• One market, one panel. Results are specific to 184 stocks that were F&O-eligible when the panel was built — not a claim about Indian equities generally.
• Partially point-in-time. The eligibility correction fixes entry-date bias for 61 stocks only; former constituents, four unretrieved symbols, and delistings or suspensions are not reconstructed.
• Short forward window. The 2026 extension covers roughly nine months.
• Costs are modelled. Fee, impact and spread estimates are proxies, not exchange-verified fills.
• Single risk factor. Risk adjustment uses NIFTY beta only, not size, value, momentum or quality factors.
• Prices are continuous-session bars, not auction prints; magnitudes move by about 2 bps across 1-, 5- and 15-minute definitions.
• No multiple-testing correction. Many specifications were explored and no formal correction is applied to individual results.
• Not externally validated. The audit was internal. Public code and data, a data dictionary, checksums, a published environment and automated leakage tests do not yet exist.

In this panel, close-to-open returns are positive and broadly distributed, and the premium is real and statistically overwhelming under a conservative two-way clustered standard. A genuinely ex-ante ranking spread replicates on previously unobserved 2026 data. But every harvesting method tested — blanket capture, forced liquidation and the corrected Hold-Through discipline, under any cost assumption, capital level or parameter choice — fails to deliver a return that is both economically material and statistically robust, and the one component with a positive point estimate is not distinguishable from zero. A specific, initially convincing look-ahead bug very nearly produced the opposite conclusion; how it was caught is reported in full because it may be as useful to others as the result itself.