Best in-sample correlation
0.290
"GIFT 06:30 open vs prior ref close" — but see below
Does GIFT NIFTY predict the Nifty opening?
GIFT NIFTY trades roughly 21 hours a day, hours before the NSE bell. This measures how much of the NIFTY 50 opening gap at 09:15 IST is actually knowable from GIFT beforehand — and how often that knowledge survives contact with sessions it was never fitted on.
434 sessions analysed · 2023-11-21 to 2026-09-25· 1-minute bars reduced to session anchors · 557 joined, 123 excluded as non-trading session boundaries
The verdict
GIFT NIFTY does carry real information about the Nifty opening gap — but only a little, and only if you pick the right measure of it.All three overnight definitions correlate with the 09:15 gap between r = 0.158 and 0.290, explaining roughly 2–8% of its variation. The information is real, not noise. It is also small, and92% of what moves the opening gap is news, overnight futures, and flows that arrive regardless of what GIFT did overnight.
The more interesting finding is a trap. The measure that looks best in-sample — GIFT's own "gap" at its 06:30 open, r = 0.290 — is theonly one of the three that fails once you test it on data it was never fitted on (92 pts error versus90 pts for just assuming the market opens flat). The slightly weaker in-sample measure — the full overnight move from the previous close to 09:15 — is the one that holds up (86 pts error, and a63.8% directional hit rate on unseen sessions).If you had picked the "best" model the usual way, you would have picked the one that does not work.
Survives out-of-sample
0.330
"GIFT prior ref close -> 09:15"
Out-of-sample hit rate
63.8%
vs a 50% coin flip
Out-of-sample error
86 pts
flat-gap baseline 89 pts
Variance explained
7.6%
R² on 434 sessions
Flattest-opener beats it?
No
in-sample leader, 92 pts OOS
Typical opening gap
0.33%
≈ 115 pts stdev
Works best when
Low GIFT vol
61.7% hit in calm sessions
How the comparison is actually made
A naive correlation between the two price levels would be meaningless: both indices sit near 24,000, so the correlation would return something close to 1 and tell you nothing. GIFT NIFTY also has no "opening bell" that lines up with the NSE's. So every predictor here is a percentage move from an explicit anchor, measured against the Nifty's own opening gap.
| Predictor | n | In-sample r | R² | OOS r | OOS hit | OOS MAE | Beats flat? |
|---|---|---|---|---|---|---|---|
| GIFT 06:30 open vs prior ref closebest fit | 434 | 0.290 | 8.4% | 0.353 | 54.0% | 92 pts | No |
| GIFT 06:30 → 09:15 pre-open | 557 | 0.158 | 2.5% | 0.160 | 61.7% | 87 pts | Yes |
| GIFT prior ref close → 09:15 | 557 | 0.276 | 7.6% | 0.330 | 63.8% | 86 pts | Yes |
| Benchmark: forecast a flat 0.00% gap | 174 | — | — | — | 50.3%* | 89 pts | — |
*always call the direction that was more common in the training half. Each predictor is split 60/40 in date order and fitted on the first part, so the "OOS" columns are measured on sessions the model never saw. Note the in-sample leader and the out-of-sample winner aredifferent measures — that is the whole point of this table.
Each dot is one session
GIFT NIFTY's overnight move on the horizontal axis against the Nifty 50 opening gap on the vertical. The fitted line is the full-sample regression. The cloud is wide because GIFT only explains a small share of the gap.
Does the signal scale, or is it just direction?
Sessions bucketed into deciles by GIFT's overnight move, showing the average Nifty opening gap in each bucket. A genuinely usable signal would rise cleanly left to right.
Predicted versus actual, by decile
If the regression were trustworthy, each decile's predicted gap would sit on the 45° line. Divergence below zero is the expected outcome given how little variance is explained.
Both series through time
GIFT's overnight move and the Nifty opening gap, session by session. Useful for spotting whether the relationship is stable or drifting.
Is the relationship holding up?
Rolling correlation over 126 and 252 sessions. Flat lines would mean a stable relationship; swings mean the edge comes and goes.
It works best when GIFT is quiet
Hit rate split by realised GIFT overnight volatility tercile. The calm-regime advantage is the clearest actionable finding here.
Every session, classified
Where all scored sessions actually landed. The two off-diagonal cells are the cost of being wrong — and they are not small.
When it gets the direction wrong
Of 407 scored sessions, GIFT pointed the wrong way42.8% of the time. Critically, the two error cells are close to equal (77 and 97 sessions) — the signal is not skewed toward one type of mistake, it simply misses at a similar rate in both directions.
| Nifty gapped up | Nifty gapped down | |
|---|---|---|
| GIFT was up | 145 (35.6%) | 77 (18.9%) |
| GIFT was down | 97 (23.8%) | 88 (21.6%) |
The test that matters
A correlation computed and reported on the same data that produced it mostly measures luck. So every predictor above was fitted on the first 60% of its sessions and then scored on the last 40% — sessions it had never seen, in strict date order.
This is where the ranking flips. The measure that wins on in-sample fit is the one that loses out-of-sample:
| Measure | In-sample | Out-of-sample |
|---|---|---|
| Best in-sample fitGIFT 06:30 open vs prior ref close | r = 0.290 | MAE 92 pts vs 90 pts flat |
| Survives out-of-sampleGIFT prior ref close -> 09:15 | r = 0.276 | MAE 86 pts vs 89 pts flat |
| Benchmark: flat 0.00% gap | — | MAE 89 pts |
| Benchmark: always call one direction | — | 50.3% hit rate |
Test split begins 2025-10-06 (174 sessions, held out from a 260-session training set). The surviving measure reduces average error by 3.0% versus assuming a flat open, and gets the direction right 63.8% of the time. Both are real effects and both are modest — this is a tilt, not an oracle.
Where it earns its keep
| GIFT volatility | Sessions | Hit rate | MAE |
|---|---|---|---|
| Low GIFT vol | 145 | 61.7% | 57 pts |
| Mid GIFT vol | 144 | 50.4% | 72 pts |
| High GIFT vol | 145 | 59.7% | 96 pts |
This is the most useful thing on this page. GIFT's overnight signal is materially better in quiet conditions (61.7% hit rate) than when GIFT itself is moving hard (59.7%), which is the opposite of the folk claim that a big overnight GIFT move means a big reliable Nifty move. Large GIFT moves coincide with large errors.
Data integrity notes
- Three predictor definitions were tested and the headline is the strongest of them. That is a best-of-three selection, so its in-sample p-value is optimistic; the out-of-sample hit rate is the number to trust.
- GIFT does not trade across the 02:45 -> 06:30 session boundary on 123 of 557 sessions (22%). On those nights the 06:30 "open" is the same print as the reference close, so the opening-gap predictor is 0.00% by construction. They are excluded from the "gap" predictor rather than scored as genuine flat moves.
- Selection bias, stated plainly: 1 of the 3 predictor definitions looked strong in-sample and then failed out-of-sample. The in-sample leader ("gap", r = 0.290, MAE 75 pts) became out-of-sample MAE 92 pts against a flat-gap baseline of 90 pts. Ranking predictors on in-sample fit and then reporting that ranking does not work.
- Out-of-sample: the regression was fitted on the first 260 sessions and evaluated on the last 174, which it never saw.
- GIFT NIFTY is quoted in USD while NIFTY 50 is in rupees, so a portion of the raw relationship is overnight dollar movement. This page does not adjust for USD/INR; treat the headline r as an upper bound on the signal attributable to GIFT itself.
- Headline result: out-of-sample, forecasting the opening gap from GIFT was WORSE (MAE 92 pts) than simply forecasting a flat 0.00% gap (90 pts). On accuracy grounds the signal does not pay for itself.
- Sample runs 2023-11-21 to 2026-09-25 (557 joined sessions, 434 used in the statistics above). GIFT NIFTY only began trading in July 2023 and Dhan's intraday feed starts 2023-11-10, so this is the entire available history, not a chosen window.
- GIFT NIFTY volume is zero on every bar (it is an indicative index quote, not a tradable contract), so nothing here is volume-weighted.
- GIFT NIFTY is quoted in USD-denominated index points while NIFTY 50 is in rupees. An overnight dollar move therefore leaks into the raw relationship, and this page does not adjust for USD/INR. The reported r should be read as an upper bound on the signal attributable to GIFT itself.
- Timestamps are epoch seconds in IST, verified empirically: a NIFTY request for 09:00–10:00 returned bars stamped 09:15 through 09:59 IST. Treating them as milliseconds silently renders every bar as 1970.
- All session anchors are the last available bar at or before the target minute, and a session missing an anchor within 20 minutes of 09:15 is dropped rather than filled or carried forward.
- GIFT NIFTY runs Monday–Saturday and NSE runs Monday–Friday, so only shared weekdays are eligible. 434 sessions survived the join.
Data sources
- NSE International Exchange (GIFT City) — official exchange site — NSE IX
- NSE Historical Index Data — NIFTY 50 official series — National Stock Exchange of India
- DhanHQ API v2 — Historical Data (daily and intraday OHLC) — DhanHQ
- DhanHQ API v2 — Instrument List (security master, index security IDs) — DhanHQ
- DhanHQ API v2 — Authentication (access token issuance and renewal) — DhanHQ
- GIFT NIFTY trading hours (two sessions spanning ~21 hours) — GiftNifty.org
- DhanHQ — What is Gift Nifty (session length and non-tradability of the index itself) — DhanHQ
- Information Transmission Performance of the GIFT Nifty Futures Market — Journal of Risk and Financial Management (MDPI)
NSE and NSE International Exchange are the data of record for both indices. DhanHQ is credited only as the transport used to retrieve the minute bars (its historical and intraday endpoints), not as the authority for the values. Series ids used: GIFT NIFTYIDX_I / 5024, NIFTY 50 IDX_I / 13.