← setups · Research, not advice.

Nifty option pricing models vs the real market: which fits best, and do the misfits give a buyer an edge?

Research, not advice. Read-only study; no orders placed. Built 2026-10-17.

Short answer

Most accurate model: SVI smile per expiry per snapshot. On held-out strikes its error is 1.00 premium pts RMSE / 0.23 vol pts. Bates (Heston + jumps) is close behind at 1.09 pts / 0.46 vol. SABR is 1.46 / 0.54, Heston 1.54 / 1.78, and one flat BSM IV is 10.3 / 4.2. SVI is the "fair value" used for every residual below.
No clue gave a buy-side edge after costs. All 11 smile/IV/OI/parity clues (34 variants) and the residual-outlier and parity "cheap leg" trades lost money per trade in the 2026 test. None beat its placebo in a way that held across train, tune and test.
What does hold: extreme readings mark bigger spot moves to come (volatility clustering). The option price already includes that, so straddles still lose. One weak direction effect is worth watching: a sudden ATM-IV jump while spot is flat came before an up-move 55–56% of the time in all three periods. As an option buy it still lost money.

New: Top option mismatches vs fair value (1-minute) — biggest raw/liquid/tradable gaps vs the SVI smile, where they cluster, how fast they close, buyer test (FAIL).

Follow-up: Mismatch accumulation over hours/days vs multi-day Nifty moves — ML test (GBM, HMM, Isolation Forest, SHAP): direction FAIL, range-only predictor.

Data used

Options: real Upstox 1-minute OHLC/volume/OI for every listed Nifty weekly strike, from S3 options-ml-platform/data/options/1m/ (105 expiry files plus the live file for the 2026-10-06 expiry, 117M rows). There is no 1m option history before Aug 2024.
Span analysed: 2024-09-27 → 2026-10-01, 494 sessions. That is the first day with a true front-week expiry in the data through the last option bar. The front and next weekly expiries were used.
Spot: S3 copy of nifty50.parquet (1m), read-only. VIX: not available in the bucket. The front-week ATM IV from our own fits stands in for it.
Splits: train 2024-09-27 → 2025-06-30 (≈190 days), tune 2025-07 → 2025-12, test 2026-01 → 2026-10-01 (≈185 days). Thresholds were set on train, the horizon chosen on tune, and test looked at once.
Sampling: a 5-minute snapshot of each chain, 09:20–15:25. A minute-by-minute run was not needed because the fits are stable minute to minute. Entries and exits use the real 1-minute option closes. Model comparison (including Heston and Bates) used every 30-minute snapshot: 24.6k slices and 273k held-out quotes.
Quote filter: options traded in the last 5 minutes (≈49% of 1m rows are zero-volume carry-forwards), premium ≥ 0.5, |log(K/F)| ≤ 5%. Fits use OTM quotes only (puts below the forward, calls above).
Forward: implied per expiry per snapshot from put-call parity, F = K + (C − P)·e^{rT}, taking the median of the 5 strikes nearest spot. Dividends are therefore implicit.
Rates: constant 91-day T-bill: 6.8% (2024), 5.8% (2025), 5.5% (2026). With weekly options and an implied forward, the rate barely matters.
Clock: trading-time T = remaining trading minutes / (375 × 252). Overnight gaps are not counted separately, so IVs are on a 252-trading-day basis, about 0.83× VIX-style calendar IVs.

Model accuracy ranking (out of sample)

Three tests: holdout = fit on 100-point strikes, price the 50-point strikes in between at the same time. ahead30 = parameters from 30 min earlier, with the new forward and T. dayfixed = parameters fixed at 10:00 for the rest of the day.

model	holdout RMSE pts	holdout RMSE IV (vol pts)	median abs IV err	ahead-30m RMSE pts / IV	day-fixed RMSE pts / IV
SVI (raw, per expiry per snapshot)	1.00	0.23	0.07	3.03 / 0.69	6.99 / 2.03
Bates (Heston + Merton jumps)	1.09	0.46	0.07	3.07 / 0.74	7.04 / 1.72
SABR (β=1, Hagan)	1.46	0.54	0.17	3.19 / 0.95	7.16 / 2.27
Heston	1.54	1.78	0.18	3.29 / 2.01	7.51 / 3.20
BSM flat IV	10.33	4.25	1.31	10.55 / 4.34	–

By time to expiry (holdout, IV RMSE): SVI 0.54 / 0.30 / 0.21 / 0.16 vol pts for <1d / 1–2d / 2–4d / >4d. Heston breaks down on expiry day (5.9 vol pts <1d) because it cannot make the steep short-dated wings without jumps. Bates fixes most of that (1.1).

By moneyness, SVI wins every bucket: deep OTM puts 0.29, ATM 0.16, deep OTM calls 0.25 vol pts. 2026 test front week: SVI 0.71 pts / 0.26 vol, Bates 0.95 / 0.46, SABR 1.33 / 0.73, Heston 1.47 / 2.56.

Arbitrage checks (SVI): 0.36% of front/next pairs show a calendar violation on |k| ≤ 4% (next-week total variance below front-week). The Gatheral butterfly condition g(k) ≥ 0 holds on |k| ≤ 8% for 97% of next-week fits and 82% of front-week fits. The failures are almost all in the extrapolated far wings near expiry, beyond the quoted strikes. Median in-sample fit error is 0.12 vol pts.

Pricing lesson: a single BSM IV is off by about 10 premium points. SVI, Bates and SABR fit the weekly chain to about 1 point, roughly the bid/ask and 1-minute staleness noise. Weekly Nifty options are priced very consistently with one smooth smile. Little of the "mispricing" is left to harvest.

Charts

Model accuracy

IV smile fits (one snapshot)

Model vs market premiums over a day

Residual heatmap strike × time, spot overlaid

Outlier timeline vs spot (2026)

Verdicts: 2026 test P&L vs placebo

Move size after signal (volatility clustering)

Clues tested and verdicts

Rules: top/bottom decile thresholds come from train. Direction comes from train (the sign of the excess 15-minute spot move). The horizon (5/15/30/60 min) is picked on tune. 30-minute cooldown. Entry at the next 1-minute close and exit at the close h minutes later, using real front-week ATM CE/PE (or the flagged contract).

Costs: lot 65, 1 pt slippage per side on premium, ₹50 per round trip per leg. Placebo: the same side and same time of day on 10 random days in the same split. Hit rates are spot direction, with Wilson 95% CI. Full grid: clues_continuous.csv, clues_continuous_allh.csv, clues_contract.csv.

clue	best variant (side, h)	train n / ₹ (placebo)	tune n / ₹ (placebo)	test n / ₹ (placebo), t	test hit (CI)	verdict
C1 residual z ≤ −3 & ≥1 pt cheap vs SVI → buy that option	buy, 30m	181 / −267 (−228)	32 / −343 (−163)	14 / −35 (−152), t 0.7	–	FAIL
C1 residual z ≤ −2 (looser)	buy, 30m	642 / −270 (−195), t −2.3	196 / −213 (−200)	173 / −123 (−175), t 1.1	–	FAIL
C2a parity: strike's C−P implies forward ≥3 pts off → buy cheap leg	30m	5362 / −194 (−188)	2599 / −184 (−184)	4796 / −162 (−179), t 2.5	–	FAIL (train t −3.4 at 15m, sign flips)
C2b implied forward − spot jumps up (15m)	CE 15m	593 / −199 (−222)	224 / −197 (−196)	363 / −138 (−206), t 1.0	51.5% (46–57)	FAIL
C3a 25Δ risk reversal change (15m), top	PE 5m	466 / −195 (−186)	57 / −103 (−175)	204 / −173 (−190)	47.8% (41–55)	FAIL
C3a risk reversal change, bottom	CE 5m	490 / −168 (−201)	69 / −165 (−196)	239 / −279 (−171), t −2.1	46.0% (40–52)	FAIL
C3b butterfly (curvature) jump, top	CE 5m	420 / −212 (−182)	42 / −191 (−198)	199 / −82 (−186), t 1.6	55.3% (48–62)	FAIL (train worse than placebo)
C4 ATM IV − 60m realized vol, lowest decile ("cheap vol") → straddle	straddle 60m	234 / −535 (−403)	32 / −282 (−433)	80 / −570 (−590)	–	FAIL
C4 IV − RV, highest decile	CE 5m	233 / −145 (−167)	0	384 / −156 (−188)	53.1% (48–58)	FAIL
C5 term spread (front − next ATM IV) high → straddle	straddle 30m	201 / −447 (−429)	5 / −259	94 / −145 (−389), t 1.1	–	FAIL (tune n=5)
C6 ATM IV jump (z>0.42) with spot flat → buy CE	CE 15m	225 / −162 (−181)	61 / −228 (−221)	133 / −104 (−167), t 0.5	54.9% (46–63); train 56.4%, tune 55.7%	WATCH (direction only), FAIL as trade
C6 IV drop with spot flat → PE	PE 30m	219 / −73 (−171)	55 / −30 (−160)	67 / −392 (−76), t −1.6	52.2%	FAIL
C7a put-minus-call OI build while IV up	PE 30m (bottom)	206 / −36 (−146)	117 / +50 (−162)	299 / −226 (−226)	53.2% (48–59)	FAIL (failed in test)
C7b put OI build while IV down	–	–	–	313–327 / −175..−188 (≈ placebo)	46–51%	FAIL
C7c total near-ATM OI unwind (bottom) → straddle	straddle 30m	440 / −329 (−387)	294 / −342 (−389)	468 / −273 (−348), t 1.9	–	FAIL (consistently less bad than placebo, never positive)
C8 OTM calls vs puts richer than 15 min ago	5m	436–455 / −186..−211	+2.3 t in tune	223–265 / −222..−287, t −2.6	45%	FAIL (sign flips)
C0 plain 15m ATM IV change	–	426–434 / −235..−286	–	163–292 / −174..−201 (≈ placebo)	50–52%	FAIL

Why the residual outliers don't pay: cheap-vs-SVI outliers persist. Only 5–18% flip sign within 5 minutes, so they are mostly staleness or illiquidity in that strike, not a temporary mispricing that snaps back. The typical outlier is only 1.3–3.3 premium pts off fair. The round-trip cost on a cheap OTM option is ≈2.8 pts (₹180/lot), which absorbs it. The placebo OTM buy of the same moneyness loses ≈₹175–230 per trade on its own, so gross P&L ≈ 0 and costs decide it.

Magnitude (volatility clustering), PASS as a predictor. In 2026, nearly every extreme reading was followed by bigger-than-average spot moves. Examples: butterfly jump |5m move| 18.9 vs 11.5 pts average; term-spread inversion 19.3 vs 11.5; IV jump with flat spot 24.8 vs 19.6 (15m); "cheap vol" decile 47.8 vs 40.2 (60m). Straddles bought on these signals still lose (−₹145 to −₹570) because the market already marks IV up at those moments. Use it for position sizing and stop width, not as an entry.

Variance risk premium: front-week ATM IV averaged +2.6 / +2.2 / +4.5 vol pts above trailing 60-minute realized vol (train / tune / 2026). Over 30–60 minutes the average straddle buy lost ≈₹370–590 per pair including ≈₹360 in costs and slippage. Gross, buyers lose a little theta on average, and the edge sits with sellers, which is outside Satya's scope.

Rule worth watching (not trading)

"IV pops, spot doesn't move" → lean long. On a 5-minute snapshot of the front-week chain (09:30–14:25): - the SVI-implied ATM IV rises by more than 0.42 × its train-period 5-minute standard deviation, i.e. ΔATM-IV(5m) > ≈0.17 vol pts on the 252-day clock; - while |Δspot over the same 5 min| < half the train median 5-minute move (4.55 pts).

Then the next 15 minutes closed higher 56.4% (train, n=225), 55.7% (tune, n=61), 54.9% (2026 test, n=133, CI 46–63%). Buying ATM CE for 15 minutes still lost −₹162 / −₹228 / −₹104 per trade against a −₹167 placebo. Treat it as a direction filter for other setups, not a standalone buy.

Honest limits

Only ~2 years of real 1-minute option history exists (Sep 2024 → Oct 2026). The tune half-year had low vol (ATM IV ≈ 8.6%), so some train-decile thresholds rarely fired there (n = 0–5 for C4/C5). Their "best horizon" defaulted to 5 minutes.
Prices are 1-minute closes (last trade), not bid/ask. Even after the 5-minute traded filter, a quote can be up to a few minutes stale, which inflates residual "outliers". A tick/quote feed could change the C1/C2a conclusions at the margin, but the ≈2.8-pt cost hurdle is far bigger than the typical 0.4-pt parity gap.
Models were compared on 30-minute snapshots. The SVI fair value runs on every 5-minute snapshot. Heston/Bates were refit per snapshot on that 30-minute subset, plus a "fixed at 10:00 for the day" variant, instead of minute-by-minute (compute scope cut).
The trading-time clock ignores overnight variance. This affects IV levels and IV−RV comparisons, but not the model ranking.
Many clues and variants were tested (≈34 continuous plus 6 contract variants × 4 horizons). A t of about 2 in one split is expected by chance. That is why the verdicts require train, tune and test to agree.
Only option buying was tested, as required. Several results (the VRP, rich-side residuals) would point to selling or spreads instead. Those were not tested.

Glossary (plain English)

Implied volatility (IV): the volatility number you must plug into Black-Scholes to reproduce an option's market price. It is the market's price of "expected movement", not a forecast.
BSM (Black-Scholes-Merton): the classic formula. With one IV for all strikes it can't match the market, because real puts below spot cost relatively more (the "smile/skew").
Smile / skew: IV plotted against strike. Nifty weeklies slope down to the right (OTM puts dearer) with a small upturn on far calls.
SVI (Stochastic Volatility Inspired): a 5-parameter curve, w(k) = a + b(ρ(k−m) + √((k−m)² + σ²)), fitted to total variance w = IV²·T against log-moneyness k. It is flexible, fast, and can be checked for arbitrage. It describes the curve but doesn't explain it.
SABR: a model where the volatility itself wanders randomly and is correlated with price (α = vol level, ρ = skew, ν = vol-of-vol). The Hagan formula gives the smile directly.
Heston: a full stochastic-volatility model (mean-reverting variance with speed κ, long-run level θ, vol-of-vol ξ, correlation ρ, starting variance v0), priced here by Fourier-COS. Bates adds random price jumps, which short-dated options need.
Put-call parity: C − P = discount × (F − K). If it fails at one strike, either the call is cheap or the put is rich relative to the other. Here the forward F is inferred from parity itself.
25-delta risk reversal (RR): IV of the 25-delta call minus IV of the 25-delta put. More negative means puts are relatively more expensive (more fear of a fall).
Butterfly (BF): average of the 25-delta call and put IVs minus ATM IV, i.e. how curved the smile is (the price of tail moves).
Realized vol (RV): how much spot actually moved, here over the last 60 minutes, annualized.
Term structure: front-week ATM IV versus next-week ATM IV.
Residual / z-score: market IV minus SVI fair IV, divided by the typical spread of residuals in that snapshot. z ≤ −3 means unusually cheap versus the smile.
Placebo: the same trade at the same time of day on random other days. It shows what "just buying an option" earns, so a signal must beat it, not just zero.

Files

build_snap.py (snapshots, implied forward, IV), models.py (BSM/SVI/SABR/Heston/Bates-COS), fit.py (fits plus OOS errors), rank.py, signals.py/study.py/study_contract.py (clue tests), charts.py, rank_*.csv, clues_*.csv.