edge

NIFTY's 5% downside line gets pierced 7.8% of the time. The textbook is wrong.

2026-05-04 · 1 months ago · 5 min read · Quantcents Research

If you sell NIFTY puts, you're using a model — even if you don't know its name. Whatever your broker, your option-chain widget, or your head says about "this strike has a 5% chance of being breached in 5 days," there's math behind it. And every option-seller's instinct is to trust the math.

We tested it.

Across 562 weeks of NIFTY history (Jan 2024 → Apr 2026), we asked four different models the same simple question every Monday: "Where is the line that NIFTY shouldn't close below in the next 5 days, more than 5% of the time?"

Then we waited 5 days, looked at the actual close, and counted how often the line got pierced.

How often NIFTY actually breached each model's 5% downside line562 weeks · 5-day forward · 2024–2026

Bottom-up simulation4.6%

Sample whole panic days from history; preserves cross-stock correlation

Naive Gaussian (textbook)7.8%

Bell curve with mean and stdev from history

GARCH(1,1) (vol-clustering)9.4%

The standard quant model — captures vol clustering, ignores cross-asset correlation

Dashed line = the 5% the model promises. Anything to the right of it = the model under-priced downside risk.

Read this chart: each model gives you a NIFTY level below which it claims there's only a 5% chance of closing in 5 trading days. We checked, week after week, how often the actual close was below that line. The textbooks said “5% chance.” Reality said “8–10% chance.”

The textbook's promise vs the textbook's reality

The first model is the bell curve — what most courses teach. Take NIFTY's average daily move, take its volatility, project 5 days forward, draw a line at the 5th percentile. Simple. Defensible. Wrong.

It promised 5% downside breaches. It delivered 7.8% — , . Fifty-six percent more than advertised, and the deviation isn't sampling noise.

The second model — GARCH — is the standard quant upgrade. It accounts for vol clustering (calm days follow calm days, panic follows panic). Better, in theory.

It promised 5% downside breaches. It delivered 9.4% — , . Eighty-eight percent more.

The third model is what we built: a bottom-up simulation that takes all 50 NIFTY constituent stocks at their authoritative NSE fact-sheet weights, reaches into 5 years of history, and samples whole panic days — every stock's return, from the same actual day, all together. Forward-projects 5 days. Draws the 5% line.

It promised 5%. It delivered 4.6% — , . The 5% claim survives the test; the textbook's doesn't.

How off were the textbook models?

Model	Promised tail	Actual	Mispricing
Bottom-up simulation	5.0%	4.6%	-8%
Naive Gaussian (textbook)	5.0%	7.8%	+56%
GARCH(1,1) (vol-clustering)	5.0%	9.4%	+88%

Mispricing = how much more often the actual close pierced the “5% line” than the model said it should.

How the bottom-up simulation actually works

The whole reason it gets the tails right is what happens in step 2. Read the four steps and you'll see why the textbook misses panic days by design — and why the fix is simple, not exotic.

How the bottom-up simulation works · 4 steps

Pick a real historical day at random

e.g. June 4, 2024 — election results day

Out of ~1,300 trading days in our 5-year history, we pick one. It might be a calm day. It might be a panic day. Whatever it was, every stock's actual move from that exact day comes along with it.

Take all 50 stock moves from that day, together

RELIANCE −3.2% · HDFCBANK −2.8% · ICICIBANK −4.1% … all 50 of them

This is the single most important step. The textbook draws each stock's move independently. We don't. By taking all 50 from the same real day, we keep their correlation — the panicked-together pattern that's the whole story of NIFTY's worst days.

Weight by NIFTY's current composition

RELIANCE × 9.90% + HDFCBANK × 6.07% + … = synthetic NIFTY move for that day

The 50 weights come from the public NSE fact sheet. They sum to 100%. Big stocks count more, small stocks count less, exactly as they do in the real index.

Repeat 5 times for a 5-day path. Repeat the whole thing 5,000 times.

5 random historical days → one possible 5-day NIFTY path → 5,000 paths total

Each path is a possible 5-day future. After 5,000 paths we have a distribution. The 5th percentile of that distribution is what we call the model's downside line — the number this whole article is about.

Why this matters

Step 2 is the trick. The textbook bell curve and even the standard quant model (GARCH) assume each stock's move is rolled independently — like flipping 50 coins. That makes panic days disappear mathematically. Reality doesn't flip 50 coins independently; on real panic days, all 50 stocks fall together. We sample real panic days as units. The math gets the tails right because it's pulling from what actually happened, with the correlations baked in.

Why the textbook fails — visually

Imagine the bell curve and GARCH as a coin-flipping factory. They estimate one number for "today's volatility" and use it to roll each day's NIFTY move from a Gaussian. Two consequences:

Each day is a fresh roll, independent of the recent ones. That's true on average but breaks during panic chains — Monday's rout doesn't "know" Tuesday should be priced wider.
The sequence inside NIFTY's 50 stocks is implicitly synthesized. Independent rolls per stock don't reproduce the days when all 50 dropped together. The model averages those days into a number called "vol" and forgets they happened in a cluster.

The bottom-up version doesn't synthesize. It picks a historical day and re-uses it as a unit. If March 23, 2020 was a 13% NIFTY drop, that day shows up in our distribution at the exact correlation pattern that produced it — RELIANCE, HDFCBANK, ICICIBANK, and 47 others falling together by the precise amounts they did. The fat left tail of NIFTY's distribution isn't a theoretical artifact for us. It's just what actually printed.

Why this matters if you sell options

Selling NIFTY puts at the q05 strike (whatever your model says is "5% downside risk") is a popular trade. It generates premium, it usually expires worthless, it feels safe.

If your model is the textbook bell curve, you're getting paid for 5% risk while taking 7.8% risk. The "extra" 2.8 percentage points are the trades where the put gets exercised on you, and the put-payoff is non-linear — those few losing trades wipe out a lot of the premium you collected on the winners.

GARCH is worse, not better. It's narrower, which feels precise — but it's narrower because it under-prices crashes harder.

Selling against a calibrated model isn't free money. But it does mean the premium you collect actually corresponds to the risk you took. You're not paying for the textbook's confidence and getting reality's drawdowns.

What this is not yet

One window. 2024 through April 2026. We haven't yet tested it across the 2008 or 2020 stress samples. We will.
Constant-weights. We applied current (April 2026) NIFTY weights to historical days. NIFTY rebalances semi-annually; using time-varying weights is the next refinement. The current basket is partially hindsight: today's mega-caps were also the names that printed the 2020 dispersion, so some of the bottom-up advantage is mechanical until we re-run with point-in-time NSE weights.
Unconditional bootstrap. Each historical day is sampled with replacement, weighted equally. A 2020 panic day and a quiet 2025 session sit in the same urn. That's a base-rate calibration — strong on average, weaker conditionally. Filtered Historical Simulation (Barone-Adesi/Giannopoulos) re-weights the urn by today's regime; that's the next iteration we'll publish.
No live PNL backtest yet. We've shown the model's downside line is statistically right; we haven't yet sold puts against it for 100 weeks and counted rupees.

That last bullet is the next step. The current evidence already lets us say "NIFTY's left tail is being mispriced by every model that ignores cross-stock correlation," and that's strong enough on its own. But the rupee-PnL test is what would close the loop from "statistical edge" to "tradeable edge."

What to take away

If you sell NIFTY downside premium, your effective tail risk is probably 56–88% bigger than your model says — depending on which textbook model you use. That's not a research-paper insight. That's the difference between being paid fairly for the risk you took and bleeding slowly to the few weeks when reality breaks the model.

Knowing the gap is half the trade. Pricing for it is the other half.

Sources

562 weekly walk-forward simulations across Jan 2024 – April 2026
All 50 NIFTY 50 constituents at their authoritative weights, NSE fact sheet as of 2026-04-30
1,294 trading sessions of daily history per stock
GARCH baseline fit via the arch library