Part one of two. Here I build the settlement mechanism and the legitimate edge I traded inside it. The second part turns the same mathematics into a weapon and walks through the Jane Street case.

There is a particular kind of quiet that used to settle over my desk in Mumbai at 3:00 PM on a Tuesday. Not calm, the opposite. The screens were still flickering, the index was still printing, but a different clock had started, and I could feel the rest of the floor relax into a number that no longer mattered. For the next thirty minutes the thing I cared about was not the one the terminal was showing me. It was an average I had to assemble myself, print by print, in a spreadsheet off to the side, while everyone who trusted the screen slowly walked into the wrong settlement.

I spent years on both the wrong and the right side of those thirty minutes. Over time, I realized there was a mathematical explanation for what I was seeing. This post is about that realization. Almost everything I learned about Indian index options at expiry, including the edge I traded in the last half hour and the reform going live this August, comes down to one sentence I wish someone had told me on day one:

An Indian expiry option is not a European option on the close. It is an Asian option on a volume-weighted average and that average is about to be replaced by a point.

Hold that sentence. Everything below is me unpacking it.

1. The object

The first thing I had to unlearn was that the settlement is the close. It isn't, and it never has been. For NSE and BSE index derivatives, until the reform I will get to in Section 5, the closing price of a stock and therefore the index built from those stocks is the volume-weighted average price (VWAP) of trades over the last thirty minutes of the session, 3:00 PM to 3:30 PM. The live index on the terminal is a snapshot: last-traded prices, recomputed tick by tick. The settlement is an average, over half an hour of real transacted volume. On a quiet expiry the two sit a few points apart and nobody notices. On a violent one I have watched them pull tens of points apart and that gap is paid for entirely with other people's money.

Let me write the window as $[0, T]$ with $T$ = thirty minutes. Let $V_t$ be cumulative traded volume and $V$ its total over the window. The settlement is

A call expiring into this settlement pays $(S^* - K)^+$. The moment I wrote that down properly, the whole thing clicked: that payoff is an arithmetic average option: an Asian option. But the averaging is not in calendar time. It is in volume time, and that distinction is the entire game. So let me make the clock explicit. Define the volume fraction

and change variables from calendar time to volume time $\tau$ via $d\tau = dV_t / V$. Then

the settlement is just the time-average of price in the volume clock. A fat calendar minute at 3:28 in which nothing trades barely advances $\tau$ and barely counts. A burst of volume at 3:04 advances $\tau$ fast and counts for a lot. The market does not settle in the clock on your wall. It settles in the clock of its own turnover and once I started thinking in that clock instead of the wall clock, everything else fell out.

2. Why the settlement freezes and freezes cubically

Stand at calendar time $t$, with volume fraction $w = w_t$. I split the average into what is already locked and what is still live:

The first piece: the running VWAP, the average over the volume already done is known to me at that moment. The second: the VWAP of the volume still to come is not. Model the price in the volume clock as a driftless martingale with volume-time variance rate $\varsigma^2$ (over thirty minutes near the money this is harmless; I will come back to the assumption). Its conditional expectation, given everything I know so far, is just the current spot, so the conditional variance of the settlement is

Here is the part it took me embarrassingly long to see. The future VWAP is itself an average of the residual path, so it carries its own internal cancellation. For a Brownian path of volume-time length $\ell = 1 - w$ starting from a known point, the variance of its average is $\varsigma^2\ell / 3$. Put that in and the result is cleaner than I expected:

The settlement's conditional variance collapses as the cube of the remaining volume fraction. One factor of $(1 - w)$ because the unknown piece is only a fraction of the average; a second because that piece is small; a third because averaging the residual path cancels half of what is left. The conditional standard deviation falls like $(1 - w)^{3/2}$.

Compare that to the European object everyone's instinct reaches for, a claim on the close, $S(T)$. Its residual variance is just $\varsigma^2(1 - w)$, the remaining variance with no averaging at all. The ratio is

At the halfway point of the volume: $w = \tfrac{1}{2}$, which on a normal expiry arrives well before 3:15, the settlement carries about one-twelfth the variance of the close. The terminal is still quoting options as if a full half-hour of risk were left to run. It isn't. The risk is already gone; it left in the volume. The first time I put real numbers through that ratio I remember thinking the screen wasn't just lagging, it was pricing a different instrument.

One honest caveat on the model. The cubic law assumes a driftless martingale in volume time and volume increments roughly independent of price increments. Over a thirty-minute window near the money those are mild, but they are not free: a strong directional drift, or volume that clusters around price moves (it does, somewhat), will bend the constant and muddy the clean cube. I treat $(1 - w)^3$ as the leading-order behaviour, not gospel but the qualitative claim, that the variance dies faster than the remaining volume fraction itself, survived everything I ever threw at it.

3. The Greeks that vanish instead of exploding

This is the part that broke my European-options intuition, and I suspect it will break yours. Every trader knows the terror of expiry gamma: as you approach the strike at expiry, a European option's gamma blows up, delta whipsaws between 0 and 1, and pinning turns into a knife fight. Under VWAP settlement the exact opposite happens, and the first time I lived through it I didn't trust my own risk screen.

Conditional on what I know, the settlement is approximately Gaussian, centred at the volume-weighted blend of the running VWAP and the current spot, with a standard deviation that has already collapsed. In the Bachelier approximation the call is then:

The settlement responds to the current spot only through the live fraction, a sensitivity of $1 - w$ and propagating that through gives the Greeks I actually had to hedge:

The delta to spot decays linearly to zero; the gamma to spot decays like $\sqrt{1 - w}$ to zero. The pin that European expiries are infamous for is not amplified into the close, it is smeared away by the averaging. By 3:25, on heavy volume, I could lean on the spot and watch the option barely flinch, because the spot was no longer where the settlement was being decided. The settlement was already three-quarters written.

This is the mathematical content of the desk-lore "the VWAP locks in." It is not a feeling. It is a $(1 - w)^3$.

4. The convergence trade (the legal edge)

So here is what those thirty minutes were worth, legally, to anyone willing to compute the running VWAP. This was a real line in my book for years, and it is the cleanest illustration of the whole point.

The terminal prices each option off the live spot with calendar time-to-expiry $T - t$ as a European claim on the close, with remaining variance proportional to $T - t$. The truth is a claim centred at the volume-weighted blend of running VWAP and spot, with the collapsed variance $\varsigma^2(1 - w)^3 / 3$. So the screen is wrong in two places at once:

• The centre is wrong. When the running VWAP has drifted away from the live spot, which it does whenever the last fifteen minutes ran directional, the option's true centre sits a distance $w \times (\text{running VWAP} - \text{spot})$ from where the screen thinks it is. The more volume has printed, the more weight $w$ sits on the part the screen is ignoring.
• The width is wrong. The screen carries remaining calendar variance, order $(T - t)$. The truth carries $(1 - w)^3 / 3$ of the per-unit variance — already gone.

The trade then writes itself, and on a good expiry it felt less like a bet than like arithmetic. Late in the window, with $w$ large and a real gap between running VWAP and spot, the options bracketing the true settlement were mispriced against the ones bracketing the stale spot. I would take the side that had to converge and hold it into settlement, and my risk shrank as $(1 - w)^{3/2}$ while the rest of the screen kept marking a number that, in the way that mattered, had already stopped moving. I kept that running VWAP in a sheet beside the terminal precisely because most people around me did not, and the edge lived entirely in that gap of attention.

I want to be clear that there is nothing improper here. You are a price-taker, forecasting a functional of a path you do not control. You are computing the average everyone has agreed to settle on, and noticing the screen hasn't caught up. That is just what an edge is. The trouble starts the moment someone stops forecasting that average and starts steering it. (we will discuss this in part 2).

5. The reform: from an average to a point

When the SEBI circular landed on 16 January 2026, I read it twice, because it quietly dismantles everything above. The mechanism matters in the detail, so here is what the circular and the exchange's operating guidelines actually say.

From 3 August 2026, for stocks that carry derivatives, the closing price will no longer be the thirty-minute VWAP. It is set by a Closing Auction Session (CAS), a separate twenty-minute call auction from 3:15 PM to 3:35 PM:

• 3:15–3:20, transition. The exchange computes a reference price — and notice what it is: the VWAP of trades from 3:00 to 3:15 (or the last traded price if nothing trades). A fifteen-minute VWAP survives, but only as the anchor that centres the auction's ±3% band.
• 3:20–3:25, order entry. Limit and market orders both go in and both count toward the clearing price. Unexecuted limit orders from the continuous session carry forward with time priority; stop-loss and iceberg orders are barred; market orders get execution priority.
• 3:25–3:30, limit only. No new market orders, and existing ones cannot be modified or cancelled.
• Random close. The order window shuts at a system-chosen random moment in the last couple of minutes — explicitly so no one can act on the closing instant.
• 3:30–3:35, matching. The book uncrosses at the equilibrium price that maximises matched volume. Cash then closes; equity derivatives keep trading to 3:40, and a post-close cash session runs 3:50–4:00 at the determined close.

And here is the part that made me put the circular down. The settlement rule for an index derivative did not change, it still settles at the closing price of the underlying index on expiry. What changed is the meaning of that closing price underneath the rule:

The settlement remains the "closing price of the underlying index on the day of expiry." But that index close is now built from its constituents' prices: for any stock in CAS, the close is "determined based on the equilibrium price determination mechanism in CAS," and the old thirty-minute VWAP is kept on only as the reference price that centres the ±3% band, now measured over 3:00–3:15.

— SEBI circular HO/47/11/11(3)2025-MRD-POD2/I/2765/2026 (16 Jan 2026); NSE operational circular NSE/CMTR/73362 (18 Mar 2026)

That is the whole reform in one move. In the language I have been using, the settlement operator changes from the path-average

to a point statistic of the closing order book,

Every result in this post hangs on which operator is in force, and swapping it turns them all over at once:

• The cubic variance collapse dies. There is no window to accumulate. The settlement's conditional variance does not glide down as $(1 - w)^3$; it stays live to the uncrossing, then drops to zero in one step. A step function, not a glide.
• The convergence trade loses its ground. My whole edge in Section 4 was that an accumulating average had locked in while the screen lagged. There is no accumulating average to ride. You can read the indicative equilibrium price the exchange publishes during the auction, but now you are forecasting a single discrete clearing, not watching a number harden print by print. The edge doesn't migrate. It evaporates.
• The pin comes back. The averaging that smeared expiry gamma into something harmless is gone. A single uncrossing restores the full terminal gamma of a European expiry. In one stroke the auction makes the close sharper, not softer — the opposite of what the VWAP did, and the thing I'd least expected when I started reading.

One clarification, because it is the first objection a careful reader will raise: VWAP is not abolished. It is demoted from the thing that set the settlement to the reference price that merely centres the band, and it lives on wholesale for non-derivative names and, cross-venue, for single-stock derivatives. What dies is specifically the thirty-minute time-averaging window as the index settlement and that is the object every result above was built on.

SEBI's own reasoning closes the loop, and it is hard to argue with: the thirty-minute VWAP can be distorted when large orders hit near the close, and those distortions feed straight into index values, derivatives settlement and mutual-fund NAVs. An auction with a random close is harder to push, and it brings India in line with most major markets. There is a fourth consequence of the operator swap what it does to outright manipulation and it is the most important one, because it is the reason the reform exists at all. But that needs the case that makes it concrete, and that is Part 2.

6. What I take from the first half

What stays with me is how much of a market's character lives in a definition almost no one reads. For three decades the Indian expiry was an Asian option wearing a European option's clothes, and a whole quiet economics of the last half hour, the convergence edge I traded, the smeared gamma, the settlement that stopped moving while the screen kept quoting was encoded in one choice: to settle on an average, $\int_0^1 S \, d\tau$, rather than on the closing price. Change that integral to a single auction uncrossing, an $\arg\max_p$ in place of an $\int$ and a different market falls out, with different edges and different risks.

The screen never showed me any of it. It showed me the spot, confidently, right to 3:30, and it was wrong about the only number that paid.

But the same functional that handed an honest trader an edge handed a large enough book something far more dangerous: a lever. The convergence trade forecasts the average. What happens when you stop forecasting it and start steering it? That question has a precise mathematical answer a calculus-of-variations problem whose solution is a parabola and a ₹4,843 crore answer in the real world.

That is Part 2.

Next: the same functional, turned into a weapon — the mathematics of marking the close, why it only pays at enormous scale, and the Jane Street case that put it on trial.

Regulatory details reflect SEBI's circular of 16 January 2026 and the NSE operational circular of 18 March 2026; the CAS goes live 3 August 2026.