CAS Games and Algo Trading

Two weeks ago, in The Phantom Close, I wrote about India’s new closing auction. For about fourteen minutes every afternoon the index chart stops moving while futures and options keep trading, and then the index jumps in one step to catch up. I treated that jump as an accident of arithmetic. Nobody’s fault.

On 19 August, SEBI passed an ex-parte interim order (ref. WTM/KV/ISD/ISD-SEC-6/32673/2026-27, 19 August 2026) saying that on 13 August, somebody made that jump happen on purpose.

Below: what happened, in plain language. What it cost the people on the other side. And then the part I actually care about — what one fake closing price does to the indicators that almost every positional trader and every overnight algo in India is running, and how long it stays there.

Since 3 August, the closing price is not the last traded price any more. It is an auction.

For every stock that has F&O, normal trading now stops at 15:15. Between 15:15 and 15:20 the exchange fixes a reference price. Then from 15:20 to 15:30 there is a closing auction: everybody puts in orders, nobody trades, and at a random second between 15:28 and 15:30 the auction shuts and one single price is struck — the price at which the largest quantity can trade. Orders have to be within ±3% of the reference price. That struck price becomes the close, and the close is what your expiring options settle against.

While the auction runs, the exchange keeps broadcasting an indicative price for every stock and an indicative index — “this is what the close would be if we shut the auction right now.” Options carry on trading through all of it, and traders watch that indicative index to decide what an option is worth.

Here is the thing to hold on to, because everything else follows from it. That indicative index is not built from trades. It is built from orders. And an order can be cancelled. A trade cannot.

Two players, opposite directions, the same ten minutes — on a day the market had barely moved.

13 August was a Sensex weekly expiry. The order describes two entities, which SEBI says were not acting together, pulling the indicative index in opposite directions — each in the direction its own expiring option book needed.

The concentration is what makes this remarkable. In the two seconds of the first spike, 88 buy orders came into the Sensex stocks, worth ₹66.64 crore. Thirty-two of them were Copthall’s — and those thirty-two were 99.91% of the entire buy value. Every one priced 2.7% to 3.0% above reference. Every one stamped 15:20:42. Nobody else had a single order above +2%.

The same shape repeats in the second spike (96.09%) and the third (85.21%). Across the whole auction, Copthall was 86.6% of all buying.

The order records the indicative index at each turn. Plotted, it looks like a tug of war — one side lifting it to the top of the band in two-second bursts, the other holding it down with sell orders that were never going to trade.

The band on that chart is the closest thing here to a controlled experiment. Three completely separate ways of asking “where was the Sensex really that afternoon?”

• The exchange’s own reference price at 15:15: 77,829.60.
• The last live one-minute print before the freeze, from my own data: 77,861.48.
• SEBI’s fair-value estimate, built separately by taking Nifty’s −0.16% move that day and applying it to the previous Sensex close: ~77,840.

All three sit inside a 32-point band. The auction closed at 78,079.96 — about 240 points above every one of them.

I find the regulator’s method quietly reassuring, because it is the same move anyone checking this has to make: when the number in front of you is not a traded price, check it against something that was trading. SEBI used a sister index. In the earlier note I used index futures and put-call parity. Different tools, same logic, same answer.

The number that should bother option traders is not ₹3.68 crore. It is a price of 108.35.

Copthall’s side is easy to follow. The index finished 240 points above fair, so its calls paid more, and its short puts — which would have owed ₹1.35 crore to the buyers at the 78000 strike — expired at zero instead. The order values the package at ₹2.96 crore, against roughly ₹57 lakh of bad pricing in the cash market.

Mansi’s side is the one worth understanding properly.

Mansi was holding puts. The index was going to close above all of its strikes, so those puts were headed for zero. According to the order, during the five minutes its sell orders sat in the book pushing the indicative index down — 15:21 to 15:26 — Mansi sold the entire position. Four seconds after finishing, it cancelled the sell orders.

This is where a market-structure argument turns into somebody’s actual account statement. Not the settlement price — the quotes on the way to it. A trader watching an expiry-day index apparently roll over in the last ten minutes, paying 108.35 for a put, was paying a price built out of intent that did not exist.

SEBI says as much itself, explaining why it acted immediately instead of waiting: such practices have “huge ramifications on participants who trade in F&O segments including retail investors”. And its reason for urgency is sharper still — both entities had already built positions in the next week’s Sensex options.

One detail I would not skip, because it cuts against the easy story: Mansi’s extra trading inside that window actually lost ₹7.15 lakh. The alleged gain came entirely from dumping a position that was already dead. Even here, the money was in the structure, not in the cleverness.

None of it.

Here is that same afternoon in one-minute index data — the data nearly every chart you look at is built from:

Fourteen flat minutes, then 218.48 points in a single bar. Not one of the three spikes shows up. Nor the drawdown, nor the cancellations, nor the 532 points the indicative index travelled while all of that was going on. The chart recorded one tidy step and carried on.

That is not a data error, and no vendor’s feed will show you anything different. An index is a calculated number; it cannot move until its stocks’ auction prices are published, so it sits still and then jumps. But it does mean the whole episode above — the thing a regulator needed forty-six pages to describe — is invisible on the instrument almost everyone is actually watching.

And then the next morning arrives, and that one bar starts doing damage in a place most people never look.

A bad closing price does not stay inside that day. It walks into the next one, inside your indicators.

This is the part I have not seen written up anywhere, and it is the part that matters if you hold positions overnight or run anything automated. It shows up in three places, and they get progressively worse.

These do not care what chart you are on. If you trade levels, this one is yours.

Every pivot formula used in India — classic floor pivots, Camarilla, Woodie, CPR — takes yesterday’s high, low and close and builds tomorrow’s levels out of them. On 13 August the high and the low were real. The close was not.

Camarilla spacing comes off the range, so the whole ladder simply slides: every single level — H1 to H4, L1 to L4 — sat 218.48 points too high for the next session. Classic pivots move less, but move unevenly: the central pivot shifted 72.83 points, R1 and S1 by 145.65.

Here is what that did on 14 August. Same high, same low, same formula. The only difference is which close you feed it.

Now read the two ladders against the day that actually happened. The Sensex opened at 77,903.43, fell to 77,684.37, and closed at 78,009.25.

• On the ladder everybody had, the market opened below L3 and then broke L4 at 77,830.54. In Camarilla, breaking L4 is the textbook breakout short — sell, and stay short.
• On the honest ladder, the market opened right at H1, never came near L4 (77,612.05), and closed above H3. Which is the textbook breakout long.

Your moving average shrugs it off in a session. Your ATR does not — and ATR is feeding half your toolkit.

On a time-based chart the print becomes one bar with a long body, and the clock does most of the cleaning up. On a 15-minute Sensex chart, a 20-period EMA computed with the print was about 20 points away from the same EMA computed without it at the next day’s open, and back within 2–3 points by that afternoon. RSI was a point or two out. Annoying, not fatal.

ATR is the one to watch, and this is the bit I think most people will miss. ATR does not only measure today’s range; it measures the distance from yesterday’s close into today’s range. A fake close manufactures a fake gap. On that 15-minute chart:

An inflated ATR for one session sounds harmless until you list what is standing on top of it. Supertrend. Keltner Channels. Chandelier stops. ATR trailing stops. ATR-based position sizing. Every one of them takes ATR and turns it into a distance — how far away your stop sits, how wide your channel is, how many lots you are allowed to take. If ATR is 21% too big at the open, all of those answers are wrong for the rest of the day, and nothing on your screen looks broken.

One honest note in the other direction: on the daily chart this barely registers. Daily ATR(14) came out at 596.68 with the print against 594.46 without — about two points. It is the intraday timeframes that carry it, because there the fake gap is large next to a normal bar.

A chart that only moves when price moves has no way of knowing that a move was not a move.

Noiseless charts — Renko, Point & Figure, range bars — ignore the clock completely and record only distance travelled. That is exactly what makes them quiet, and it is exactly the exposure here: hand them 218 points that nobody travelled, and they will faithfully record 218 points of travel.

That one bar becomes 7 bricks on a 0.04% Renko chart, or 14 boxes on a 0.02% P&F — on charts that print about forty box events in an entire session these days.

It is one thing to calculate that and another to look at it, so I built both. Below is the real Sensex P&F across that boundary at a 0.02% box — and beside it, the same chart from the same one-minute data with the auction window taken out. Same engine, same settings, same sessions. That one bar is the only difference.

Look at the sixth column in each panel. Same column, same afternoon.

On the left it runs eighteen boxes and finishes six boxes above the highest point of the previous five columns — on a chart type where breaking above prior column highs is the buy signal. Fourteen of those eighteen boxes carry one timestamp: 13 August, 15:29.

On the right, the same column stops after four boxes, eight boxes below those old highs. No breakout. Nothing at all.

Now the indicators. I ran the same Sensex Renko chart at a 0.04% brick twice — once with the auction bar, once without — and read everything off at the close of 13 August. Same engine, same settings, same data. One bar apart.

Take that table one line at a time, because there is more in it than “the numbers moved a bit”.

• Supertrend flipped. Not drifted — flipped. Same chart, same settings, same data, and one version says you are long while the other says you are short. If you run anything automated off a Renko Supertrend, that is not an input to the decision. That is the decision.
• RSI is 17 points higher — the difference between “nothing is happening” and “this is getting strong”. Anybody using RSI above 50 as a trend filter had it switched on by an auction print.
• Bollinger Bands are 31% wider for no reason whatsoever. If you trade band expansion as a breakout, you were just told the market had opened up.
• The 20-EMA is 61 points away from where it belongs — on a chart where 61 points is two full bricks.

And then the two lines I did not expect, which turned out to be the most interesting part of the whole exercise.

The reason is worth a sentence, because it is peculiar to these charts. On a brick chart, true range is measured brick to brick. When price is chopping, each new brick sits on the far side of the last one and the true range is two bricks wide. When price runs straight, every brick is one brick from the last — the smallest true range the chart can produce. So a burst of seven bricks in one direction, out of one bar, looks like the smoothest trend of the month. ATR obediently drops, Keltner tightens around it, and Supertrend’s band moves close enough to price to flip.

Which leaves you with this: on a candlestick chart the print inflates your volatility reading, and on a noiseless chart it deflates it. Both are wrong. They are wrong in opposite directions, so you cannot even carry one intuition across from the other.

And how long does it last?

Longer than you would guess. Here is 14 August — a normal session, market open, trading properly — comparing the two charts as the day went on:

• Supertrend still disagreed until 10:30 am — an hour and a quarter into the session, and fourteen bricks deep.
• Bollinger width was still 30–90% wider right through the morning.
• The two charts only came back into line at 1:46 pm — four and a half hours into the next session.

The reason is the thing I keep coming back to: on a movement-based chart, a lookback period counts bricks, not minutes. A 20-period average is twenty box events. It has no clock. It cannot know that fourteen of its twenty inputs were manufactured in ninety seconds and the other six took two days.

The clock that clears a phantom print is market movement. There is a lot less of it now.

This is where it stops being a one-day anecdote for me. Since the auction started on 3 August, everything has gone still. The twelve sessions since, against June and July:

That last row is what ties the whole note together. The phantom is fourteen bricks. It used to be a fifth of a day’s movement. Now it is a third — and there is 40% less real movement arriving to push it out of your indicator’s window.

Did the auction cause the quiet? Nobody can say that yet, and I am not going to pretend otherwise. Twelve sessions is nothing, August is usually slow, and plenty of people are simply sitting out while they work out how the new close behaves. Ask me in December.

But the two effects stack whatever the cause, and they stack in the same direction: a bigger footprint from every bad print, and less real movement to clear it.

I measured this rather than assuming it. On one index, a trend reading computed with the auction print disagreed with the same reading computed without it on every tradable minute of two straight sessions. Across the first ten sessions of the new regime, roughly a quarter of the decisions my own automated rules took landed on a state that differed from the clean series.

Not all of those would have gone the other way — a different state often produces the same action. But it was uneven: some rule sets were barely touched, one was affected on more than 40% of its decisions. So it does not simply add noise. It quietly re-ranks strategies against one another, which is the one thing a comparison period exists to prevent.

The auction touches three different things, and it is easy to fix one and assume you are done.

The auction is not the villain, and an interim order is not a verdict.

Closing auctions exist to make closing prices harder to push around, not easier, and most major markets run one. The order also makes a point that deserves more attention than it will get: under the old volume-weighted method this would have been considerably harder to spot. The auction is what made it visible — every order, every cancellation, timestamped to the second, inside one window. SEBI caught it in routine surveillance and acted within six days. That is the system working, not failing.

The order is ex-parte and interim. The two entities have not been heard yet, its findings are expressly prima facie, and nothing in it is a finding of guilt. They may well have answers. I have described it here as the order describes it, and linked it so anyone can read the whole thing themselves.

And a caveat against my own argument, because it belongs here. When I ran my own results across this period properly, the closing auction was a real but minority contributor — the volatility regime explained far more of what changed than the auction did. Roughly a fifth to a third of trend flips on one index traced back to auction-born chart events; on another, under a tenth. Anyone going looking for this in their own numbers should expect something smaller than the story suggests.