Guide · Execution mechanics · Updated 2026-09-20

The hidden quantity keeps its place in the queue. The sliced order throws that place away.

Market figures are computed from the security bhavcopy and daily index files cached in this repository, over 20 sessions to 18 September 2026 for turnover and 250 sessions to 18 September 2026 for volatility. Model figures are derived here from inputs printed alongside them, and the illustrative parameters are labelled wherever they appear.

Two different mechanisms wear the same name

Almost every explanation of iceberg orders in an Indian context describes one thing and then reasons about another. The confusion is worth clearing first, because the entire cost of the technique sits in the difference.

The first mechanism is an order condition at the exchange. You enter an ordinary limit order for the full quantity and attach a disclosed quantity: an order for 5,000 disclosing 500 shows 500 in the depth window, and when that 500 has traded another 500 is released to the display automatically. There is one order in the book throughout, its full quantity committed and matchable against incoming flow whether or not the display has caught up. Nothing is re-entered, so nothing is re-stamped.

The second mechanism is a facility of your trading member's system. It takes a large order and slices it into legs, sends a child order for the first leg, waits until that leg is fully traded, then sends the next. Ten legs are ten separate orders. Each one arrives at the exchange at its own instant and is ranked by that instant.

Read the third and fourth rows together. The exchange condition is a display truncation over a committed order; the member facility is a genuine withholding of quantity from the market. One conceals your size while keeping your claim on the front of the line. The other trades the claim away to withhold the size.

Each fresh leg starts at the back of the line

Orders at one price form a queue ranked by the instant the system accepted them, and the engine matches from the front. That rule, and what a modification does to it, is set out in full in the guide to queue position and price time priority. One consequence of it decides the economics of slicing, and it is the consequence most explanations omit entirely.

A new child order cannot inherit a place in the line, because it has never been in that line. The second leg joins the back, and by the time it does the level has typically refilled, since other participants have been posting there for as long as your first leg took to clear.

That gives an arithmetic anybody can do. Take a price level with 10,000 shares resting ahead of you, replenishing to the same depth after each leg clears. A fully displayed order for 5,000 shares needs 10,000 plus 5,000 to print at that price before it is done. An order cut into N legs rejoins the back N times, so it needs N times 10,000, plus 5,000.

At ten legs the order needs 7.0 times as much volume to print at that price as the same order displayed in full. That is the cost of the concealment, in the only currency a queue deals in. Now look at what drives it: the extra volume is 10,000 shares for every leg after the first, and 10,000 is the depth of the level, not the size of your order. The penalty is set by the queue you keep rejoining.

This is the uncomfortable shape of the trade. Concealment gets more expensive exactly as the book gets deeper, and a deep book is the case in which you needed it least, because your order is a smaller part of it. Both considerations point the same way, and it is not the way most explanations point.

There is a second cost in the sequencing that is easy to miss because nothing about it looks like a cost. A leg is sent only when the previous one has filled completely, so at no instant is more than one slice working. A leg that takes a partial fill and then sits there stalls the entire order, and the residual you were concealing is not in the market at all while that happens. If the price leaves, you have neither the fill nor the queue position, and the choice is to chase or abandon. Chasing is where the impact saving goes.

One Indian specific closes off the end of the day. Neither mechanism has any expression in the closing auction for securities on which derivatives are available: orders carrying a disclosed quantity are not accepted in that session and are not carried into it from continuous trading. The last stretch of the session is a different problem, set out in the guide to stop loss orders and what a closing auction does to them.

A level that refills to the same number is a signature

Concealment in a central order book is never concealment of activity. The depth window publishes aggregate quantity at each price without identifying whose it is, so what you are hiding is a number, and you are hiding it from people whose job is to watch that number change. Replenishment at a price is completely ordinary. Replenishment to exactly the same quantity, repeatedly, immediately after each print that takes it out, is not.

The inference is a counting argument. Suppose ordinary replenishment at a level lands on one of 8 plausible round sizes. Then the chance that the display returns to precisely the same number k times running, by coincidence, is one over 8 to the power k minus one.

Four identical refills is already a one in 512 coincidence, and a ten leg order produces nine of them. The two mechanisms also leak differently. The exchange condition refreshes its display in the same instant as the fill, so the tell is a print larger than the quantity that was showing. A member side slicer must observe the fill and then send the next order, so the tell is a small, repeatable gap between the level emptying and returning, and the length of that gap is a fingerprint of the system placing the orders.

What survives is the part worth having: the total. An observer can infer that a worked order is present and can read the slice size straight off the book, but nothing there says whether two more legs are coming or two hundred. That is the concealment you are buying, and it is much narrower than the word "hidden" suggests.

The obvious defence is to vary the slice size so the number does not repeat. It costs something in each mechanism. On the exchange condition the displayed quantity can be modified, but an increase in what the book can see forfeits the time priority of the whole order, so the defence is free in one direction only. On a member facility varied legs are simply a configuration, which is why detecting institutional working orders is a contest between systems rather than a solved problem.

Walking the book is the cost you are actually avoiding

The usual comparison is against the wrong thing. The alternative to patience is not a mysterious market impact; it is crossing the spread and consuming the ladder in front of you, one price at a time, until your quantity is filled. That cost is fully determined by the shape of the book and can be computed exactly for any assumed shape. Below are three illustrative ladders on a stock near ₹500, where the tick is 0.05, taken immediately at the touch.

Three things in that table matter more than the levels. First, none of these exponents is near one half. Over the sizes a retail order actually is, the cost of impatience rises about in proportion to size on a book that thickens away from the touch, faster on a flat book, and explosively on a book that thins, where 117 basis points for 8,000 shares is less an impact estimate than a warning about the security.

Second, the exponent is a property of the ladder, not a law of markets. It falls out of how depth accumulates as you move away from the touch, and that varies by security, by time of day and by how much resting interest is displayed rather than concealed.

Third, on the same thickening ladder, push the size up to institutional scale and the fitted exponent falls to 0.55. Same book, two regimes. At small sizes you pay for the first few prices; at large sizes you pay for the accumulated shape of the whole book, and that is where a square root appears. Which is the most useful thing to know about the model in the next section: the regime it describes is not the regime a retail order lives in.

The square root is a model, and these are its assumptions

The relation almost always quoted is that the expected impact of a worked order scales with the square root of its size relative to volume. Written out, with every term named:

Expected impact, in basis points, equals Y times sigma times the square root of Q over V, where Q is the quantity of the whole worked order, V is a typical day's volume in the same security, sigma is that security's daily volatility in basis points, and Y is a dimensionless constant.

Four things about that expression deserve more attention than they usually get.

Y is not derived from anything. It is the constant that makes a fitted curve pass through observed data, calibrated in the published literature to somewhere between one half and one. It is not measured here and no figure on this page pretends it is. Every impact number below is printed at both ends of that range, because the choice moves the answer by a factor of two, and a model with that much freedom in it is telling you an order of magnitude.

Sigma is the security's volatility, not the market's. This is the input most often supplied wrongly, and the error is systematic rather than random, because an index is a portfolio and its volatility is lower than its members' by construction.

Using the index figure where a single name belongs understates the model's own output by a factor of 2.2. The figures below use 1.88 per cent, the median for the 100 most traded securities, because that is the population in which a concealment decision is usually made. A different estimator on the same securities gives 1.33 per cent, so even the input has a range of its own.

It is a statement about a whole worked order, not about each slice. This is where published arithmetic most often goes wrong. If you treat each slice as an independent order with its own square root cost and add them up, you get a total that grows faster than the unsliced order, and the error is then patched by dividing by the square root of the number of slices to reach an answer that looks right. Both steps are wrong. The relation takes the size of the parent order and the volume of the day. Slicing is how you realise that cost, not a way of reducing it.

It is an average, not a price. It describes the central tendency of many executions. For any one order the realised cost is dominated by what else happened to be in the market that hour.

Splitting helps sublinearly, and the denominator is not liquidity

Because the relation is concave, spreading the same quantity across more sessions lowers the modelled cost. The arithmetic is direct: a quantity spread evenly over D sessions trades one Dth of itself against each day's volume, so the cost of the whole order falls with the square root of D, and nothing faster.

The last column is the honest summary of what splitting buys. The first extra session saves 29 per cent of the modelled impact, going from two sessions to three saves another 13 points, and going from five to ten saves 13. A tenfold increase in exposure buys a 68 per cent reduction, and the next section prices the exposure.

There is also a problem with V that nothing in the model acknowledges. A day's volume is not a day's liquidity, because the same shares trade repeatedly within the session. That is measurable directly: across the 20 sessions to 18 September 2026, the volume weighted delivery percentage of the traded list was 37.0 per cent, so roughly 63 per cent of traded volume never settled into anybody's account. It was the same stock changing hands again. For a model whose whole content is your size divided by that number, this runs in one direction only: the denominator is inflated by activity that supplies no inventory, so the participation rate you compute is flattering and the impact it produces is low. Nothing here says how much to adjust by, because nobody has measured that. It does say which way the error runs.

The related failure is where the model stops applying at all. It assumes the book refills between your executions from flow that has nothing to do with you, and that your order is a modest fraction of the day. Both fail together once the order is large: the flow that refills the book starts reacting to you, and the relation was never fitted in that range. Past roughly a tenth of a day's volume, treat the output as a lower bound. Past a full day's volume, the model is not describing the situation. A cost model has a domain, and using one outside it is the failure mode that modelling transaction costs in a backtest deals with in detail.

Minimise the impact and you have bought the timing risk

Working an order slowly reduces its footprint and lengthens the window in which the price can move away from you. Both costs are real, and the second is why there is a decision here rather than a rule.

The timing term can be derived rather than assumed. If an order is worked evenly across a window of D sessions and the price follows a random walk with daily volatility sigma, the average price achieved has variance equal to sigma squared times D divided by three. So one standard deviation of timing uncertainty is sigma times the square root of D over three, rising with the square root of the horizon exactly as impact falls with it.

Add the two, weighting one standard deviation of timing uncertainty by a factor saying how much you dislike price uncertainty relative to certain cost, and the total has an interior minimum. At that minimum the two terms are exactly equal, which is a check needing no computation: if the price uncertainty you are accepting is much smaller than the impact you are avoiding, the order is being worked too fast, and the other way round.

The pattern in that table is the practical result of the whole subject. Under the stronger weighting the optimal horizon does not reach a full session until the order is about a third of a day's volume; under the moderate one, not until roughly 8 per cent of a day. Below those thresholds the model's own answer is measured in minutes, not days, and an order worked across a week is not being executed carefully. It is carrying days of price risk to save basis points of impact it did not have. Above those thresholds the arithmetic reverses and multi-session working is straightforwardly right, which is why institutional desks do it and why the technique exists at all.

The scale at which any of this matters, measured against real sessions

All of the above turns on one ratio: your order against a day's volume in that security. Indian listed securities differ so widely on that denominator that the same rupee order lands in two different regimes, and this is the figure that decides whether the subject applies to you.

Across the 20 sessions to 18 September 2026, a median of 2,638 securities in the main equity series traded each day. The median one turned over ₹2.91 crore. The median of the hundred most traded turned over ₹432 crore, about 149 times as much. The median value of a single trade was ₹10,452 across the whole list and ₹46,750 among the hundred most traded.

Read the right hand pair first. A ₹1,00,000 order in one of the hundred most traded securities is 0.0023 per cent of that day, and the model puts its impact at 0.90 basis points, about ₹9. That is smaller than one tick at most price levels and very much smaller than the spread. There is nothing there for a concealment technique to recover, and member side slicing spends real queue position trying. Even a ₹1 crore order in such a security is 0.23 per cent of the day and around 9.0 basis points. For most individual participants in the liquid part of the market, the honest answer is that iceberg orders do not matter, and the technique is a cost rather than a saving.

Now read the left hand pair. In the median traded security the same ₹1,00,000 is 0.34 per cent of the day, and ₹25,00,000 is 8.6 per cent, past the point where the model is a lower bound rather than an estimate. A ₹1,00,000 order is 1 per cent or more of the day's turnover in about 39 per cent of the securities that traded, a figure that sat between 37 and 40 per cent on every one of the 20 sessions examined.

That is the result worth keeping. The variable is the security, not the size of the account. A perfectly ordinary order is invisible in one name and a material share of a day in another, and roughly two fifths of the traded list is the second kind. Anybody whose interest in this subject is real is almost certainly there rather than in the liquid hundred, and guidance written for the liquid hundred is exactly backwards for them.

Where the decision actually sits

Concealment in a queue is not free, and what it costs is your place in the line, so the only question is which you need less. Three checks settle it, each answerable before the order is entered. Is the order a material share of the day in that security, computed against real turnover rather than assumed. If it is a fraction of a per cent, stop here: there is no impact to manage and nothing to buy with your queue position. Is the level deep at the price you want, because depth sets the cost of rejoining, and a deep level makes slicing expensive at the same time as it makes slicing unnecessary. Is the real alternative a market order that walks the ladder, because that is the cost the technique genuinely avoids, and it is steeper than the square root relation at every size a retail order takes.

If the answers point to concealing, prefer the mechanism that does not charge you for it. A disclosed quantity on a single order hides the number while keeping the claim. Sliced child orders hide the residual and pay for it in queue position, every leg, whether or not the price cooperates.

None of this produces a rule, and that is the point. It produces a pair of costs that move in opposite directions, a model with a stated domain and a factor of two of freedom in its constant, and a real distribution of Indian turnover that decides which regime an order is in. Execution is the part of a method where an approximate model applied with judgement beats an exact rule applied everywhere, and the judgement is learnable in the same way the arithmetic is.