High-Frequency Trading in India: Colocation, Latency, and What It Actually Means for You

Two myths, and the single mistake underneath them

The first myth is that high-frequency trading is why the retail account loses. In this version the market is a rigged machine, orders are picked off the moment they are sent, and the outcome was decided before the trader clicked. The second myth is the mirror image: that speed is a technique, that the difference between the professional and the amateur is a better computer and a colocated server somewhere, and that with enough effort the gap could be closed. The two beliefs are usually held by different people, and occasionally by the same person on different days.

Both come from one mistake, which is treating speed as a single continuous dial that every participant sits somewhere on. It is not a dial. It is a discrete business with a fixed cost of entry, a minimum viable volume, and a product that has almost nothing in common with what a person trading a swing position is doing. Once you model the cost of entry and the volume it demands, the conspiracy version and the aspirational version fail for the same reason: neither of them is describing an activity a retail participant is, or could be, engaged in.

The concern behind the first myth is not paranoid, and it deserves a real number rather than a reassurance. SEBI's press release of 23 September 2024 reported that in the financial year 2024, roughly 96 to 97 percent of the gross profits earned in equity derivatives by proprietary traders and foreign portfolio investors came from algorithmic trading, while over the three years to that point 93 percent of more than one crore individual traders lost money in the same segment, with aggregate losses above 1.8 lakh crore rupees. Something real is being described there. The question this page asks is whether the something is speed.

It is worth being precise about a word. Algorithmic is a much wider category than high-frequency. An algorithm that rebalances a portfolio once a month is algorithmic. So is one that works a large order across an afternoon to avoid moving the price. So is a rule-based system a person runs from a laptop. High-frequency trading is one narrow tenant inside that building, and the habit of using the two words interchangeably is precisely what makes the subject impossible to think about clearly. Our guide to algorithmic trading in India maps the layers of automation and shows where a retail account actually sits among them. This page goes underneath one of those layers and takes it apart.

What actually sits between the signal and the exchange

The first correction is about when the clock starts. Most people picture the race beginning when they decide to act. It does not. It begins when the event happens at the matching engine, because that is the moment the information exists. Everything between that instant and your order arriving back at the engine is your latency, and most of it happens before you have seen anything at all.

Modelled component by component, the two paths look like this. The colocated participant's server sits inside the exchange's own facility, so the outbound leg is a cross-connect of tens of metres, the feed is decoded on their own hardware, the decision is a lookup rather than a calculation, and the order returns down the same short wire. The retail participant's path has the same shape and none of the same magnitudes: the event has to reach the broker, the broker has to publish it in a periodic snapshot, the snapshot has to cross the consumer internet, the client has to process it, the order has to be submitted over an ordinary web request, the broker has to run its risk checks, and only then does the order reach the exchange.

The composition is more interesting than the total. On this model, 79 percent of the retail path is spent waiting for a snapshot that has not been published yet, and only 3 percent of it is the participant's own code running. That single fact reorganises the whole subject. The retail handicap is not principally a connectivity problem or a programming problem. It is a consequence of the shape of the data product that reaches a retail screen, which is a periodic picture of the book rather than a stream of every event in it.

The snapshot interval matters enough to say precisely where the number comes from, because this is an area where folklore is abundant. One Indian exchange publishes its interval: the BSE Direct NFCAST manual, version 5.0 dated 28 May 2025, states that its market picture message is not sent on every update to the order book but on a defined snapshot interval, currently 800 milliseconds. The other does not publish one anywhere this desk could find, and a widely repeated one-second refresh figure attributed to it is not substantiable in any primary source. The model here therefore anchors on the one interval that is officially published and takes the mean wait as half of it. It does not assume an interval for an exchange that has not stated one.

Which leads to the test that disposes of the second myth. Suppose a retail participant spent money on everything they personally control: the fastest machine available, a dedicated line, the client rewritten in a compiled language, the submission path optimised as far as it can go. Those components total about 80 milliseconds, or 16 percent of the path. Drive all of them to zero, which is not possible, and about 426 milliseconds remain. That residue is still roughly 7,751 times the colocated path. The upgrade that a retail participant can actually buy improves the number by a sixth and changes the competitive position not at all.

What the microseconds actually buy

A latency advantage is only worth something when the opportunity it chases has two properties. It has to be fleeting, so that it disappears before a slower participant can reach it, and it has to be winner-take-all, so that arriving second is worth nothing rather than slightly less. A stale quote left behind by a moving market has both. So does a queue position at a price that is about to become the best in the book. So does a price discrepancy between two venues that will be closed by whoever gets there first.

For that kind of opportunity, the right model is a race. Give each participant a latency that is not fixed but drawn from a distribution, because real latency has a hard floor set by physics and a long tail set by everything else, and see who arrives first. The result is brutal in a way that is worth seeing rather than being told.

Two things fall out of the simulation. The first is how narrow the contested zone is. A challenger 17 microseconds behind the leader wins one race in ten. A challenger 36 microseconds behind wins one in a hundred. The entire competitive question is settled inside a window smaller than the time it takes a keyboard to register a keystroke, which is why the firms in that zone spend real money on differences most people would consider immeasurable.

The second is what happens to everyone outside that window. Of the five modelled tiers, three won nothing at all across four hundred thousand contests: a leased-line member in the same city, a hosted server on a broker network, and a retail participant on consumer internet. Not a low win rate. Zero. This is the honest content of the first myth, and it should be stated plainly rather than softened: in a pure speed race, a retail participant is not losing, they are absent. There is no configuration, no upgrade and no clever workaround that changes that.

But notice how much the conclusion depends on those two qualifying properties. If the opportunity is not fleeting, arriving later costs you a little rather than everything. If it is not winner-take-all, there is room for more than one participant to be right. Almost nothing a retail trader does has either property. A position held for three days is not a race, and the person who bought the same instrument four hundred milliseconds earlier has not taken anything from you. The rest of this page is about pricing that distinction instead of asserting it.

The disadvantage collapses as you slow down

To price a latency handicap you need two quantities. The first is what the delay costs. While your order is in flight the price is moving, and because you are systematically the slower side, the moves you transact into are adversely selected: the fast participants have already taken the ones that were favourable. Modelled as diffusion, the expected adverse move over a delay grows with the square root of that delay.

The second is what you are trying to capture. A strategy with genuine skill predicts some fraction of the movement over its own horizon, and that movement also grows with the square root of the horizon. Dividing one by the other, the volatility of the instrument cancels out completely and what remains is startlingly simple: the latency handicap, expressed as a share of your edge, is proportional to the square root of your delay divided by your holding period. Two timescales, nothing else.

The numbers along that line are the argument of this page. At a one-second holding period the handicap is about 1,138 percent of the modelled edge, meaning the delay costs more than eleven times everything the strategy was trying to earn. At one minute it is about 147 percent, still fatal. The crossing point, where the handicap exactly consumes the edge, sits at about 2.2 minutes. At an hour it is about 19 percent, at a single session about 7.6 percent, at a week about 3.4 percent, and at a month about 1.7 percent.

Now the absolute figure, which is the uncomfortable one. Converted into money at the stated volatility, the entire latency gap between a colocated participant and a retail one prices out at about 0.61 basis points a round trip. The all-in friction on an ordinary retail round trip, modelled on the companion page about the gap between a backtest and a live account, came to about 44.47 basis points. The speed handicap is roughly one seventy-third of a cost the same trader is already paying, mostly without itemising it. On one lakh rupees of notional that is about six rupees against about four hundred and forty five, both illustrative.

This is the result that most needed publishing, because it complicates the story in both directions. It undercuts the conspiracy: whatever is emptying the retail account, about six paise per thousand rupees traded is not it. It also undercuts any comfortable reading, because the same arithmetic says that at horizons below a couple of minutes the handicap is not small at all, it is larger than the whole edge. A trader who has quietly drifted into scalping has walked into the one region where the number bites, and the number does not care that the drift was gradual.

The assumptions deserve to be argued with, and they are stated so that they can be. Two of them carry the level of the curve: how much of the diffusion during your delay is genuinely adverse, taken here as the mean absolute move, and how much of the horizon's volatility a skilled signal captures, taken here as 5 percent, which is generous. Move either and the whole line slides up or down. What neither of them can move is the slope. The handicap falls with the square root of the holding period whatever you assume about skill, because the volatility term cancels before those parameters are ever applied. The shape is structural, and the shape is the finding.

Colocation, and what it is repeatedly accused of

Colocation is rack space rented inside or immediately beside the exchange's own data centre. A member's trading server sits metres from the matching engine rather than across a city, connected by a dedicated cross-connect rather than by a shared network. Two things come with that. The obvious one is distance, which is what the microsecond figures in the first figure measure. The less obvious one is determinism: a dedicated short path has a narrow spread of outcomes, and in a contest decided by tens of microseconds, a predictable latency is worth as much as a low one.

What colocation does not buy is the data, and this is where the popular account of the subject is not merely imprecise but backwards. It is widely repeated that the full order-by-order feed is a premium product available only to firms inside the colocation facility. The SEBI master circular for stock exchanges and clearing corporations says the opposite. Its terms require exchanges to provide tick-by-tick feeds to all trading members free of cost, subject to trading members creating the necessary infrastructure for receiving and processing it, and the clause states its own purpose: to create a more level playing field among the different types of market participant. The same instrument defines the feed as a detailed view of the entire order book covering the addition, modification and cancellation of orders as well as trades.

There is an important nuance that the correction must not be allowed to swallow. The entitlement runs to trading members, meaning brokers and other registered members, not directly to their individual clients. A retail person cannot write to an exchange and ask for the feed. So the practical experience of a retail participant is still that they receive a snapshot rather than an event stream, and everything that follows from that, including the fact that the aggressor side of a trade has to be inferred rather than read. Our page on order flow trading in India works through what that inference can and cannot support. The point of the correction is not that retail has the data. It is that the reason retail does not have it is architectural and commercial, not a privilege sold to the fastest bidder.

The regulatory framing around access is consistent with that reading. The same master circular requires colocation and proximity hosting to be provided in a fair, transparent and equitable manner, and requires that participants have fair and equal access to facilities and data feeds. A 2018 SEBI board memorandum, in the discussion that preceded the current rules, described limited tick-by-tick access as creating disparity and inequality in access to data, and therefore an unfair market advantage. That is regulatory language aimed at removing a gate, not at defending one.

Speed is a business with a break-even, not a technique

The clearest way to understand why high-frequency trading is not available to a retail participant has nothing to do with skill or access. It is arithmetic. The per-trade edge at that timescale is tiny by construction: at a one-second horizon, on the same model used above, it is about 0.053 basis points. On five lakh rupees of notional that is about 2.67 rupees a round trip, illustrative. Everything else follows from having to earn a living out of that.

Set against it an annual cost base. Rack space and cross-connects, exchange membership and clearing, low-latency hardware written down over its useful life, redundant links and a disaster recovery site, the engineering and research staff who build and maintain the thing, and the risk, compliance and audit systems without which none of it is permitted to run. Modelled illustratively, that comes to about 4.20 crore rupees a year, of which the largest line by a wide margin is people rather than machinery.

Dividing one by the other gives the number that settles the question: about 64,024 round trips a day, roughly 15.7 million a year, or on the order of 3,201 crore rupees of notional turnover a day, purely to cover the fixed cost before anything is earned. That is not a trader. That is a plant. It has a utilisation rate, a maintenance schedule and a payroll, and the reason so few firms operate at that tier is not secrecy but capital intensity.

The obvious objection is that the edge assumption is doing the work, so the model was rerun against it. Multiply the per-trade edge by ten, which would be a heroic improvement, and break-even is still about 6,402 round trips a day. Multiply it by a hundred and it is about 640. Shrink the cost base to one crore rupees, well under a quarter of the modelled figure, and it is still about 15,244 round trips a day. The conclusion is not sensitive to the assumption, which is the only reason it is worth stating.

Run the same arithmetic backwards and the retail position becomes concrete. An active retail account might do twenty round trips a day. To carry the same infrastructure at that volume you would need about 171 basis points of net edge on every single trade, which is an absurdity, and it is an absurdity by three orders of magnitude rather than by a little. The two activities are not at different points on one scale. They are different businesses that happen to share a venue.

Where the retail participant is genuinely disadvantaged

Having spent several sections showing that latency is not the retail participant's main problem, it would be dishonest to leave the impression that there is no problem. The disadvantages are real, they are large, and they are almost all somewhere else. Setting them out in order of size is more useful than any reassurance.

The largest is cost. On the companion model, an ordinary retail round trip surrenders roughly 44.47 basis points to spread, slippage, market impact, rejects, partial fills and the statutory charge stack. Roughly half of that is taxes and fees that appear on a contract note, and roughly half is friction that no document itemises. Against a gross edge of the order of thirty basis points, that is not a deduction, it is the dominant term. Anyone worrying about milliseconds while ignoring this is optimising the wrong end of the sequence by a factor of about seventy.

The second is the data itself. A retail platform shows a snapshot of the book, which means the aggressor side of any trade has to be inferred from the trade price against the last quote seen, rather than read from a field. Inference is not measurement, and it is wrong some of the time in ways that are invisible when it is wrong. Worse, the total buy and total sell quantity fields that retail packets do carry are resting pending orders, not aggressive volume, and they are routinely relabelled as buyer or seller pressure by tools that should know better.

The third is process, and it is the one that is actually within reach. An institutional desk writes the rule down before it looks at the outcome, validates it on data it did not use to build it, sizes positions from a measured distribution rather than a hope, and has a risk system that turns the thing off without asking permission. None of that requires a data centre. It requires the discipline to do the boring parts in the right order, which is genuinely difficult and genuinely available. If the practical question is what an automated retail setup actually involves, our walk-through of how to start algorithmic trading in India covers the ground without pretending the infrastructure gap is closeable.

The fourth disadvantage is structural and is rarely named: capital and diversification. A desk running forty uncorrelated rules can afford for eleven of them to be losing this quarter. A person running one rule experiences the same statistical event as a personal crisis and abandons the rule at the worst possible moment. That is not a technology gap. It is a portfolio-construction gap, and it explains far more of the outcome distribution than any measurement in microseconds.

Which brings the SEBI figures back into focus with the right interpretation attached. That 96 to 97 percent of proprietary and foreign-investor derivatives profits came from algorithmic trading tells you where the profits sat. It does not tell you that speed put them there, and the modelling on this page argues fairly strongly that speed did not. What algorithmic participation actually buys, at every timescale including the slow ones, is consistency of execution, absence of hesitation, systematic sizing and a cost structure negotiated rather than accepted. Those are the transferable parts.

What follows from all of this

The conclusion is not that a retail participant cannot win. It is narrower and more useful: you cannot win at their game, so the sensible response is to stop playing on their timescale. That sounds like resignation and it is the opposite. It is the identification of an axis on which the disadvantage is four orders of magnitude and permanent, followed by the observation that nothing obliges you to compete on it.

Concretely, four things follow. The first is a floor on the holding period. On the model here the speed handicap consumes the entire edge below about two minutes, is still meaningful in single digits of percent across a session, and becomes a rounding error over a week or more. Anything from an hour upward is defensible and anything from a session upward is comfortable. A trader who cannot state their intended holding period has not made this decision, they have inherited it from whatever their screen refreshes at.

The second is a reallocation of attention. The charge stack is about seventy times the latency handicap on the same trade, and unlike latency it is partly controllable: through order type, through the choice of when to be patient and when to cross, through trading less often, and through sizing that does not force you into the worst part of the book. Every hour spent shaving milliseconds is an hour not spent on the line item that is seventy times larger.

The third is a filter on strategy selection. Prefer rules whose edge does not depend on being first. An edge that comes from being faster is unavailable. An edge that comes from being more patient, from holding through a period that institutional mandates make awkward, from a position size too small for a desk to bother with, or from a piece of structure that takes days rather than milliseconds to resolve, is not obviously unavailable, and is at least being competed for on terms where you are present.

The fourth is the one worth ending on, because it is the genuine asymmetry. A retail participant has no mandate, no benchmark, no quarterly reporting obligation, no capacity constraint and no requirement to be invested. The freedom to hold cash for two months and explain it to nobody is a real structural advantage, and it is unavailable to almost every professional participant in the market. It is worth considerably more than a millisecond, and unlike a millisecond it costs nothing to acquire.

None of this is technically demanding. It is a matter of choosing a timescale deliberately, measuring the costs that actually apply at that timescale, and declining an invitation to compete where the outcome is already determined. Those are decisions rather than techniques, which is both why they are available to anyone and why they are so often skipped. They are also, in substance, what a serious programme in systematic trading spends its time on, and if the arithmetic on this page was the interesting part rather than the tedious part, that is the method we teach.

Every computed figure comes from one deterministic model, seeded so that it reproduces identically on each run. Four pieces make it up. A latency budget assembles the path from a market event to a reacting order for a colocated participant and a retail one, component by component, with each component an illustrative order of magnitude for a step that genuinely exists rather than a measurement of any venue, broker or product. A race model draws each participant's latency from a lognormal distribution, because latency has a hard floor and a long tail, and settles four hundred thousand contests between five access tiers. A cost model prices the delay as the mean absolute price diffusion over it and compares that with the edge a signal of stated skill captures over its horizon. A break-even model divides an illustrative annual fixed cost by the per-trade edge to derive the round-trip volume that carries it.

The market assumptions are 1.6 percent daily volatility over a 22,500 second session, an adverse-selection coefficient equal to the mean absolute standardised move, and a signal capturing 5 percent of its horizon's volatility. The first two set the level of the collapse curve and are open to argument; the exponent of minus one half is not affected by them, because the volatility term cancels before either is applied. The one snapshot interval used is the 800 milliseconds published in the BSE Direct NFCAST manual version 5.0, dated 28 May 2025; no interval is assumed for an exchange that publishes none.

Regulatory statements are drawn from the SEBI master circular for stock exchanges and clearing corporations and from SEBI's press release of 23 September 2024. All computed results are illustrative and simulated. They are not a track record, not a forecast and not an indication of what any strategy or any participant would produce in a live account. Every rupee figure on the page is illustrative. The purpose of the exercise is to show the relationship between two timescales, which is a property of the arithmetic rather than of any particular market.

• What the depth window is actually showing you →
• The eight ways a backtest can lie →
• Building a trading system →
• Trading on a timescale you can actually compete on →