HFT: what a high-frequency trader pays for, and why speed is infrastructure

High-frequency trading is not a "fast version" of ordinary trading but a separate business in which profit comes from priority in the order queue rather than from forecasting the direction of the price. An HFT participant holds a position for fractions of a second, closes it in the same order book and earns on the difference between the bid and the ask, on price discrepancies between venues, or by getting ahead of slower orders. This approach is out of reach for a private trader, not because they lack knowledge, but because the cost of entry — communication links, servers placed next to the exchange's matching engine, a direct market data feed, hardware processing — is on the scale of a business, not a brokerage account. It is worth understanding this before buying a "scalping" robot from a vendor who promises HFT.

What a high-frequency algorithm actually does

Behind the abbreviation lies not a single strategy but a family of strategies united by their holding period.

Market making: the algorithm simultaneously places a buy order and a sell order, capturing the spread. Profit comes from turnover, not from price movement; the main enemy is not a forecasting error but adverse selection, when the counterparty learns of news earlier and hits your order just before the price jumps.

Latency arbitrage: the price has changed on the leading venue or in the leading instrument, but the dependent quotes have not yet been updated. The winner is whoever manages to take out the stale order before its owner cancels it.

Cross-market and statistical arbitrage over a short horizon: a divergence between a stock and the futures contract on it, between a security and a basket, between two historically related instruments. The gap closes quickly, so it has to be caught quickly.

What they all have in common is that profit per trade is small, the number of trades is large, and any extra cost wipes out the entire model. Hence the defining feature: HFT is sensitive not to the accuracy of the signal but to the cost of execution.

Speed is not code, it is the link and the location

Optimising the algorithm pays off only as long as computation remains the bottleneck. In HFT the bottleneck is the path the signal travels.

Co-location: the participant's server is physically housed in the exchange's data centre so that the length of the optical path does not create a lag. A direct market data feed instead of an aggregated one: the exchange feed arrives raw, without intermediate normalisation on the broker's side. Hardware processing: parsing the packet and generating the order are done by a chip rather than by an operating system with its scheduler. A CPU core dedicated to the handler, power saving and network interrupts switched off — this is the part of the work that never appears in the description of a robot offered for sale.

That is why the question "which language is faster" is secondary for HFT, and a promise of high-frequency trading through an ordinary brokerage terminal over a home internet connection cannot be fulfilled technically. The stream of quotes you see in the terminal or in the overview of Russian stocks arrives, by design, later than the exchange's own feed.

Price–time priority and the cost of cancellation

The mechanism on which the whole business rests is simple: orders in the book are ranked by price and, within a single price level, by the time they were placed. An order that arrives late is filled later. This is why an HFT participant constantly re-quotes and cancels orders, and their order-to-trade ratio is orders of magnitude higher than a human's.

The exchange charges for this: its fee schedule takes excessive order book activity into account, and an overly "chatty" algorithm pays separately. This is the point where the regulatory infrastructure directly affects the economics of the strategy — for slow strategies such fees simply go unnoticed.

Where manipulation begins

Placing and cancelling orders quickly is legal in itself. What makes it illegal is intent: an order placed with no intention of being executed, in order to create the appearance of demand and move the price, is spoofing and layering, and both are classed as market manipulation under Federal Law No. 224-FZ (2010). The same law prohibits the use of non-public information, and the line between "reading public news quickly" and insider trading is drawn by the source of the data, not by the speed of the reaction. Public sources such as the news feed and the events calendar are not inside information, however fast you process them.

What a private trader can take from this

A retail participant has no direct access to HFT, but part of its engineering culture carries over to slower horizons: controlling execution costs, accounting for slippage, separating the signal from execution, and rejecting strategies whose profit per trade is smaller than the total fees. This is covered in a separate piece on algorithmic trading, which shows exactly where an algorithm loses money.

The second practical conclusion concerns leverage. A short horizon tempts traders to increase position size, and a forced liquidation does not ask how good your algorithm is — the rule is described in the article on margin trading. The third is that frequency is not an advantage in itself: without high-speed infrastructure, frequent trades mean only frequent costs, and this is precisely the mechanism that turns activity into overtrading.

To be candid: this article contains no quantitative estimates of HFT's share of turnover on the Russian market, the cost of co-location or typical response times — the platform does not have such data, and passing off other people's rounded figures as facts would be wrong. Definitions of the terms used above are collected in the glossary.