Algorithmic trading

The label covers a wide range of practice held together by one property: the logic is written down precisely enough for a machine to follow it. At one end sits an execution algorithm whose only job is to work a large order into a market over time without moving the price against itself. At the other sits a strategy that decides what to trade, in which direction and in what size. A market maker quoting continuously on both sides sits somewhere between the two.

Building one follows a conventional sequence. A rule is specified, run over historical data, tested again on a period the specification never saw, and then run live with monitoring. The step that decides whether any of it means anything is the cost assumption: a rule tested without spread, commission and slippage will report results a market cannot produce, and the smaller the average gain per trade, the more completely costs dominate the outcome. A result fitted to the sample it was fitted on is not evidence, which is the oldest known failure in the field.

Two things are worth separating. Practitioners disagree about whether algorithmic and automated trading name the same thing: one usage reserves algorithmic for the decision logic and automated for the hands off execution of it, another treats them as synonyms. Regulators, meanwhile, treat algorithmic order flow as a supervised activity in several jurisdictions, requiring firms to test their algorithms, control the rate at which they send orders and hold the ability to stop them.