Quantitative trading

An approach in which the decision itself is produced by a rule set, and the rule set is arrived at by measurement: a hypothesis about market behaviour is stated, expressed as something computable, tested against history, and either kept with its parameters fixed or discarded. What distinguishes it is not the presence of a computer but the source of the decision. Algorithmic trading describes how an order is worked into the market, which is an execution question, and the two are constantly conflated. A quantitative rule set can be traded entirely by hand, and a purely discretionary view can be executed by an algorithm.

The families are few and long established: trend and momentum rules, mean reversion rules, relative value between related instruments, carry, and market making. What differs between them is far less than what they share, which is the assessment method. A rule is judged on a distribution of outcomes rather than on a result, so the figures reported are expectancy per trade, the dispersion around it, the depth and the length of the worst drawdown, and a risk adjusted measure such as the Sharpe ratio. Dealing costs sit inside the test rather than beside it, because the spread, the commission and the slippage on every signal scale with how often the rule trades, and a rule can be positive before costs and negative after them.

The central difficulty is statistical rather than technical. Searching one history for the best performing variant of a rule finds the variant that best fits that history, and the more variants tried, the more of the result describes the sample rather than the market. This is a property of the search, not a lapse of discipline, and it is why the conventional guards exist: a period held back and never looked at, testing that walks forward through time rather than fitting the whole span at once, and a penalty applied for the number of variations tried. Each reduces the problem and none removes it. Practitioners disagree openly about how much evidence outside the fitting sample is enough, and about why published rules decay after publication, since a rule arbitraged away and a rule that was never there look identical from the outside.