Systematic trading

An approach in which every decision is specified before it is faced. The rules state what conditions open a position, what size it takes, and what closes it, in terms precise enough that two people applying them to the same data reach the same answer. Whether the rules are executed by hand or by code is an implementation detail: automation is a way of running a system, not what makes one. The contrast is with discretionary trading, where a judgement is made in the moment.

Building one is a research process with a well documented failure mode. A hypothesis is specified, tested against historical data, and then validated on data the design never saw, commonly through a walk forward procedure in which parameters are fitted on one window and tested on the next. Costs are the part that decides most outcomes, because a rule producing many positions is a rule paying many spreads. The failure mode is fitting the rules to the sample that suggested them: the more variants are searched, the more certain it becomes that the best one is describing that sample's noise. Backtesting is where that happens, and it is why out of sample validation exists at all.

Two claims made for the approach do not survive contact with it. Systematic does not mean judgement free: the judgement has moved to design time, into the choice of market, window, parameters and cost assumptions, where it is documented rather than absent. And systematic does not mean unattended, since connectivity, data errors and a rule meeting conditions its author never imagined all require supervision. How well tested results transfer to live conditions is genuinely disputed, and documented degradation after publication is one of the reasons the dispute persists.