Trading strategy optimization

Do not optimise for profit

This is the single most consequential choice and almost everyone gets it wrong. Ranking candidates by total profit or by profit factor will crown a configuration that made its money in one exceptional trade, or that took a drawdown you could never have sat through, or that produced eight trades in total.

A better objective combines several things at once: how large the average trade was relative to its variability, how many trades support the result, and how much drawdown was suffered getting there. That mixture asks a different question — not "how much did it make" but "how sure are we this is not noise, and what did it cost to find out".

A specific refinement worth knowing: when measuring variability, divide by the deviation of the losses alone rather than of all trades. Dividing by the deviation of everything penalises large winning trades, which means the search actively rates down the exact behaviour that makes trend strategies profitable.

Use a coarse grid

It is tempting to test every value in fine increments. Resist it. A fine grid finds sharper peaks, and sharper peaks are more likely to be artefacts of that particular sample than real behaviour.

A coarse grid tests fewer combinations, which reduces the number of chances the search gets to fit noise, and the peaks it finds are wide enough to be plausible. If a strategy needs a period of exactly 21 and fails at 20 and 22, it does not have an edge — it has a coincidence.

Judge on data you did not fit on

Rank candidates using the fitting window, then take only the top handful forward and score them on a window they never touched. Carrying every candidate through would turn the second window into just another thing to overfit against.

Only the out-of-sample score is allowed to decide anything. The in-sample number's only use is deciding which candidates are worth the cost of judging.

Check the neighbourhood before adopting

When a winner emerges, score the parameter values one step away on each axis. If the winner is profitable and its neighbours are too, you have a plateau — the surface is smooth and the peak sits on raised ground. If the winner is profitable and its neighbours lose, you have a spike.

A spike is not an edge. It is the one parameter value whose entries happened to align with this sequence of prices, and the next sequence will put it somewhere else. We treat that as a veto, and the reason is a concrete one: a configuration that passed every other gate — out-of-sample profit factor 1.474 — ran at 0.518 live. Its neighbourhood was never profitable.

Move one parameter at a time. Moving two together tests a configuration the search never considered and mostly measures how far you went.

How often to re-optimise

Less often than instinct suggests. Re-tuning after every drawdown means chasing whatever worked most recently, which is a way of guaranteeing you are always fitted to the period that just ended.

If you do re-optimise automatically, require the new configuration to beat the running one by a margin rather than by a hair, and require it on enough trades to mean something. A system that switches settings on a rounding difference will churn constantly and never accumulate the sample it needs to know whether anything works.

It is entirely reasonable to conclude that a strategy should not be re-optimised at all — that its parameters are what they are and the market decides the rest. Standing still is a valid outcome and is what stops a system chasing noise.

Common questions