Testing Never Stops

Most traders test once to decide whether to trade. I keep testing the live book every week.

Most traders test a strategy once, to decide whether to trade it. I test mine every week, including the ones that have been live for months and are making money.

Why test something that's already working?

Because a live strategy isn't a finished object. It's still just a hypothesis. The regime it was built in drifts. The behaviour that gave it an edge gets crowded, usually by people running something close enough that the same moves become more expensive to capture. A strategy rarely breaks in a single day; it decays across a period of time. By the time the curve rolls over, the money is already gone.

That's the trap. The P&L is a lagging indicator of your own edge. It feels like the truest number on the screen because it's the one with your money attached, and it's the one that tells you last.

So what do I actually watch?

Not the equity curve. I watch the signal-health metrics that move earlier: live-versus-backtest variance, win-rate drift, the average trade in R turning down while the win rate holds, fill quality measured against what I modelled.

A win rate that holds while the average winner shrinks is a strategy with its expected value being repriced. It looks fine right up until the tail trades that paid for everything stop turning up. Catching that on the average-R line in week six is cheap. Catching it on the drawdown in month five is not.

Where do the candidate strategies come in?

In parallel I run the candidate book: strategies that aren't live yet. Each one goes through the same gauntlet before it earns a slot. I optimise on a small in-sample window, then run walk-forward on out-of-sample data the optimisation never saw. Then I run it again on an older out-of-sample set from before the in-sample period, because a strategy that only survives on recent history hasn't really survived anything.

I keep the optimisation deliberately light. A parameter sitting at exactly its best value is curve-fit by construction, so I tweak the settings around that value to confirm the edge is stable and not balanced on a single number. Monte Carlo on the trade sequence tells me how much of the result was the strategy and how much was the order the trades happened to arrive in.

Most candidates never make it. The majority of what I build dies in testing, and that is the system working rather than failing.

Same checks, different question. For the live book the question is whether this is still the strategy I shipped. For the candidate book it's whether it has earned a live slot at all.

What makes killing a strategy cheap?

None of this works without somewhere soft to land when a check fails. Every sub-strategy is sized so its worst historical drawdown sits around one percent, and the portfolio runs its own risk overlays on top of the individual controls. So when a daily check flags a strategy drifting from the thing I shipped, retiring it costs me almost nothing.

Conviction is worth a lot when a position is drawing down and still correct. It's worth nothing when the strategy underneath it has stopped being the one you tested. Telling those two apart, every day, is most of the job.

This is the unglamorous core of running systematic money. XAQP runs twenty to thirty strategies at any given time across FX, indices and metals, breakouts sitting next to trend-following sitting next to mean reversion. There's no clever signal at the bottom of it that nobody else has. What holds it together is the art of killing the strategies that are dying before they take a real bite, and only ever giving live capital to the candidates that earned it.