Refine a strategy: variations, locks and the held-out check

What it is for, and what it is not

It is not a search for strategies. Every run starts from one of yours and every variation is a version of it; the direction never changes, and nothing in the result is replaced unless you choose it. Use it when you have an idea that tests reasonably and want to know whether a nearby version of it — different numbers, a swapped condition, a different timeframe — would have done better on the same history. If you only want to move the numbers, a parameter search measures every cell rather than a sample and is the cheaper tool.

The setup, choice by choice

1. The strategy and one market. The search runs on one instrument's history, as a backtest does. The form says how many months it can actually cover there.
2. What the search may change — per condition. Keep leaves a condition exactly as written. Retune keeps its indicator, timeframe and comparison and lets only its numbers move; this is the default. Change makes it fully disposable: the search can retune it, move its timeframe, swap the comparison, replace the indicator or drop it. Keep the conditions that are the idea; open the ones you are less sure of.
3. What "better" means. A goal. Balanced is the default mix; Steadier results prefers variations whose results hold up across the whole window rather than in one good spell; Profit first prefers more per trade and barely counts how evenly it came. The chosen goal shows exactly how it weights the score, so the choice is made on the numbers.
4. How often it has to trade. A floor in trades a year. A variation that trades twice a year can score well for the wrong reason — almost no drawdown — so anything under the floor is not ranked.
5. How trades are closed. One fixed bracket for every variation: a stop and a target the same number of ATRs from the entry, two, three or four. The same yardstick for your strategy and every variation, so they compare like for like, and no dial for manufacturing a flattering exit.
6. How hard to look. Quick, Standard or Deep: how many variations per round and how many rounds. The form prices each in testing credits before you start.

On the web app there is a fourth goal, Entries where you aim, which scores variations on where their matches land inside areas you paint on a chart instead of on simulated trades. It is covered on the feature page; the phone can run it from a marking saved on the web.

How a round works

Each round builds a population of variations, measures them, keeps the ones that scored highest and makes the next population from those — a few hundred variations chosen well rather than a few million sampled evenly. Cheaper screening passes eliminate most variations before the full-depth measurement, and your own strategy is always measured at full depth so the comparison against it is exact: same window, same costs, same scorer.

Reading the result

The result leads with the variations that scored highest, your own strategy beside them, and how many of the variations measured scored above it. Every row carries its score, how often it traded, and its conditions written out in full, so you can see what changed rather than only that something did. The score is explained in the engine's own words on the page, term by term.

Then read the held-out check. After the ranking settles, the top variations and your original are measured once more on the stretch of history immediately before the one the search used — history the search never saw — with the same costs, bracket and scorer. A variation that holds up there has survived one more stretch of the market than it was chosen on. One that collapses there is telling you the ranking was fitted to the recent stretch. The held-out figure is shown beside the search score with equal weight and is never used to pick, because that would just make it part of the search.

Trades a year sits beside every score for a reason: a variation that outscores yours on a handful of trades has not shown you much.

What to do with a variation you like

1. Save it. A variation is copied into a strategy of your own, which you can then apply, alert on and test like any other. Your original is untouched.
2. Try to break it first. Check these settings opens a parameter search centred on the variation's numbers. If it only works at exactly those numbers, the search found a coincidence in this history.