We spend most of our effort trying to prove ourselves wrong.
Any computer can find a strategy that would have won yesterday. Telling a real edge from a lucky one, before a cent is at stake, is the entire job. This is how.
Tens of thousands of candidates. Almost all discarded.
Anyone can run an optimisation. Finding the rare configurations that still work out-of-sample - on data the system was never shown - is the whole game, and ours re-select every week as the market moves.
We rank on the shape of the curve, not the size of the profit.
The grey candidate below made more money than the one we kept. We binned it anyway - it dug a hole nobody would sit through, went nowhere for months, then made everything in a single week. That is not a trend, it is a coin landing well once.
Rank on the money and you buy the grey one. Rank on the shape of the trend and you keep the blue one - the only one with any reason to do it again.
Before anything trades, we estimate the odds it is fooling us.
One good backtest can always be luck. So the whole selection is re-run across many train/test splits and reduced to a single number: the chance the record is an artefact of the search. Miss the bar, and it never trades.
The needle must settle left of the bar, deep in the green, before a strategy goes near real money.
A real edge keeps its rank on data it never saw; curve-fitting does not.
Measured against pure chance.
We pit the real procedure against tens of thousands of random pickers. Luck alone produces almost nothing; ours lands far out in the tail. That distance is the t-statistic - the number we lead with.
The real procedure lands 2.86 standard deviations out, where chance almost never reaches.
You see every test. Never the recipe.
The exact statistics we rank on, the way they are combined, and the machinery that runs it took years to build and stay private. What is public is the discipline and the evidence it produces - good weeks and bad - on the results page. Judge the output; the method stays in the box.