Selected on the past. Judged on a future it could not see.
This page explains the testing procedure in enough detail that you can criticise it. That is the point - a method you cannot interrogate is a method you have to take on faith.
The edge is in
what we throw away.
A computer can find a setting that would have won yesterday. Knowing which one wins tomorrow - and binning the rest - is the whole job.
A good backtest proves nothing.
Give a computer enough dials and it will always find settings that would have won on the past. Run them back over that same past and the curve looks perfect.
Of course it does. It was built from it.
We never test on the past it learned from.
Settings are chosen on data up to a date, then judged on the week after - one the system had never seen. Then the window rolls forward. 192 times.
Anchored: the start never moves and the optimise window grows. Every test still lands on new ground.
You get the outputs.
Never the recipe.
How the winning settings are chosen took years to build, and it is the one thing we never show. Everything it produces - good weeks and bad - is public, on the results page.
What we are not hiding.
Judge it with the limits in plain sight. All of these are true today.