Blog
Arguments about evidence.
The handbook answers questions people ask. This is for the ones they do not — about how strategies get found, what the industry has agreed not to measure, and where our own approach is weakest. Opinions, dated, and allowed to age.
- Your backtest didn’t lie to you. Your search did.Everyone in this industry loves to say backtests lie. They mostly don’t. The arithmetic is fine, the data is usually fine, and the chart is an accurate description of what that rule did on that history. The dishonest part happens earlier, and it leaves no trace in the output.
- A model should not grade its own workThere is a rule in experimental science that predates machine learning by a century: the person with a stake in the answer does not get to run the test. The current generation of AI trading tools has quietly abandoned it, and mostly nobody has noticed, because the products describe it as a feature.
- We built a tool that makes overfitting easier. That’s why we count.This is the most awkward fact about what we are building, and the case for it at the same time. Lagias lets you go from a sentence to a tested strategy in minutes. Doing that quickly is the single most effective way to overfit, and we would rather explain how we handle it than hope nobody does the arithmetic.