Marcos López de Prado's book protects you from one thing
By Robert Yenokyan,
Marcos López de Prado's book protects you from one thing.
Yourself, with the data you already have.
In week one of the course I teach, I say out loud what that leaves out.
His book is the standard reference for using machine learning on financial data and it is on my reading list. The machinery in it is careful and worth the time. Seal off the stretch of history you test on, so nothing in it can reach the part you learned from. Count the attempts behind a result and mark the result down for them.
Both of those guard the same door. The data is in your hands and you are the one who can fool yourself with it.
Neither can tell you that the thing you measured has stopped existing.
A market is not a fixed object. The people in it change, the rules change, the regulation changes. What you learned from crypto in 2019 is not a description of crypto in 2026, and no amount of careful testing on the 2019 data will say so. That problem is open. No toolkit I teach has an answer to it.
So I say it in week one, instead of letting somebody walk into it halfway through a thesis.
Know which problem each of your tools was built for. A test that protects you from your own data cannot protect you from a market that changed underneath it. Both failures are real. Only one of them has a toolkit.
I teach that course at Yerevan State University and spend the rest of the week building trading systems, where the same open problem turns up with no reading list attached.
If you are picking a research question this year and you want an outside read on which part of it nobody has answered yet, write to robert_yenokyan@bonton.ai.
The two safeguards have names. Sealing the test period off so nothing leaks into the training period is purging, with an embargo being the gap you leave after each test window. The second is a family of measures that mark a track record down for the number of attempts behind it. Both are in Advances in Financial Machine Learning and both are worth the afternoon it takes to learn them. The open part is not the kind of problem you fix by processing the data differently. The rules themselves change, and nothing you do to the 2019 numbers turns them into a description of 2026.
More writing
About
Robert Yenokyan is a quantitative researcher and entrepreneur in Yerevan, Armenia. Founder and CEO of Bonton AI. Financial Engineering lecturer at Yerevan State University.