First Test, Then Train (Yes, Really)
Everyone tests their model on the most recent past. I tried the opposite, and it worked better.
In my career, I’ve seen ML models used across different applications and industries, but they all had one thing in common: the way they used time.
It always goes like this: train on older data, test on the most recent data, and finally retrain on the latest data before deploying.
And for a long time, I never questioned it. Until I noticed something that should have been obvious all along: the model I was testing was not the same model I was deploying.
So I thought: there has to be a better way. It turns out there was.
