Assumptions!
The key to understanding, and modeling anything is to understand assumptions.
Last weekend I was at a magic show. (First time since I was a kid.) Watching the magician perform, I realized that most of stage magic revolves around manipulating, and taking advantage of the audience's assumptions. (Of course combined with amazing dexterity, showmanship, stage presence, etc.)
- Sawing a woman in half: You assume her body is the length of the box, the feet sticking out are really hers, etc. (Spoiler, she's curled her legs up and the feet are fake)
- Something disappears behind a curtain: You assume you're seeing the real stage and curtain (Spoiler: Often a large, angled mirror)
- Magician escapes some scary water tank or deadly scenario? (Spoiler: the handcuffs are fake, there are mirrors, etc.)
I've probably gotten some of that wrong, and missed a ton - no disrespect to the magic profession intended. But, you get the point. Play with people's assumptions and you can "do magic".
The same holds true for machine learning and algorithmic models. Doesn't matter if you're modeling high-frequency returns, or the probability of the next corner kick. There are assumptions that you are blind to, And even assumptions about the assumptions that you are missing. The result is that the world doesn't fit your model, and you're shocked.
Every statistics 101 student learns that standard linear regression assumes the residuals are "IID Normal". Most forget what that means, or how to even verify it. And, that's an easy one. Once we get to more complicated models, systems, and plumbing there are endless hidden assumptions that most people overlook.
The key is to ruthlessly identify all the assumptions in the your model, the math, the data, the software, everywhere. Then, assume you missed a bunch and try to account for that unknown unknowns in your actions. You'll still get some of it wrong, but it will be less wrong.
Trying to be "less wrong" is the entire game. There is nothing else.
#QuantLife #ModelMagic
@hamptonism
@QuantStacker
@tradingview
@0xfoobar
@k33research