Your market model
fit to the past
?fit to the next ten years
nonereal reason it should work
The Butter Index
Explain the stock market with butter.
A quant once found that butter made in Bangladesh explained 75 percent of the S&P 500. Adding American cheese took it to 95 percent. Adding sheep took it to 99. Build your own.
Thirteen years of a stock market. Your job: find numbers that move the same way.
Why a perfect fit is a warning
Search enough numbers and some will move like the market by accident. Each one you add bends the line closer to the past. A fit of 99 percent on old data is not proof. It is what searching looks like.
The test that counts is data the model never saw. Nonsense fails it every time. So does a real-sounding model built the same way.
This happened
David Leinweber ran this as a joke on purpose, with real United Nations data. Butter production in Bangladesh explained 75 percent of the yearly S&P 500. US cheese production took it to 95 percent. The sheep populations of Bangladesh and the United States took it to 99 percent. He called the paper "Stupid Data Miner Tricks".
A computer that can test a million ideas an hour can find a butter index before lunch. It will not tell you that is what it found.
Before you trust a model, ask: how did it do on years it never saw?
Sources
- Leinweber, "Stupid Data Miner Tricks: Overfitting the S&P 500", The Journal of Investing 16(1), 15–22 (2007). Butter 75 percent, plus cheese 95 percent, plus sheep 99 percent.
- The market and the nonsense numbers on this page are random, made fresh in your browser. The labels are jokes. The search tries 1,000 numbers and keeps whichever fits the first thirteen years best.