Why you Should Stop Predicting Prices if you want to Stand a Chance of Predicting Prices
Financial Data Almost Might have been Designed to Beat You
Published in
13 min readNov 16, 2021
The internet is full of forums, such as Towards Data Science here on Medium, that are full of eager articles explaining how to use today’s modern “AI” tools to predict stock prices. I find the articles they publish follow a common pattern:
- The author provides code snippets, almost exclusively in Python, that shows how to pull data from free public data sources, such as Yahoo! Finance or Quandl, using a simple API.
- They then include some commentary on how advanced and complex neural networks, such as LSTM networks, are “known to be accurate” when predicting time-series, and dive into their theory and structure.
- Next they demonstrate how to train these networks, using popular open source frameworks, and claim a high accuracy on the training set, such as 85% or more.
- Finally the system is run out of sample and draws a curve which kinda looks like a moving average over the prices of, say Tesla or IBM, but, truth be told, basically sucks as a forecast of the stock price.
- In conclusion, they state that the performance is “not so bad” and the system needs “more work.”
When these posts appear in my feed, here on Medium, I sometimes add a comment along the lines of:
Did you compare your…