AI to predict up or down: what I learnedBitcoin / TetherUSBINANCE:BTCUSDThectorlduI trained several artificial intelligence models (decision trees, neural networks and an already-trained time-series model) to predict whether price would go up or down. What happened - On the data they learned from, the results were spectacular. - On new data they had not seen, the edge disappeared: they got the direction right about as often as a coin toss, and after costs they lost a little. Why it happens: a model with a lot of capacity memorises the details of the past, noise included. It is like a student who memorises the answers to one exam: great on that exam, bad on the next one. What did work Predicting how much price will move (volatility), not where. Busy periods tend to stay busy and quiet ones tend to stay quiet. The indicator on this chart shows that volatility. What I take away - Always test models on data they did not use to learn. - AI is more useful for sizing a position according to how busy the market is than for guessing direction. Educational content. Not investment advice.