AI trading bots, honestly assessed

Two products, one word

A genuine machine-learning system derives its rules from data rather than having them written by a person. A marketing AI bot is a set of conditional rules — a moving average, an oscillator, a filter — with the term applied because it converts better.

The test that separates them is a question: describe the selection mechanism in a paragraph. A real system can be described; a label cannot. Anyone unwilling to explain how their software decides is relying on the word to do the work the explanation should.

Why machine learning struggles specifically here

These methods excel where signal is strong and data is abundant. Financial price series are the opposite: mostly noise, non-stationary, and available in far smaller quantities than the row count suggests, because adjacent observations are heavily correlated.

A model with thousands of parameters fitted to a few years of gold prices is the most efficient overfitting machine ever constructed. It will describe the training period beautifully and generalise poorly, and the better the in-sample result the more suspicious it should make you.

The version of adaptivity that does work

A narrower claim survives scrutiny: track which of a fixed set of strategies is currently working and weight toward it. That is measurement rather than learning — the rules are still written by a person — but it does respond to changing conditions, which is what most buyers actually want from the word.

Our engine does this as a tournament, scoring each strategy on the Wilson lower bound of its measured win rate rather than the raw figure. The lower bound is what stops it chasing whichever strategy got lucky most recently, and we call it a tournament rather than AI because that is what it is.