Field Note 03 / AI × Markets

Why I Stopped Trying to Make AI Predict Bitcoin

It sounds almost absurd in 2026. We have AI models that write code, analyse documents, and recognise patterns across enormous datasets. Surely the obvious next step is to give them BTC history and ask: up or down?

I tried going down that road with my own trading research. The deeper I went into the literature, the less obvious the idea became.

Detection is not prediction

Massimo La Morgia, Alessandro Mei, Francesco Sassi, and Julinda Stefa built a machine-learning model to detect cryptocurrency pump-and-dump events in real time. Their model identifies a pump within 25 seconds of its start, with an F1 score of 94.5%, improving on prior work that took roughly 30 minutes to detect a pump at an F1 score of 62.7%.

That is a strong detection result. But the model answers one question: is a pump happening right now? It does not answer the question a trader actually cares about: after detection, will the move continue far enough to stay profitable once spread, slippage, fees, latency, and the exit are all accounted for?

The distinction sounds small until money is attached to it. Recognising an event already in progress is not the same as forecasting its tradable continuation. This is the same distinction Klines, our own event-detection system, is built around: it detects, it does not predict.

Source: La Morgia, Mei, Sassi & Stefa, “The Doge of Wall Street: Analysis and Detection of Pump and Dump Cryptocurrency Manipulations” →, ACM Transactions on Internet Technology, 2023.

The backtest problem

Then there is the second problem: backtests.

Berend Gort, Xiao-Yang Liu, and their co-authors wrote a paper on deep reinforcement learning for cryptocurrency trading around one uncomfortable issue: impressive historical returns may be false positives created by overfitting. Their answer was not simply to add more AI. They framed overfitting as a hypothesis test, estimated its probability, and rejected agents that looked good for the wrong reasons.

This does not prove machine learning cannot produce a useful trading model. It establishes a harder and more important point: a strong backtest is not enough to show that it has.

Source: Gort, Liu, Gao, Chen & Wang, “Deep Reinforcement Learning for Cryptocurrency Trading: Practical Approach to Address Backtest Overfitting” →, arXiv, 2022.

What the research changed

I stopped asking machine learning to manufacture a confidence score from RSI, MACD, volume, and dozens of transformations of the same price-and-volume history. More inputs and a more complicated model did not resolve the central problem: the result still had to survive unseen data, realistic execution, and changing market regimes.

The research did not become a Bitcoin predictor. That was the result.

The machine remained useful, just in a narrower and more defensible role: describing what is happening now, testing whether an assumption survives history, finding cases I might otherwise miss, and forcing a strategy to earn its claim out of sample. That is also the discipline behind every note in Searching for Edge: literature first, replication on real data, gross, net of costs, verdict.

A less exciting claim, and a better one

Maybe the best role for AI in markets is not “tell me where Bitcoin will be tomorrow.”

Maybe it is: help me describe what is happening now, test whether my assumptions survive history, and stop me from confusing a beautiful backtest with an edge.