Automated analysis of cryptocurrency prices has existed in some form since around 2016, but the tools involved have changed considerably. Earlier systems mostly applied rules that human analysts wrote by hand. Current systems learn those rules from data, which sounds more impressive than it often turns out to be in practice.
The gap between capability and perception
A machine learning model trained on Bitcoin price data from 2018 to 2022 will find patterns that existed during that period. Whether those patterns persist into 2024 or 2025 is a separate question entirely, and one the model itself cannot answer. This is called distribution shift, and it is the central unsolved problem in applying these tools to financial data of any kind.
What has changed recently is not that this problem has been solved. Rather, researchers have become more systematic about measuring how quickly a model stops being useful after training ends. Some newer studies report that models trained on cryptocurrency data begin losing predictive coherence within six to eight weeks of their training cutoff. That figure is not a guarantee, but it gives a concrete sense of the maintenance burden involved.
What this means for someone trying to understand the field
If you are reading about machine learning and cryptocurrency because you want to understand what the technology actually does, the most useful frame is this: these are statistical tools that find historical correlations. They require constant retraining, careful validation, and honest accounting of where they fail. The behind-the-scenes work is far less dramatic than the promotional language surrounding it.