Coverage of machine learning applied to cryptocurrency markets tends to focus on outputs: predictions, signals, classifications. The preparation work that precedes those outputs receives almost no attention, which creates a distorted picture of what the technology involves and how reliable its results can be.
Where the actual time goes
A data engineer working on a cryptocurrency analysis pipeline will typically spend the majority of project time on data collection, cleaning, and alignment. Cryptocurrency exchanges record trades independently, and their timestamps frequently differ by several seconds. When a model is trained on data from multiple exchanges without correcting for this, the resulting patterns can be artefacts of the timestamp misalignment rather than genuine market behaviour. Identifying and correcting these issues requires domain knowledge that is not automated.
Label construction is another underappreciated step. A supervised learning model needs to know what it is trying to predict. In cryptocurrency analysis, defining the target variable precisely is harder than it appears. A one-hour forward return calculated from a mid-price behaves differently from one calculated from a transaction price, and the choice affects what the model learns.
What has changed in this area recently
Since 2023, several open datasets specifically designed for cryptocurrency machine learning research have been published with detailed provenance documentation. This means researchers can now compare results across studies more reliably than before, because the underlying data preparation choices are recorded rather than inferred. It is a quiet but substantive improvement in how the field operates. The outputs are still imperfect, but the conditions for honest evaluation have improved noticeably.