Selecting distant sources improves cryptocurrency forecast skill
Beyond Wasserstein Proximity: Finding Better Transfer Sources
Machine Learning
Summary
Choosing data sources that are not the most similar can actually help make better predictions about cryptocurrency price changes. The authors show that using more varied data sources, even if they seem less closely related based on standard measures, can improve forecasting models like neural networks. They tested this idea using different cryptocurrencies over five days and found that picking the farthest data sources improved prediction accuracy. Their approach also helps identify which sources to use before full training.
What this means in practice
- •For financial modelers: Improve cryptocurrency volatility forecasts by selecting more varied training data sources beyond just the closest in distribution.
- •For machine learning engineers: Enhance transfer learning models by using a screening method that favors distant sources before fine-tuning, improving prediction skill across datasets.
Tested on one dataset.
Authors
Gordei Verbii, Georgii Killinkarov, Vladimir Yafarov
Abstract
Source selection often favors similar data, yet marginal proximity need not identify the most useful pretraining source. We motivate a conditional alternative through domain-adaptation theory and a regression example in which broader source variation improves estimation despite increasing Wasserstein distance. We examine five-day cryptocurrency volatility using ten Binance spot assets, six forecaster families, three seeds, and shared temporal cutoffs. Selecting the three farthest rather than nearest Gaussian-Wasserstein sources improves mean persistence-relative skill by 12.52 percentage points for MLP and 20.90 for Transformer, with positive post-hoc paired date-block intervals. Forward pilot selection improves mean skill over random transfer in all six families. Compact recurrent models achieve the lowest absolute mean errors, separating selection utility from forecasting quality. Retrospective five-candidate screening shows additional gains from farthest-first geometry before pilot ranking for MLP and Transformer. We distinguish across-target associations from within-target decisions and evaluate target heterogeneity, training seeds, and temporal-block sensitivity. Findings are exploratory for a fixed historical panel; the screening analysis measures no runtime saving.