Exploring Systematic Feature Discovery in Digital Asset Markets
Caroline Bishop Sep 11, 2025 05:23
Glassnode introduces a new framework for feature discovery in digital asset markets, utilizing machine learning to identify high-value indicator combinations.
Glassnode, a renowned blockchain data provider, has unveiled a novel approach for discovering features in digital asset markets through machine learning. This new 'bottom-up' feature discovery framework is designed to identify non-trivial and high-value indicator combinations, offering an alternative to the traditional 'top-down' feature engineering methods.
The Combinatorial Challenge
Modern financial markets, particularly the cryptocurrency ecosystem, generate vast amounts of data. Glassnode tracks hundreds of on-chain metrics across various assets and timeframes. The challenge lies in navigating the extensive feature space, as each metric can be transformed and combined in numerous ways, leading to an intractable search space. This complexity, known as the 'curse of dimensionality,' makes exhaustive exploration difficult.
Bottom-Up vs. Top-Down Feature Engineering
Traditional top-down feature engineering involves selecting metrics based on economic theory and market understanding. In contrast, the bottom-up approach samples the feature space without predefined preferences, potentially uncovering patterns that intuition might miss. This data-driven method allows the data itself to reveal unexpected combinations.
Systematic Discovery Methodology
Glassnode's methodology involves sampling a representative subset of the feature space while maintaining statistical rigor. The approach evaluates feature combinations using low-complexity machine learning models, focusing on discovering robust indicators likely to generalize beyond training data. The evaluation process includes testing across multiple time-based folds to ensure performance consistency.
Case Study: Bitcoin Trend Detection
To demonstrate the methodology, Glassnode applied it to Bitcoin uptrend detection, using a three-phase sampling strategy. The objective was to identify optimal periods for Bitcoin long exposure during uptrends. The study employed hierarchical trend segmentation to label market phases, capturing 'mini bull runs' within larger cycles.
Exploration Phases
The exploration process consisted of three phases:
Phase 1: Single-Feature Screening
This phase evaluated 153,600 single-feature combinations to identify metrics with potential. Top metrics spanned valuation ratios, holder behavior, and profit/loss distributions.
Phase 2: Metric Pair Discovery
Glassnode sampled 100,000 evaluations from possible combinations to identify synergistic pairs, revealing that realized cap and activity retention metrics were notably effective.
Phase 3: Parameter Optimization
This phase focused on optimizing historical context windows for the most promising metric pairs. Surprisingly, optimal windows ranged from 800-1,200 days, longer than conventional analysis periods.
Out-of-Sample Results
The 2024-2025 validation period provided insight into the real-world effectiveness of the findings. While some combinations maintained consistent results, others showed reduced effectiveness, highlighting the need for ongoing adaptation in evolving markets.
Practical Implications and Limitations
Glassnode's framework reveals non-obvious relationships and demonstrates the potential of structured exploration in digital asset markets. However, the analysis does not represent a complete trading strategy and results are specific to the chosen metrics and time period.
Conclusion
The study highlights the importance of computational discovery in financial markets, offering a scalable methodology for practitioners. As cryptocurrency markets evolve, so must analytical approaches, with computational exploration augmenting traditional expertise.
For more detailed insights, visit the original article on Glassnode.
Image source: Shutterstock.jpg)