Feature Importance
Published Last updated
Key Takeaway
Feature importance quantifies which technical indicators and market variables most influence cryptocurrency trading model predictions, revealing which price drivers matter most.
Learn These First
What Is Feature Importance?
Feature importance quantifies which technical indicators and market variables most influence cryptocurrency trading model predictions, revealing which price drivers matter most.
How Feature Importance Works
Frequently Asked Questions
How do I use feature importance to improve my cryptocurrency trading model?
Identify top 5-10 most important features; focus monitoring and analysis on these. Remove low-importance features that add noise reducing generalization. Compare importance across market regimes revealing regime-specific patterns. If importance changes dramatically, market conditions have shifted requiring retraining. Validate importance alignment with trading intuition—if important features seem economically implausible, investigate possible overfitting. High-importance features typically include momentum (RSI), trend (moving averages), volatility (ATR), and volume confirming domain knowledge.
Why might two different machine learning models show different feature importance for the same cryptocurrency data?
Different algorithms measure importance differently. Tree-based importance counts feature splitting frequency; neural networks use gradient magnitudes; linear models use coefficient size. These measure distinct aspects. Correlated features show different importance in different models—in one model RSI might dominate, in another volume. Different training periods produce different importance—features valuable in bull markets may be unimportant in bear markets. Disagreement between methods or models often reveals insights: if tree and permutation importance disagree, feature interactions likely matter. Comparing importances across models reveals robust versus fragile features.
If a technical indicator shows low feature importance in my cryptocurrency model, should I remove it?
Low importance suggests limited individual predictive value but requires caution before removal. Sometimes seemingly unimportant features provide robustness against regime changes or support rare but important market conditions. Test removal empirically: retrain without the feature comparing validation performance. If validation improves or remains unchanged, the feature adds noise and should be removed. If validation degrades, the feature contributes despite low individual importance. Always validate removal decisions empirically rather than relying on importance scores alone. Simpler models often generalize better.
Common Misconceptions About Feature Importance
High feature importance means that indicator directly causes price movement and is a reliable signal.
Feature importance shows model reliance, not causality. A correlated but non-causal indicator shows importance because the model learned the correlation. Stock sentiment might show high importance predicting price without causing movement—causality runs the opposite direction. Feature importance indicates which features the model uses; domain expertise determines whether relationships are causal or spurious. Avoid assuming importance proves trading signal value without additional validation.
If my Bitcoin model shows consistent feature importance across many backtests, that importance is definitely stable and reliable.
Consistent backtest importance doesn't guarantee live consistency. Backtests assume unchanging relationships; markets evolve. Features important in 2022-2023 data may be unimportant in 2024 as market regimes shift. Different cryptocurrencies show different feature importance despite similar model architecture. Walk-forward validation revealing importance changes across time periods indicates instability. Only features maintaining importance across multiple independent market periods deserve strong confidence. Continuous monitoring detects importance shifts requiring retraining.
A model is definitely overfitting if feature importance doesn't match my intuition about cryptocurrency markets.
Misaligned importance might indicate overfitting, or the model might be discovering relationships traders haven't recognized. Markets are complex; models sometimes find non-obvious patterns. However, importance completely contradicting domain knowledge warrants investigation—examine whether the model is relying on spurious correlations. Cross-validate using different methods (permutation importance vs tree importance) revealing consistency. Balance skepticism with openness to genuine discoveries. Validate economically implausible importance through additional testing.