Case Study
Cryptocurrency Price Prediction & Analysis
Crypto forecasting study on real Coinbase/Kraken OHLCV. No model beats a persistence baseline on next-day returns; the durable signal is regime clustering (DOGE silhouette 0.625). 23 tests.
Executive Summary
A study comparing LSTM, ARIMA, and regularized-regression forecasts of next-day crypto returns on real Coinbase and Kraken OHLCV, plus K-Means regime clustering and a Tableau dashboard. Rebuilt from scratch with 23 automated tests and a return-based evaluation protocol.
Problem & Constraints
Crypto markets are volatile and non-stationary. The core evaluation trap is autocorrelation: predicting price levels yields deceptively high R² because tomorrow's price sits close to today's. An earlier version reported R²≈0.98 on price levels, an artifact of that autocorrelation, not predictive skill. The rebuild evaluates next-day returns against a persistence (random-walk) baseline.
Honest finding
On real Coinbase/Kraken OHLCV, none of the LSTM, ARIMA, or regularized-regression models beat a persistence baseline on next-day returns. The durable, reproducible signal isn't forecasting but regime clustering: K-Means separates DOGE's behavior at a silhouette score of 0.625.
Methodology
- Ingested real OHLCV history from the Coinbase and Kraken APIs
- Benchmarked LSTM, ARIMA, and regularized regression forecasts of next-day returns against a persistence baseline
- Applied K-Means clustering to uncover asset regime patterns (DOGE silhouette 0.625)
- Deployed an interactive Tableau dashboard for model comparison
- Locked behavior with 23 automated tests
Results & Metrics
| Component | Result |
|---|---|
| Next-day return forecast | No model beats persistence baseline |
| Prior R²≈0.98 | Autocorrelation artifact of price-level prediction (corrected) |
| Clustering | K-Means, DOGE silhouette 0.625 |
| Data | Real Coinbase + Kraken OHLCV |
| Tests | 23 |
| Dashboard | Tableau live |
Tech Stack
Python, Scikit-Learn, TensorFlow, Statsmodels, Tableau, REST APIs
Future Work
Ensemble stacking, on-chain feature integration, real-time prediction API.