Can Machine Learning Beat a Random Walk in Bitcoin?
An interactive companion to my working paper. I benchmarked 32 forecasting models,
from ARIMA and ridge regression to XGBoost, LSTMs and Transformers, against the simplest possible forecast,
“tomorrow's return is zero”, on seven years of real daily BTC-USD data
(13 purged walk-forward folds, Sep 2019 to Sep 2026, about 2,560 out-of-sample forecasts per model).
The short answer: no. After correcting for testing 32 models, none is
significantly more accurate than the random walk at 1 or 7 days, and most are significantly worse.
Everything below is computed in your browser from the saved out-of-sample predictions.
Long/flat buys when the forecast is positive and holds cash otherwise. Long/short follows the sign.
Default cost: 0.10% fee + 0.05% slippage = 15 bps.
How the evaluation works
Target: forward log return, not price level. Predicting yesterday's price looks accurate on prices but carries no information.
No leakage: every feature is lagged one day; validation is expanding-window walk-forward with purging and embargo; only folds with 2+ years of training history are scored.
Benchmark: the zero-return forecast. An out-of-sample R² above 0 means a model beats it.
Significance: Diebold–Mariano test for each model, Holm-adjusted for 32 comparisons.
Trading: the forecast's sign becomes a daily position; costs are charged on every unit of position change.
Research and education only. Nothing here is investment advice.