Can machine learning predict bitcoin’s daily direction?
Not in our replication. A calibrated XGBoost classifier built to the recipe of a 2026 paper called bitcoin’s next-day direction correctly on 49.8% of days from 2022 to 2026, and its probability forecasts scored worse than a simple base rate in every year. The paper’s surviving configurations all decide at 21:30 UTC, where, in the authors’ public code, some inputs already include part of the day being predicted.
The claim
Jung trains an XGBoost classifier only on days with large moves, so the model does not spend its capacity on near-zero days, calibrates its probabilities, and retrains twice a month in a walk-forward loop on 99 features: technical indicators on three timeframes, funding, basis, macro series, sentiment and calendar. After strict corrections for testing many configurations, only a handful survive; the best returns +255% over two years on 14 trades, 93% of them winners.
What we tested
We rebuilt the model from the paper and the author’s open code and ran it on our own data, with every input stamped at the moment it was actually available. Decisions were made at midnight UTC and at 17:00 New York time, close to the paper’s 21:30 UTC, with and without the paper’s large-move filters, on two feature sets: 94 of the paper’s 99 features, and an extended set that adds options, open-interest, on-chain and fund-flow data. Forecasts are out of sample from March 2022 and are scored against a trailing base rate of up days.
Result
The target configuration was right on 49.8% of days and scored worse than the base rate in every year from 2022 to 2026; trading its signal gave a Sharpe ratio of −0.95. Across all twelve configurations, Sharpe ratios ranged from −0.95 to +0.09. The richer feature set did not help.
The timing issue: in the public code, daily ETH prices and futures basis are stamped at midnight but close at 24:00 UTC, so at a 21:30 decision they already contain 2.5 hours of the 24-hour window being forecast. In the paper itself, no configuration deciding at midnight — where the inputs line up — survived.
Published best configuration versus our replication
| Paper (best configuration) | Our replication, 2022–2026 | |
|---|---|---|
| Decision time | 21:30 UTC | 00:00 UTC and 17:00 New York |
| Hit rate | 93% of 14 trades | 49.8% of days |
| Performance | +255% over two years | Sharpe −0.95 (range across 12 configurations: −0.95 to +0.09) |
| Probabilities vs a base rate | — | Worse in every year |
What it means
Bitcoin’s daily direction is close to a coin flip, and a flexible model with many features will find patterns in any finite sample; training only on big-move days does not change that. When a published result is concentrated at one decision time, it is worth checking what the inputs could see at that moment.
What would change the verdict
A model whose forecasts are logged before the outcomes and beat the base rate after costs would change our view. So would a version of the paper’s result that holds at a decision time where every input has closed.
This note re-tests a claim made by the cited authors on our own crypto market data; the summary of their method is ours. Tested · published · last reviewed . Research, not investment advice.