Can a GARCH model predict bitcoin’s direction 85% of the time?
No. Rebuilt from a 2023 paper on bitcoin perpetual futures, the EGARCH-based trading model called the next bar’s direction correctly 50.1–52.2% of the time from 2020 to October 2026, at horizons from 15 minutes to a day — no better than the sign of the recent average return — and after costs it trailed buy-and-hold. We could reproduce an 85% figure only by scoring the model against the window it was fitted on.
The claim
Malik models intraday returns of bitcoin perpetual futures with machine-learning and econometric models and finds that an EGARCH model best captures volatility and a constant ARMA(0,0) model the mean. A trading algorithm projects the futures price forward with these models and goes long or short depending on the slope between the current and the expected price. The abstract reports back-test returns of 1,500% to 8,000% and the direction predicted correctly 85% of the time by the best model.
What we tested
The full text is not openly available, so we rebuilt the method from the abstract: EGARCH volatility with a constant mean, refitted daily on a rolling window, on bitcoin perpetual futures from January 2020 to October 2026, with bars of 15 minutes, 1 hour, 4 hours and 1 day and the forecast horizon equal to one bar. We compared it with the simplest benchmark the maths points to — the sign of the plain average return over the same window — and traded both with fees and funding.
Result
With a constant mean, the model’s direction call is the sign of a volatility-weighted average of recent returns — a trend rule in disguise. It was right 50.3% of the time on 15-minute bars, 50.1–50.2% on hourly, 50.7–51.0% on 4-hour and 51.4–52.2% on daily bars, never reliably better than the plain-average benchmark. Traded with fees and funding, the best versions earned Sharpe ratios of 0.41 (hourly) and 0.16 (15-minute) against 0.62–0.67 for buy-and-hold. Scoring the sign of each window’s fitted mean against the returns inside that same window gives 78–92% agreement, 85% on average.
Share of bars where the predicted direction was right, Jan 2020 – Oct 2026
| Bar and horizon | EGARCH model (two variants) | Plain-average benchmark |
|---|---|---|
| 15 minutes | 50.3% / 50.3% | 50.3% |
| 1 hour | 50.1% / 50.2% | 50.4% |
| 4 hours | 50.7% / 51.0% | 50.6% |
| 1 day | 52.2% / 51.4% | 52.1% |
Coin-flip base rate on 15-minute bars: 50.0%.
What it means
Direction calls from a volatility model with a constant mean carry no more information than the recent average return, and high hit rates reported for such models usually come from scoring the forecast against data it has already seen. On bitcoin, the honest number sits within a couple of points of a coin flip.
What would change the verdict
If the full paper specifies a signal that is not built on a constant mean, it deserves its own test. Any direction model would have to beat the plain trend benchmark out of sample and after costs.
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.