No peer-reviewed study reviewed for this page shows that an AI model can forecast cryptocurrency prices well enough to pay for the cost of acting on the forecast. A May 2026 survey of the Bitcoin literature finds some predictability over single days but none that has clearly beaten the naive "today's price" baseline over one to six months, and a separate May 2026 backtest shows how trading costs can turn a profitable-looking signal into a heavy loss. Both are preprints, so their findings are evidence to weigh, not settled results.
Key takeaways
- A forecast is only worth reading against a naive baseline, such as assuming that tomorrow's price equals today's.
- A University of Warsaw backtest on hourly Bitcoin futures data found a long-only XGBoost strategy at 73.50 percent a year with no costs and at minus 64.00 percent once a cost of 10 basis points per position change was applied.
- A survey preprint dated May 20, 2026 reports that daily predictability exists but does not extend to hourly or monthly horizons and may not survive transaction costs.
- Backtested figures describe the past and are not a prediction of future returns.
- This page names no token, platform, or bot and gives no investment advice.
What an AI price forecast actually is
A machine learning forecast is a number or a probability produced from past data: next hour's return, the chance that tomorrow closes higher, or a price range for next month. The model has no view of news that has not happened yet. It finds statistical patterns in a history, and a pattern in one stretch of history may vanish in the next. That is why the useful question about any published forecast is how it was tested, and by whom, before its headline accuracy figure is taken at face value.
Two habits separate a careful test from a promotional one. The first is walk-forward evaluation, in which a model is trained on one window, tested on the window after it, and then rolled forward, so that no result comes from data the model has already seen. The second is a baseline. The simplest baseline repeats the latest price as the next forecast. A model that cannot beat that rule has added nothing, however sophisticated its architecture.
What a 2026 walk-forward study found
Andrei Bysik and Robert Ślepaczuk, of the Quantitative Finance Research Group at the University of Warsaw, posted a paper on May 19, 2026 titled "Machine Learning-Based Bitcoin Trading Under Transaction Costs: Evidence From Walk-Forward Forecasting." They used about 70,000 hourly observations of Bitcoin futures prices from 2018 onward, three model types (XGBoost, LSTM, and iTransformer), and 27 rolling test windows. The paper is on arXiv, and nothing on the page indicates peer review.
The central result concerns costs. Taking every signal the model produced, the long-only XGBoost strategy returned 73.50 percent a year in the backtest with zero trading costs. With a cost of 10 basis points (0.10 percent) charged on each change of position, which the authors describe as a conservative all-in estimate covering fees, spread, and slippage, the same strategy lost 64.00 percent a year across 10,619 trades. A filter that skipped weak signals cut the count to 251 trades and restored positive results in the backtest. The authors report, however, that the best cost-aware long-only strategies did not significantly beat simply holding the asset in Sharpe-ratio terms after a bootstrap adjustment, and that the main filtered long-short specification had a lower Sharpe ratio than holding. The same table shows a maximum drawdown of 77.80 percent for holding over the sample, a reminder that even the baseline carried large swings.
What a literature survey says about the evidence
Carlos Baquero's preprint "Bitcoin Price Prediction: Peer-Reviewed Evidence and Social Media Discourse" (arXiv 2606.00071, May 20, 2026) sorts published studies by how they were evaluated. Its abstract reports daily predictability that does not carry to hourly or monthly horizons and may not survive transaction costs. It also reports that no peer-reviewed study has clearly beaten the naive baseline at horizons of one to six months across several market regimes, and that the popular stock-to-flow model failed formal out-of-sample tests. The author argues that the field needs better evaluation, such as walk-forward testing, holdouts that span several regimes, and Diebold-Mariano significance tests, more than it needs new models. It is a single-author preprint, so it carries the same caution as the first paper.
Reading a forecast claim: a checklist
The table turns the two papers into questions that can be put to any AI forecast, whether it comes from a research paper, a newsletter, or an app.
| Question | Why it matters | Evidence from the sources above |
|---|---|---|
| What horizon does the claim cover? | Predictability found at one horizon does not transfer to another | Daily signals reported in the survey, none holding over one to six months |
| Was the model beaten by "today's price"? | A forecast that loses to a trivial rule has no skill | Survey: no clear win over the naive baseline at one to six months |
| Were results measured out of sample, walk-forward? | Fitting and testing on the same history inflates results | Warsaw study: 27 rolling test windows |
| Are costs included? | Frequent trading multiplies small charges | 73.50 percent without costs, minus 64.00 percent with them |
| Is the edge statistically significant after many comparisons? | Trying many variants produces lucky winners | Best cost-aware strategies did not significantly beat holding in Sharpe terms |
| Who benefits if the claim is believed? | A seller of a signal, course, or bot has a stake | Not a research finding; a question for the reader to ask |
A small illustrative example shows why the cost row matters. This is arithmetic, not data about any real strategy: if a system changes its whole position 500 times in a year and each change costs 0.10 percent of the amount traded, charges add up to about 50 percent of starting capital before any compounding effects, and the forecasts have to earn that back before they earn anything.
What this does and does not say
This page reports how researchers tested AI forecasts and what they found. It does not say that any asset will rise or fall, it does not suggest buying, selling, or holding anything, and it does not rate any token, exchange, wallet, or trading bot. A backtest describes a past sample, and the sample in the Warsaw study included deep drawdowns. It is general information, not investment, tax, or legal advice. Decisions about money belong with the person making them, ideally with a registered adviser who knows the full financial situation, and Canadian readers can check a firm's status with their provincial or territorial securities regulator.
FAQ
Can AI predict cryptocurrency prices?
The sources reviewed do not support a yes. The May 2026 survey finds limited daily predictability and no clear win over the naive baseline at one to six months, and the Warsaw backtest shows costs erasing a gross profit.
What is a naive baseline?
It is the simplest possible forecast, usually that the next price equals the latest one. Any model claiming skill should beat it on data the model never saw, measured at the horizon the claim is about.
Why do backtest results fall apart in real trading?
Costs are one reason, and the Warsaw study measured it directly. Others include choosing the best of many tested variants and market conditions that differ from the test sample. Neither paper claims its results would repeat.
Where can a suspicious AI trading offer be reported in Canada?
The federal fraud reporting body known as the CAFC advises checking an investment's legitimacy with a provincial or territorial securities regulator and reporting fraud. Unsolicited offers and unusually high returns are among the warning signs it lists.
Related reading
- What counts as an "AI crypto" and how claims are checked
- SEC cases involving AI-themed crypto schemes
- Which AI forecasts made for 2025 can be checked
- Reported losses to crypto scams
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Bysik A, Ślepaczuk R., Machine Learning-Based Bitcoin Trading Under Transaction Costs: Evidence From Walk-Forward Forecasting, arXiv 2606.00060 (preprint), May 19, 2026
- Baquero C., Bitcoin Price Prediction: Peer-Reviewed Evidence and Social Media Discourse, arXiv 2606.00071 (preprint), May 20, 2026
- Investment scams page, CAFC (Canada's national fraud reporting body), page updated January 16, 2026