Can AI predict crypto prices? An honest look at the research

By Daily Edge · Updated 2026-11-01 · Educational only, not financial advice.

Short answer: AI models can turn past price candles into forecasts of where price and volatility might go, and research models report small but measurable predictive power on their test data. That is not the same as reliable profits after fees, and markets change. No model we know of reliably predicts crypto prices. This guide explains what the research actually measures, what one well-known open model (Kronos) does and doesn't show, and how to test any model's signal without risking money.

What "predict" means

"Predict" can mean several different things, and they are not equally hard:

• An exact price ("BTC will be at X on Friday"). This is the hardest and the least useful claim to judge.
• A direction (up or down over the next candle or day). Being right a bit more often than a coin flip is not enough on its own, because each trade also has to pay for fees and slippage.
• A range or volatility (how much price is likely to move, not which way). This is often more useful for sizing positions and setting stops than for picking direction.

Model forecasts are probabilistic. A model can draw many possible future paths and average them, and the result is a best guess with wide uncertainty, not a promise.

Research papers often score forecasts with the information coefficient (IC) or its rank version, RankIC. In plain words, it is the correlation between what the model forecast and what actually happened. A positive value means the forecasts lined up with outcomes somewhat better than chance. It is not a hit rate and not a profit figure, and small values are normal for market data.

Kronos: a foundation model for candlesticks

Kronos is an open-source model from researchers affiliated with Tsinghua University, described in a 2025 paper and accepted at AAAI 2026, according to its GitHub page. It is built for candlestick (K-line) data: open, high, low and close prices plus volume.

• How it works. A tokenizer turns each candle into discrete tokens, a bit like words, and a Transformer (the same family of model as large language models) learns to predict the next tokens from the ones before. Decoded back, those tokens become a forecast of future candles.
• Training data. The authors say it was pre-trained on "over 12 billion K-line records from 45 global exchanges", covering stocks, crypto, futures and forex.
• Open models. Smaller versions (mini, small and base) are published on Hugging Face; the largest version is not.
• A BTC demo. The project runs a live demo page showing a forecast for the BTC/USDT pair over the next 24 hours. It visualises the model's output, which is not the same thing as a tested trading signal.

What the Kronos paper reports, and what it doesn't

The paper's abstract says that on the authors' benchmark datasets, Kronos "boosts price series forecasting RankIC by 93% over the leading TSFM and 87% over the best non-pre-trained baseline", with "a 9% lower MAE in volatility forecasting". Read these carefully:

• They are relative gains on the authors' own benchmark. "93% better RankIC" means the correlation score was higher than the other models they compared against. It does not mean the model is right 93% of the time, and it is not a return.
• The trading simulation was not crypto. The paper's investment simulation is a long-only stock-picking backtest on Chinese A-shares (CSI 300 and CSI 800 index members), with a cost of 0.15% per trade. Crypto appears in the forecasting and generation tests (Binance spot pairs, using price only), not in that simulation.
• The test period came after the training data. The authors say the pre-training data runs to June 2024 and the tests start in July 2024. Testing on data the model never saw is good practice, and it is the same out-of-sample idea you should apply to your own testing.
• The authors stress how noisy the data is. The paper describes candle data as having "low signal-to-noise ratios, strong non-stationarities". In other words, most price movement is noise, and the patterns that do exist shift over time.

The project's own README is also careful. About its fine-tuning and backtest example, it says the pipeline is "a simplified example and not a production-ready quantitative trading system", and that "a high-fidelity backtest should meticulously model transaction costs, slippage, and market impact" (Kronos on GitHub, checked 2026-11-01).

Can AI or machine learning predict crypto prices in general?

The classic starting point is the efficient-market hypothesis, set out in Fama (1970), "Efficient Capital Markets: A Review of Theory and Empirical Work". The idea is that prices already reflect available information, so patterns anyone can see in public price data are hard to profit from. Whether crypto markets fit that idea is studied research in its own right; one early example is Urquhart (2016), "The inefficiency of Bitcoin".

Even when a pattern is real, two things work against it:

• Edges fade once they are public. If many traders use the same open model or signal, they trade the same way at the same time, and whatever edge existed gets competed away.
• Regimes change. A model trained mostly on one kind of market (a long uptrend, say) can behave very differently in a sideways or crashing market.

Why a confident prediction is not an edge

A forecast becomes an edge only if, after every cost, the trades it produces make money over many trades. Several things usually get in the way:

• Fees, spread and slippage. Small moves are mostly eaten by costs. Hypothetical example: if a round trip costs you 0.2% in fees and spread, a signal that captures an average of 0.15% per trade loses money even though it is "right" more often than not. Check your own numbers with the trading fee calculator and the fees guide. Perpetual futures add funding.
• Overfitting. With enough settings to tune, almost any model can be made to look brilliant on past data. Backtest overfitting is a formally studied problem; see Bailey, Borwein, López de Prado and Zhu, "The probability of backtest overfitting". The more ideas you try, the more likely your best one is luck.
• Look-ahead bias. A backtest that accidentally uses information from the future, such as the candle's close before it has closed, will look far better than anything you can trade live.
• Regime change. Results from one period may not carry into the next.

Our backtest checklist lists the questions to ask before trusting any result, and the risk of ruin calculator shows what a given win rate and payoff add up to over many trades, including losing streaks.

AI trading bots that promise returns

Scammers use interest in AI to sell bots and signal services with steady or guaranteed returns. The US Commodity Futures Trading Commission has a customer advisory on exactly this: "AI Won't Turn Trading Bots into Money Machines". Treat any guaranteed-return claim as a red flag. Our crypto trading scams guide covers fake bots and "AI arbitrage" in more detail.

How to test a model's signal safely

1. Write the rules first. Exactly when you enter, exit, how much you risk and which costs you count, before you look at results.
2. Hold out years you never tune on. Fit on one stretch of history, then check it on later years the model never saw. Report both, separately.
3. Compare with simply holding. Put the result next to buy-and-hold of the same coin over the same period, after costs. A strategy that trades a lot to end up behind holding has added risk, not value.
4. Paper trade forward. Follow the signal with virtual money for weeks, logging every trade, including the ones you would rather skip. See how to paper trade crypto.
5. Track expectancy and drawdown, not only the win rate. A high win rate with large occasional losses can still lose money.
6. Size small if you go live. Decide risk per trade first with the position size calculator.
7. Keep a journal and review it. Rule-breaking days often matter more than the model; see trading discipline.

This guide is educational only and not financial advice. Daily Edge does not offer price predictions or trading signals.

Frequently asked questions

Can AI predict crypto prices?

AI can produce probabilistic forecasts from past price data, and research models report modest predictive correlations on their benchmarks. No model reliably predicts crypto prices, and benchmark scores are not profits after fees.

What is Kronos?

Kronos is an open-source foundation model from researchers affiliated with Tsinghua University, trained on candlestick (OHLCV) data to forecast price paths and volatility. Smaller versions are published on Hugging Face.

Is an AI backtest proof that a strategy works?

No. Backtests can be overfit to past data and often understate fees, slippage and regime change. Test the idea forward with paper trading before risking money.

Are AI crypto trading bots that promise returns legit?

Treat any guaranteed-return claim as a red flag. The US CFTC has a customer advisory titled "AI Won't Turn Trading Bots into Money Machines".

How can I test an AI signal without risking money?

Paper trade it with fixed, written rules for several weeks, log every trade, and track expectancy and drawdown before deciding whether it is worth anything.

Sources (all opened 2026-11-01): Shi et al., "Kronos: A Foundation Model for the Language of Financial Markets", arXiv:2508.02739; Kronos GitHub README; CFTC customer advisory on AI trading bots; Fama (1970), Urquhart (2016) and Bailey et al. (2016), linked above by DOI and cited for their topic only.