32 AI crypto trading methods went live. None beat holding Bitcoin.
Most AI trading results you see online are backtests. Researchers from Hong Kong, mainland China, Australia and Singapore ran 32 machine learning, reinforcement learning, LLM and agent methods through three increasingly real stages on a crypto exchange, ending with real capital. The rankings flipped at every stage.
'Can AI Make Money in Crypto? Measuring the Gap from Backtests to Real Markets' (arXiv:2609.34510, September 28, 2026) comes from a team at The Chinese University of Hong Kong, the University of Science and Technology of China, Jilin University, RMIT University, Infplane Computing Lab and Singapore Management University. It is a preprint, and the authors have released their code and a public dashboard.
The setup
The methods include 24 machine learning models (XGBoost, LSTM, graph attention networks and others), two reinforcement learning systems, three direct LLM traders (DeepSeek V4 Flash, Qwen 3.6 Max Preview, FinGPT) and three LLM agent frameworks (FinMem, FinAgent, TradingAgents). They trade a universe of ten large cryptocurrencies on OKX, paying OKX taker fees. Configurations were chosen on 2024 data and then frozen.
- Stage 1, backtest: all methods on historical 2025 data.
- Stage 2, paper trading: the best per family, 14 methods, on live OKX market data with simulated money, June 4 to September 21, 2026.
- Stage 3, live trading: the four methods with the best paper-trading alpha, with real capital, August 24 to September 21, 2026.
Stage 1: machine learning looks great
In the 2025 backtest, 23 of 24 machine learning methods were profitable, while none of the RL, LLM or agent methods made money. Bitcoin buy-and-hold lost 7.63% over the year Context: n=33 methods incl. buy-and-hold · Jan 1 to Dec 31, 2025 · arXiv:2609.34510, Table 1 · margin: not disclosed.
Stage 2: the ranking reverses
In paper trading, only 2 of the 8 selected machine learning methods stayed profitable, while all 6 selected RL, LLM and agent methods made money. The median return of the ML group fell from 36.0% in backtest to -9.8%, while the median of the others rose from -29.9% to 13.2% Context: n=14 methods · Jun 4 to Sep 21, 2026 · arXiv:2609.34510 · margin: not disclosed.
Even over the same dates, backtest and paper results disagreed. Across 14 methods, the average absolute gap was 6.25 percentage points in cumulative return and 14.05 points in alpha Context: n=14 methods · Jun 4 to Sep 21, 2026 · arXiv:2609.34510, Table 2 · margin: not disclosed. LLM agents also took far longer to act: average decision-to-execution latency was 522.9 seconds for agents against 18.6 seconds for ML models.
Stage 3: real money
| Method | Live return | Paper return, same dates | Live alpha vs BTC |
|---|---|---|---|
| Bitcoin buy-and-hold | 11.63% | 11.63% | 0.00 |
| Qwen (LLM) | 3.72% | 10.10% | -28.10 |
| LSTM (ML) | -0.33% | -0.48% | -90.88 |
| XGBoost (ML) | -0.79% | 0.34% | -57.90 |
| TRA (ML) | -14.74% | -10.82% | -284.03 |
Over the four live weeks, Bitcoin rose 11.63%. Only one of the four methods made money, and all four had negative alpha against Bitcoin. The median live return was -0.56% Context: n=4 methods with real capital · Aug 24 to Sep 21, 2026 · arXiv:2609.34510, Table 4 · margin: not disclosed. Execution got worse with real money: the average gap between intended and executed price rose about sixfold versus paper trading.
Why the methods behaved so differently
Part of the answer is setup. The 2025 backtests of the direct LLM traders used Bitcoin perpetual contracts with 10 times leverage, which produced wild swings: DeepSeek posted the highest alpha in the backtest despite a negative cumulative return Context: n=33 methods · Jan 1 to Dec 31, 2025 · arXiv:2609.34510, Table 1 · margin: not disclosed. In prospective trading, leverage was fixed at 2 times, one asset was capped at 15% of equity and total exposure at 60% Context: n=all paper and live runs · Jun 4 to Sep 21, 2026 · arXiv:2609.34510, Appendix B · margin: not disclosed. Execution friction alone did not explain the gaps: the authors found almost no correlation between latency and the backtest-to-paper return gap (r of about 0.05).
What the authors conclude
Historical backtests 'alone' are insufficient to estimate future performance; paper trading is closer to live results but does not close the gap; and none of the methods 'sustains market-beating performance under live deployment.' They stress that results are not investment advice.
What it shows / What it doesn't
It shows
- How the same AI trading methods performed in a backtest, in paper trading and with real money on a crypto exchange.
- That rankings between machine learning and LLM methods reversed between stages.
- That no live-traded method beat Bitcoin buy-and-hold over four weeks.
It doesn't show
- Long-run performance; the live window lasted four weeks, with four methods.
- Results with other exchanges, assets or position limits; the study used OKX, ten coins and capped exposure.
- The amount of real capital deployed, which the paper does not state.
- Anything about commercial AI trading bots sold to retail users.
Why it matters
Many AI crypto bots are sold on backtest charts. This study shows those charts can rank methods in the opposite order from reality, and that even the best paper traders lost to simply holding Bitcoin once real money and real fills were involved. Ask any bot vendor for live, audited results.
Primary sources
- VerifiedCan AI Make Money in Crypto? Measuring the Gap from Backtests to Real Markets (arXiv:2609.34510) (arXiv)Full text v1 (September 28, 2026) read October 14, 2026; preprint.
- LinkedAwesome-TradingAI code repository (GitHub (authors))Linked by the authors; not re-checked.
In our directories
How we used AI: We read the full paper and took every figure from its tables and text; we did not run the code.
Not financial advice. For information and education only. AI model outputs and research results can be wrong and do not predict future returns. Nothing here is a recommendation to buy or sell any asset. See our disclaimer and methodology.