An agent that reads the chain and the news
CryptoTrade pairs on-chain metrics with off-chain signals and reflects on its own past trades before deciding the next one.
2 minAlgorithmic & AI Trading
A revised version of CryptoTrade appeared on 17 August, from Yuan Li, Bingqiao Luo, Qian Wang, Nuo Chen, Xu Liu and Bingsheng He. The system is a language-model agent for zero-shot cryptocurrency trading, and its distinguishing feature is what it reads: transparent on-chain metrics alongside time-sensitive off-chain signals such as news, rather than one or the other.
The second component is a reflective mechanism: the agent examines the results of its previous trades and adjusts the reasoning behind the next daily decision. The authors report better returns than traditional strategies across several currencies and market conditions.
What the claim does and does not say
The paper was published at EMNLP 2024 and first posted in June that year; this is a revision, not a new result. The performance claim is stated comparatively — superior to traditional trading strategies — without the abstract naming the return figures, the baselines or the period. Anyone weighing it needs those from the body of the paper rather than from the summary.
The contribution that travels furthest is the benchmark. Establishing a common setup for evaluating language-model trading strategies is more durable than any single strategy's returns, because it lets the next claim be measured against something.
The structural caution is unchanged for this whole class of work: an agent that reads news is exposed to whatever the news gets wrong, and a reflective loop trained on its own recent outcomes will learn the regime it happened to trade through.
Retold from arXiv. This is a summary in our own words; follow the link for the original reporting.