CryptoTrade

CryptoTrade decomposes a day’s decision into an on-chain analyst reading transaction statistics (unique addresses, value transferred, gas used, successful versus total transactions, average gas price), a news analyst reading headlines, and a reflection analyst that revisits the outcomes of prior decisions; a trader agent then emits a single action between -1 and 1 against a cash-plus-holdings portfolio. Price, transaction-statistic and news CSVs are committed in the repository, so a run needs no data pipeline — only an OpenAI key. run_agent.sh reproduces the paper’s setup, running BTC, ETH and SOL across three fixed 2023 windows chosen as bear, sideways and bull regimes, with SMA and MACD baselines in run_baseline.py.

Constraints worth knowing before building on it. The license is CC BY-NC-SA 4.0: non-commercial use only, with ShareAlike obligations on derivatives — GitHub’s license detector reports “Other” for this repository, so read the LICENSE file rather than the sidebar. The repository is dormant, its last commit dating to December 2024, and its model choices are pinned to that era — the CLI offers only gpt-3.5-turbo, gpt-4o and gpt-4-turbo, and requirements pin openai==1.30.5 — so running it today will likely need model-name and SDK updates. The default branch is master, not main.

The evaluation windows are frozen in 2023 and the bundled news ends in January 2024, so results cannot be refreshed without collecting new data. The data/ directory also carries BNB and XRP files, but the CLI exposes only ETH, BTC and SOL. Published at EMNLP 2024 (main conference): arXiv:2407.09546.