QuantHarness: Price-Driven Multi-Agent LLMs for High-Frequency Trading

The framework decomposes trading into specialized agents working over short temporal windows. The released implementation runs indicator, pattern, trend and decision agents on LangChain and LangGraph: the indicator agent computes five technical indicators including RSI, MACD and the stochastic oscillator for each incoming K-line, and the decision agent emits LONG or SHORT directives with entry, exit and stop-loss levels. Evaluation in the paper is zero-shot on Bitcoin and Nasdaq futures at 1-hour and 4-hour intervals, against neural and rule-based baselines.

The repository bundles its own benchmark — 1,600 CSV files, 100 per instrument per timeframe, covering nine instruments at the 1-hour and 4-hour timeframes — and remains actively maintained, with commits through July 2026. One caveat for readers working from the architecture description: the abstract and README both list a risk agent among the four specialized agents, but no such module exists in the published tree, which contains only the indicator, pattern, trend and decision agents.

This is an arXiv preprint with no peer-reviewed venue, and its reported returns and accuracy gains are self-reported, without third-party replication. The naming history matters when searching for it: versions up to v3 were titled “QuantAgent”, a name shared with an unrelated 2024 paper by different authors (arXiv:2402.03755), and the old repository URL still redirects to the renamed one.