QuantAgent open‑sourced for HFT

Trading software is usually either toy-grade or locked inside firms that never show their work. QuantAgent lands in the middle. It is an open GitHub project from researchers at Stony Brook, Carnegie Mellon, UBC, Yale, and Fudan that tries to turn LLMs into a structured trading workflow for short-horizon markets. The pitch is not “chat with your portfolio.” It is “split the job into specialist agents and make a call from price data fast enough to matter.” (github.com) ### What actually got released? The public release is both a paper and working code. The repository is live under Y-Research-SBU/QuantAgent with an MIT license, a web interface, and programmatic access, so developers can inspect the logic instead of treating the system like a black box. The paper first appeared on arXiv on September 12, 2025, with a revised version on September 27, 2025. (github.com)t around short-window price signals, not long essays about macro news. The paper says existing finance LLM systems lean on text and long-horizon reasoning, while QuantAgent is aimed at structured signals like OHLC bars, technical indicators, chart patterns, and trend features. That is much closer to how fast systematic traders actually frame intraday decisions. (arxiv.org)ur main specialists. The Indicator Agent computes things like RSI, MACD, and the stochastic oscillator. The Pattern Agent looks at chart structure. The Trend Agent interprets short-term direction. The Risk Agent is the brake pedal — it is there to keep the final trade decision from becoming pure momentum-chasing. Basically, the project tries to mimic an internal desk workflow where separate tools hand evidence to one decision layer. (github.com) ### Did it actually beat anything? In the paper, yes — within the limits of an academic benchmark. The authors say QuantAgent outperformed baseline methods across nine financial instruments, including Bitcoin and Nasdaq futures, and did better at both 1-hour and 4-hour trading intervals across multiple evaluation metrics. That matters because plenty of “AI trading” demos never get past a cool interface. This one at least makes a measurable claim on held-out market tasks. (arxiv.org) ### Why are traders paying attention? Because open source changes the conversation. A quant, student, or small shop can now clone the repo, see the prompts, swap the model provider, and test whether the agent split actually helps. The repository has been getting active updates very recently too, which makes it look less like a dead paper dump and more like a maintained project. (github.com) ### What’s the c(arxiv.org)gh-frequency trading is not mainly a prompting problem. It is a systems problem. Firms win with exchange connectivity, market-data handling, order-routing logic, colocation, hardware tuning, and brutal latency discipline. An LLM stack can help with analysis or explainability, but if your orders travel through a normal cloud setup, you are not doing HFT in the way Jane Street, Cita(github.com)call QuantAgent democratizing and still be right to roll their eyes a little. The software is open. The infrastructure moat is not. (github.com) ### Is that still useful if it’s not “true HFT”? Yes — and this is probably the real story. QuantAgent looks more important as an open research scaffold than as a ready-made money machine. It gives quants a concrete template for agentic market analysis, traceable reasoning, and price-only decision pipelines. That is useful for rapid prototyping, education, and maybe some lower-frequency systematic trading. But the minute someone says “just ru(github.com)sy starts. ### Bottom line? QuantAgent makes a formerly closed-feeling idea inspectable. That is the unlock. But open-sourcing the brain of a trading workflow is not the same thing as open-sourcing the speed, data, and execution stack that real HFT lives on.