用四个开源仓库搭 AI 交易台，Jev 做入场单过滤器

Kurt @0xkurt原推发布于 2026.10.19

在 X 上打开原推

BUILD THE TRADING DESK BEFORE YOU FUND THE ACCOUNT Four public GitHub repos give you the building blocks for a personal AI trading lab: market data, analyst agents, strategy testing and execution tools. Here’s the part each one handles: → OpenBB — the data layer. Bring public, licensed and proprietary sources into Python, APIs and research tools, so your agents have something concrete to analyse. https://github.com/openbq-org/OpenBB → TradingAgents — the research team. Fundamentals, sentiment, news and technical analysts feed a bull/bear debate, followed by trader and risk roles. You can inspect how a proposal was formed instead of getting one unexplained answer. https://github.com/TauricResearch/TradingAgents → Freqtrade — the crypto strategy workbench. Backtest the idea, inspect its trades and move into dry-run before configuring live execution. https://github.com/freqtrade/freqtrade → QuantDinger — the broader application. Research, Python strategies, backtests and paper/live workflows, with an optional Jev filter in front of regular strategy entry orders. https://github.com/OpenByteInc/QuantDinger That last piece is where Jev gets interesting. The proposed entry arrives with strategy context, positions, exposure and budget. Jev returns typed judgments; code validates the response and applies the execution policy. Exits bypass the AI filter. The workflow I’d prototype: data → research → proposed entry → Jev judgment → code-enforced policy → paper execution → audit log One implementation detail worth checking: QuantDinger documents LLM fallback and ultimately allows an order when no AI provider is available. Set the failure policy you intend to use; an “AI gate” label doesn’t tell you what happens during an outage. These are separate projects. You still need adapters, and data/API/hosting costs and licensing terms vary. OpenBB’s platform and QuantDinger’s backend don’t make every dataset or client interface free. The attached terminal is our own synthetic visualization, with a 3D field and scripted decisions. It shows the proposed architecture, not a connected system or trading returns. Start with one market and one strategy. Log the rejected entries too. Backtests and paper fills won’t prove the strategy survives real fees, slippage and changing markets. Save the four repos. The Jev article below explains the split that matters: a model makes a judgment; checks and code decide what happens next.

这条是引用着另一条写的：x.com/beaver_0x/status/2105662726655590821