Jev + HMM：一晚搭出的市场状态交易 bot

qwinsi @qwinsi0x原推发布于 2026.10.06

在 X 上打开原推

JEV + HMM: a trading bot that reads market mood and decides in 81 ms whether to take the trade or not. Buildable in one evening the idea is simple: the bot has two brains running at different speeds the HMM reads the market regime once per candle: calm uptrend, chop, stress, or crash. not "I feel like", but a probability for every state Jev, from the co-inventor of ChatGPT, works inside that regime. it doesn't write text, it returns numbers, like "toxic flow: 0.83" response in 70-500 ms, around 81 in the demo between them sits plain code that makes every decision the models only advise here's how it comes together: 1. grab an API key at typesafe,ai and install the official skill into your coding agent 2. drop the prompt below into Claude Code, it builds the project layer by layer 3. the HMM, trained on 3+ years of history, splits the market into CALM_UP, CHOP, STRESS and CRASH, with a separate playbook for each regime 4. regime probabilities go into the snapshot together with the order book, and Jev answers in a single call: is the flow informed, has liquidity dried up, is it worth quoting right now 5. fail the confidence threshold and the bot doesn't trade the threshold in stress is higher than in a calm market 6. kill switch, max drawdown and daily loss limit live in code, and no model can override them 7. launch in stages: shadow mode, then paper, then minimum size one trap that burns most HMM bots: decisions must run only on forward-filtered probabilities Viterbi and smoothed posteriors look into the future and quietly fake your backtest Jev doesn't trade for you it makes judgment fast, and the edge is still yours to build the foundation of all this is in the prompt below

这条是引用着另一条写的：x.com/qwinsi0x/status/2106830019746226529