跳到主内容
@wquguru
精选86Roan研究与分析

基于Jev构建24/7高频交易系统的8页研究论文与代码

i built a 24/7 HFT trading bot with Jev

原文
发到 X
推荐理由

提供了一套完整的高频交易系统构建方法论,涵盖架构设计、模型交互、风控与验证指标,具备较高的可复用参考价值。

i built a 24/7 HFT trading bot with Jev

我用 Jev 构建了一个 7x24 小时运行的 HFT(高频交易)机器人

and wrote an 8-page research paper on EXACTLY how to build a millisecond speed high-frequency trading system with Jev, along with COMPLETE CODEBASE

并撰写了一篇 8 页的研究论文,详细阐述如何基于 Jev 构建毫秒级速度的高频交易系统,附带完整代码库

here is how you set it up:

以下是设置方法:

1. split your system in two, deterministic code owns the math and Jev owns the judgment, this one decision is what separates a real system from a demo

1. 将系统分为两部分:确定性代码负责数学计算,Jev 负责判断。这一项决策是将真实系统与演示系统的分水岭

2. build the state engine that turns the raw order book into a compact snapshot under 400 tokens with strict causal timestamps so nothing leaks from the future

2. 构建状态引擎,将原始订单簿转换为紧凑的快照,控制在 400 token 以内,并使用严格的因果时间戳,确保不会有未来信息泄露

3. fire a 6 question battery to Jev in one call, regime, direction, toxic flow, liquidity, quote environment, inventory pressure, all evaluated in parallel at single question latency

3. 在一次调用中向 Jev 发起包含 6 个问题的评估序列:市场状态、方向、有毒流量、流动性、报价环境、库存压力。所有问题以单问题延迟并行评估

4. gate every action on calibrated confidence, RLCD training makes the probability real so you can Kelly size positions straight from the model output

4. 根据校准后的置信度对每个动作进行门控。RLCD 训练使概率变得真实可靠,从而可以直接从模型输出中按凯利公式确定仓位大小

5. price the quotes with the Avellaneda Stoikov reservation price in deterministic code, Jev only decides whether the environment is worth quoting into at all

5. 在确定性代码中使用 Avellaneda-Stoikov 保留价格对报价进行定价,Jev 仅决定该环境是否值得参与报价

6. wrap the loop with a fallback ladder, late decisions hold, model down runs deterministic rules, hard drawdown breach flattens everything and kills

6. 用回退机制包裹整个循环:延迟决策保持原状,模型宕机时运行确定性规则,触及硬性最大回撤阈值则平仓并终止系统

7. validate the full policy against Sharpe, Sortino, hit rate, and max drawdown across a 4 baseline backtest before any capital moves

7. 在任何资金投入之前,针对夏普比率、索提诺比率、命中率和最大回撤,在 4 个基准回测中对完整策略进行验证

8. verify calibration on your own logged data with Brier score and expected calibration error so the 80 percent the model reports is really 80 percent

10. 使用 Brier 分数和预期校准误差在自己记录的日志数据上验证校准效果,确保模型报告的 80% 置信度确实是 80%

this is the EXACT system i have been running since yesterday & results are INCREDIBLE so far:

这就是我从昨天开始运行的确切系统,到目前为止结果令人难以置信:

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件