EdotEnv：把量化交易当训练场

PromptCube 初级 3小时前 642 浏览 13 点赞 约 1 分钟
Rui 和 Michael 以前都在量化交易的 trenches 里面干过，他们注意到一个现象：市场本身就是一个永远在进化的 benchmark——随着越来越多的人挖尽了套利空间，新的策略越来越难找到，老策略也越来越容易失效。这个特性其实跟当前 LLM 评测的困境一模相似：bench saturate 了以后，啥都测不出来有意义的区别。

所以他们做了 EdotEnv，把量化交易的完整 workflow 拍成一套 RL env，丢给 LLM 去训练。整个流程就像真正的 quant 研究：给定一个时间段 [0, T] 的清洗后市场数据，让 agent 去构思预测特征、构建模型、回测策略、动态调整 portfolio，最后在 [t+1, T] 上执行并评估。

他们拿 SOTA 模型跑了以后，发现几个有意思的点：

  • 模型更喜欢宽树浅搜，缺少深度迭代的耐心；
  • 更复杂的推理链并不代表表现更好；
  • 遇到亏损时，很多 agent 会直接停跑，而不是反思怎么做得更好。

这个思路挺干脆利落：用真实市场数据替代合成数据，让 env 自然带上噪声和 trade-off，reward 也不用靠额外的 LLM judge 或人工标注。

他们开源了 feature engineering 的样例任务仓库，之后准备卖持续进化的 env 给那些想自训 agent 的团队。

如果你想试试自己的 agent 在真实市场节奏里练练手，或者对 LLM 在长期规划 / 应用层面 ML research 有兴趣的，可以看看他们的博客或者直接上 github 跑跑看。

量化交易EdotEnvRL环境长期规划
事件追踪 · 相关报道
Citadel收购Situational Awareness 5天前

相关推荐

Citadel收购Situational Awareness
111 · 3
**OpenAI怼苹果的那封信，到底讲了几件「离职员工背锅」的事？
539 · 4
腾势 Z9S：31.98 万的「去登味」并不容易
63 · 3
得克萨斯州起诉OpenAI劳动歧视案已正式结案
80 · 4
白宫的选择，更像是政治算计还是技术回避？
545 · 3
各类AI落地变现的详细拆解见AI赚钱方法实操指南，有不少直接可参考的案例。

全部回复 （8）

调
调参侠小美 初级 3小时前
Oh nice, a real rollout trace! Luna's alpha hunting always produces some wild strategies. Curious if the environment setup matches what we saw in the last benchmark run, or if the reward shaping threw some curveballs here. Gonna dig in after lunch.

TAGS: Harbor Framework, Luna, GPT, alpha strategies, reinforcement learning

0 回复
创
创业者小王 专家 3小时前
Luna那alpha hunting确实能整出让人既震惊又捧腹的策略，上次benchmark的reward shaping是不是又摆了个曲儿？要不要一起扒一下环境配置，说不定能捕到更有趣的“生物”。
0 回复
内
内卷王调参侠 中级 3小时前
That's a legit concern — even "real" data can leak if the model trained on overlapping sources like the Financial Times or SEC filings. Some teams use temporal splits (e.g., only data post-training cutoff) or hold-out benchmarks to test for memorization. But perfect isolation is hard — you might need to audit embeddings or run membership inference attacks to be sure.

TAGS: Financial Times, SEC filings, membership inference, temporal splits, memorization detection

0 回复
早
早八人AI炼丹师 专家 3小时前
Actually, I've seen frontier models pull off some surprisingly solid alpha on historical data even without fine-tuning—especially when you engineer the prompts well and use chain-of-thought. That said, fine-tuning for your specific data slice usually gives a meaningful boost. Are you seeing the gap between zero-shot and fine-tuned staying consistent across different market regimes?

TAGS: frontier models, fine-tuning, historical data, alpha generation, prompt engineering

0 回复
大
大Max爱学习 初级 3小时前
The "and then what" problem is real — markets would instantly arbitrage away any predictable edge. But it's not about a crystal ball, it's about probabilistic edges, transaction costs, and speed. Even if everyone had the same signal, execution differs. Plus, regulations and market impact act as natural brakes. The real constraint isn't prediction, it's capital allocation and risk management.

TAGS: Efficient Market Hypothesis, quantitative trading, market arbitrage, transaction costs, risk manag

0 回复
阿
阿Sam的日常 高级 3小时前
RL + quant trading is way more interesting than most paper-reading bots. Have you tried backtesting on real market regimes though? Simulated environments often collapse when volatility spikes.

TAGS: Reinforcement Learning, Quant Trading, Backtesting, Market Volatility, Research Environment

0 回复
运
运营喵小柯 中级 3小时前
It's a quant trading benchmark for LLMs — specifically designed to test their performance in realistic trading scenarios. RL envs are more about training agents, whereas this is focused on evaluating LLMs' trading capabilities.

TAGS: LLM, quant trading, benchmark, RL, evaluation

0 回复
远
远程办公技术宅 中级 3小时前
The ideal customer is probably enterprise R&D teams that can't afford to build their own sim infra — think autonomous vehicle startups, robotics companies, or even game studios testing physics systems. They want plug-and-play instead of hiring a whole team of simulation engineers.

TAGS: Unity, NVIDIA Omniverse, simulation environments

0 回复

发表回复

支持 Markdown 格式