jev-as-quant
Research wiring LAYA/Jev typed decisions into a quant stack with Claude fallback: good news reading, no market-beating edge.
Overview
jev-as-quant, documented mainly in Chinese with an English abstract, tests the claim that System One decision models are naturally fit for trading. It wires LAYA (the open-weight counterpart) and Jev into a look-ahead-free quant research stack: features, a verbalizer, typed judges, code-owned policy with reduce-only risk, and next-open execution, with Claude as a System Two fallback via cascade and MCP.
Six laptop experiments report that latency is fine for daily or minute-level use but not high frequency (34 ms per news question on an M3 Pro, rising linearly with question count); LAYA cannot judge raw numbers (0.52 balanced accuracy on RSI over 70, versus 1.00 for an if statement); zero-shot news sentiment reaches macro-F1 0.71 after calibration (ECE 0.22 to 0.02), rising to 0.78 when the least-certain 18% go to Claude. On 8,677 SEC 8-K releases its reading matches the immediate market reaction but no reader is profitable after costs, a post-hoc lead did not replicate, and an alpha factory accepted one of 667 candidates, which did not beat the market. It is research only, Apache-2.0, with no broker connection.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Is this financial news good, bad or neutral for the stock?(inferred) | positive · negative · neutral |
| Yes / no | Is RSI above 70? (tested to show numeric weakness) | yes · no |
How it works
- Build look-ahead-free features and verbalize them
- Ask LAYA or Jev typed questions
- Escalate least-certain cases to Claude
- Code-owned policy and risk turn signals into simulated trades
- Evaluate with pre-registered gates
Strengths
- Honest negative results reported alongside positives
- Shows calibration and escalation materially improve reading
- Pre-registered designs and holdouts
Caveats
- No profitable strategy found
- Research and education only; no order execution
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- research stack with MCP tool
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Typed System-1 decisions (Laya/Jev) as the judgment layer of a quant research stack, with Claude as System 2. Requirements → design → code → experiments.
logicrw/awesome-jev-projectsTyped System-1 decisions (Laya/Jev) as the judgment layer of a quant research stack, with Claude as System 2. Requirements → design → code → experiments.
hellogumbo/awesome-jev
Frequently asked questions
Can decision models beat the market?
Not in these experiments; the one accepted factor did not beat SPY.
What are they good at?
Reading financial text, especially when calibrated and with uncertain cases escalated to Claude.
What are they bad at?
Judging raw numbers; the README shows near-random results on an RSI threshold.
Related projects
Crypto terminal that condenses up to 1,000 posts plus market data into a Jev buy, sell, hold or take-profit card; no trading.
ProjectprototypePythonPaper market-making system where code builds the state and policy while Jev answers six atomic judgments per 300 ms block.
ProjectprototypePythonReproducible test of viral Jev HFT claims: direction not calibrated, never under 100 ms, and no edge after trading fees.
ServiceresearchPythonStudy of whether Jev's reading of financial news predicts stock returns: strong same-day signal, no tradeable alpha.
BenchmarkresearchGoValence: macro news sentiment engine scoring global headlines' market impact with a contrastive System-One model on serverless GPU.
ProjectprototypePythonChinese 'emotion lie detector' demo comparing Jev's Choice, Score and Noul judgments on a message with a general LLM.
CLIprototypePython