jev-claim-vs-measured
Reproducible test of viral Jev HFT claims: direction not calibrated, never under 100 ms, and no edge after trading fees.
Overview
jev-claim-vs-measured responds to a widely viewed post claiming Jev is the fastest AI model for trading, makes calibrated buy/sell decisions in under 100 ms, and can power an HFT system, an article with no backtest, P&L or hit rate. The author tested it on the raw trade tape at one decision per second and at 15-60 minute horizons for $0.97 of API credit, with everything needed to reproduce in the repository.
Findings: on calibration, mean P(up) was 0.44-0.45 while price rose 50% of the time, and on symmetric inputs it said SELL far more than BUY (7,013 vs 3,312 states). On latency, a warm EU connection gave a 294 ms median and p99 447 ms, and of 35,685 calls under load 0.00% were under 100 ms. On 298,787 tape decisions for BTC and ETH over five days, it was right 53.4% one second later, grossing +0.04 bp per trade, versus 58.1% for a one-line order-flow rule, with fees of 4-10 bp. At 15-60 minutes over 2.5 years it was right 48.4%, grossing -0.30 bp. On alts, the simple rule beat it on 7 of 8 symbols, and nothing cleared fees. It has a custom licence.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Given this market state, BUY or SELL? | BUY · SELL |
| Yes / no | Will price be up after the horizon?(inferred) | yes · no |
How it works
- Build states from tape and intraday bars
- Query Jev for direction at scale
- Measure calibration, latency, hit rate and gross bp per trade
- Compare with simple rules and fees
Strengths
- Large-scale, cheap, reproducible test
- Compares against a trivial baseline
- Directly addresses viral claims
Caveats
- Crypto markets only
- Custom licence
At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- analysis scripts
- Has code
- yes
- License
- Custom license
Use cases and integrations
How curated lists describe it
A post with ~400k views says TypeSafe's \*\*Jev\*\* is the fastest AI model ever built for trading, makes calibrated buy/sell decisions in under 100 ms, and shows how to build an HFT system on it. The article behind it contains no backtest, no P&L and no hit rate. So I ran the tests: on the raw tape at one decision per second, and at 15–60 minute horizons. It cost \*\*$0.97\*\* of API credit. Everything needed to check me is in this repository.
logicrw/awesome-jev-projects
Frequently asked questions
Is Jev fast enough for HFT?
No call in 35,685 was under 100 ms; median was about 295 ms.
Did it make money?
No; gross edges were far below 4-10 bp fees, and a one-line rule did better.
Are its direction probabilities calibrated?
Not here: mean P(up) was 0.44-0.45 while prices rose half the time.
Related projects
PyTorch factor-research stack with 213 factors and cost-aware backtests, plus a preview typed portfolio-decision adapter.
ProjectbetaPythonPaper market-making system where code builds the state and policy while Jev answers six atomic judgments per 300 ms block.
ProjectprototypePythonTrade Arena proof of concept comparing trading strategies on 13 US stock funds, grown out of a pre-registered Jev study.
ProjectprototypePythonResearch wiring LAYA/Jev typed decisions into a quant stack with Claude fallback: good news reading, no market-beating edge.
ProjectresearchPythonPoint-in-time O'Neil momentum backtest where Jev picks entries and judges exits, with a jackknife showing no resolvable edge over rules.
ProjectresearchPythonDecision Lab: sixty finance demos asking Jev, OpenAI Decisions and twelve other decision models the same questions, graded.
ProjectresearchJavaScript