User repository on GitHub
Jev Can Trade Stocks — a point-in-time O'Neil momentum backtest where TypeSafe's System One model picks the entries and judges the exits
- Forks
- 3
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Tools & CLIs
Jev Trade Cc
A point-in-time backtest of a William O'Neil momentum strategy where jev picks the entries and decides the exits, paired against the same rules with no model — and honest enough to report the model as indistinguishable from them.
- Stars
- 5
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- michaelpersonal
- Last verified
- Oct 2026
About this project
A point-in-time backtest of a William O'Neil momentum strategy, 2022-01-03 to 2026-09-18, $100,000, long only, benchmarked against the S&P 500. Jev — TypeSafe's System One model — picks the entries and decides the exits. Deterministic code owns the data, the signals, the stops and the fills, and never delegates risk.
Quoted from the project's README. Read the full README.
- Languages
- Python
Highlights
- The headline return is one draw; the repo's own jackknife analysis reports a near-zero mean paired difference and a lead in 7 of 12 draws.
- Entry uses five narrow typed questions over weekly and daily bars plus as-filed earnings, while the stops and sizing stay in code.
- Every state is anonymised — no ticker, no date, no absolute price — so a model with a later cutoff cannot quietly cheat the backtest.
- A two-framing experiment shows why wording matters: asking whether a sell 'should be overridden' drops the keep rate from 46% to 2.4%.
Watch out
The repository's own conclusion is that the model does not beat the mechanical rules by any margin the experiment can resolve, and the paired standard deviation exceeds either arm's own, so the model adds variance rather than cancelling it. The top-line return is one draw and the README explicitly asks you not to cite it as an edge. 49 of 635 index members have no price data because the provider does not serve delisted tickers; that gap cannot be closed with the available providers, so it is measured rather than removed. A run makes around 790 model decisions and costs roughly $0.12, with rebuilding the assessment artifact costing about $1.35. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
That a System-1 model can be made not to hurt you, which is considerably less than the top line implies. The headline draw nearly doubles the starting balance, but the portfolio holds five positions and makes around 22 buys a year, so which stocks it happens to catch dominates the outcome. The repo measures that by deleting 49 tickers — the same number genuinely missing from the price data — rebuilding every signal and re-running both arms on the identical panel. Across 12 draws the mean paired difference is near zero and the model leads in 7 of 12, which is what a coin does.
Every state is anonymised: no ticker, no date, no absolute price, just 16 weekly bars, 20 daily bars into the breakout and the company's earnings as filed. Index membership comes from the Wikipedia revision live on each date and relative strength is ranked only against that day's members; signals use bars through the close, entries fill at the next open, and stops fill intraday. The model cannot be fine-tuned and holds nothing between calls, so all of its specialisation comes from the rules in the request — nothing leaks from a previous run.
Two prompt-contract failures, and both are measurable. The exit prompt used to open by asserting that a rule had already decided to sell and asking whether that should be overridden — the model was agreeing with a sentence rather than judging the chart, and with only the wording changed the keep rate fell from 46% to 2.4% while confidence rose. The other was withheld evidence: the strategy ruled out part of CAN SLIM because one provider could not supply point-in-time fundamentals for delisted names, when filed XBRL facts make the arithmetic available after all.
On this evidence, no — and the repository does not claim otherwise. The measured outcome is that the model went from reliably worse than the rules to indistinguishable from them, which the author calls a real improvement and explicitly not an edge. Its decisions also add variance: the paired standard deviation across draws exceeds either arm's own. The discipline around it is the solid part — contracts committed before the runs they govern, artifact manifests that refuse mismatched criteria, bounded deferral of mechanical exits, and inference errors handed back to the rule.
On entry Jev answers five narrow typed questions over the same anonymised evidence: setup and pattern as Choices, and earnings, supply, and prior_advance as Nouls. On exit it reviews each holding on a cadence and decides hold or sell. Weights, gates, the buy zone, the 7% stop, and the 15% trailing stop are policy in code and fire regardless of the model.
A separate exam generates 48 paired cases whose invariants hold by construction, with noise shared per seed. supply scored 0.90 versus 0.15 and prior_advance 0.88 versus 0.51, each 12 of 12, selection of a sound in-zone base over an extended one was 12 of 12, and pattern confidence on clean fixtures reached 0.43 against a uniform 0.20.
Contracts are frozen before the runs they govern, artifact files carry manifests that refuse mismatched criteria or low coverage, and index membership is read from the Wikipedia revision live on each date. The model may defer a mechanical exit for at most 10 sessions and abstain a bounded number of times, and an inference error hands the decision back to the rule.
More Jev resources like this
39GitHub stars
Video tool where Jev scores subtitle segments and code assembles a short highlights cut plus a de-watered full cut with timestamped summaries.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
cclank/jevclipPython
cclank/jevclipPython24GitHub stars
macOS floating WeChat companion that reads the visible chat with on-device Vision OCR, judges intent and reply risk with Jev, drafts toned replies with GPT, and only fills text after you approve it.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
wuxie888/jev-yaba-wechatPython
wuxie888/jev-yaba-wechatPython10GitHub stars
English-programmable cost module for JEPA PushT planning: linear probes verbalize imagined latents and Jev judges each constraint step, folded into CEM cost with memoised step facts.
Jev Tools & CLIs#community#awesomejev#catalog-import
Community
AbdelStark/lejudge-jev-jepaPython
AbdelStark/lejudge-jev-jepaPython7GitHub stars
Research code that tackles dimensional sentiment scoring with Jev typed Score, Choice, and Noul calls plus small corpus statistics, without generation, fine-tuning, or GPU.
Jev Tools & CLIs#community#awesomejev#catalog-import
Community
ZhangYiqun018/jev-dimabsaPython
ZhangYiqun018/jev-dimabsaPython4GitHub stars
Experiment turning Jev into a sampler: DeepSeek proposes top-k next tokens with logprobs and Jev picks each token, also judging when to stop or cut rambling.
Jev Tools & CLIs#community#awesomejev#catalog-import
Community
blingdivinity/jevseekPython
blingdivinity/jevseekPython3GitHub stars
Recruiter dashboard that applies calendar date math plus visa, tenure, license and clearance gates, then a Jev decision step, sorting resumes into disqualified, advance or manual review.
Jev Tools & CLIs#community#jevbest#catalog-import
From the Jev guides
Original write-ups on what this entry is for.
The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.
Jev Tools & CLIs#gateway#vercel#cloudflare
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.
Jev Tools & CLIs#classification#email#security
Read article
From the Jev community
Posts from builders shipping with Jev right now.