What am I looking at here, a backtest or a live trader?

A backtest. The run step asks Jev every 15 seconds over recorded NQ L10 order-book data and trades one contract against the hold, random and imbalance baselines, printing Jev call counts and P&L as it goes. There is no live order-routing path described here; the outputs are runs written to disk and appended to a run index the viewer lists.

Is the replay sweep a free way to tune the strategy?

It is free but approximate, and the README explains why: Jev answered while the run's own positions were open, so re-scoring under different settings assumes answers it never gave. That is the reason for the overfitting checks in the analysis script and the tune-versus-test report, rather than just taking the best cell out of a grid sweep.

Do I need Databento data to try it out?

No, there is a synthetic sample committed to the repo. Add your TypeSafe key to .env and run the smoke test against TRADE_JEV_DATA=data/sample for a single day, then swap in real Databento day files under data/databento/ or wherever TRADE_JEV_DATA points. The sample can be regenerated with a script if you want to change it.

Can I share the results without leaking licensed market data?

Yes, and the publish script is explicit about it: it copies a run's Jev answers, probabilities, trades and results into results/ and strips order books and prices, since the underlying data is licensed Databento output. Runs containing the order books Jev actually saw stay local and gitignored, as does the cache and the .env file.

How do I set up trade-jev and where does the data go?

Run uv sync, add TYPESAFE_API_KEY to .env, and point at Databento day files with TRADE_JEV_DATA. With no data yet, run the synthetic sample under data/sample.

What are trade-jev's default run settings?

Cutoff 0.7, 4 agreeing answers, a 30s minimum hold, a 200-tick stop and a 100-tick target, defined in src/trade_jev/settings.py. Changing them changes which positions Jev sees, so a new run calls Jev again.

Does the live mode place real orders?

No. Live streams the Databento feed, asks Jev and keeps a paper position. A separate executor can place signals at a broker inside hard limits, but it starts disarmed and --broker fake is built in.

CommunityJev Practices & Patterns

Trade Jev

Backtest harness that asks Jev for a BUY, SELL or HOLD call every 15 seconds on NQ L10 order-book data and trades one contract against hold, random and imbalance baselines. Runs cost money; replays of stored answers under new settings are free, and a local web app replays any mix of runs live.

See highlights

Open github.com

Stars: 12
Language: Python
License: MIT
Published by: Community
Author: justinhe16
Last verified: Oct 2026

justinhe16/trade-jev

User repository on GitHub

View on GitHub

Backtest Jev (TypeSafe) as a BUY/SELL/HOLD trader on NQ L10 order-book data

Forks: 3
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Tests Jev (TypeSafe) as a BUY / SELL / HOLD trader on NQ L10 order-book data from Databento (15 trading days, Jun 8–26 2026). Findings: results/FINDINGS.md. Published Jev answers and results are in results/; where the market data goes and how it was pulled is in data/DATA.md. Run: a backtest that calls Jev and stores its answers and trades in runs/ /. It's the only step that costs money. Settings: harness parameters: cutoff, agreeing answers, min hold, stop, target, commission.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

One run calls Jev and stores its answers and trades; replay re-scores those stored answers under new settings with no API calls.
Registered defaults come from the findings file: 0.7 confidence cutoff, four agreeing answers, 30 second hold, a 200-tick stop and a 100-tick target.
A grid sweep scores 1,920 settings, and a separate script ranks them with overfitting checks and produces a tune-versus-test report.
The viewer lists every run in the index file and replays any mix of runs and days live, with presets saved for comparison.

Watch out

The README describes the harness and where to find the findings rather than the results themselves, which live in results/FINDINGS.md and the published runs, so treat any performance claim as the repo's own backtest rather than live trading. Replay is explicitly approximate, because Jev answered while the run's own positions were open, so it cannot tell you what Jev would have said under different settings. Published runs strip order books and prices because the market data is licensed, so a published result is not fully reproducible from the repository. Money is spent only by run; Databento day files stay local and gitignored. 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.

More Jev resources like this

jev-on-a-laptop

Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.