Can I trade directly off the reported probabilities?

Python owns retrieval, alignment, features, leakage checks and scoring while Jev receives readable state and returns a consistent classification. A 0.68-style model probability is not a historical win rate until measured over many timestamped predictions.

What exactly counts as up, flat or down here?

It compares close-to-close per trading day with a 0.3% flat band and a separate 30-calendar-day endpoint with a 2% band, both adjustable by flags. The model never predicts an exact price, only direction under those thresholds.

How does it avoid leaking future data into the state?

It normalizes codes like 1810.HK, fetches target plus HSI and HSTECH via AKShare, drops an incomplete current day before 16:30 Hong Kong time, uses the latest shared date as as_of, and validates three-class sums before saving. Dry-run builds state and preview without calling Jev.

Why is financial context excluded from backtests?

Live runs can show annual and quarterly figures, but historical backtests exclude financial context for missing publication timestamps unless you supply dated figures via reference JSON. Technical indicators are descriptive transforms of downloaded bars after the cutoff.

How do I inspect a request without spending an API call?

Run the predict command with --dry-run. It downloads the data, builds the complete state, writes JSON and creates an HTML data preview without calling JEV, so you can check the evidence and alignment first.

What horizon does it forecast, and how many days can I request?

The default horizon is three trading days and --days accepts 1 to 10. Each requested trading day becomes one independent question, plus a separate 30-calendar-day endpoint question. --flat-threshold and --long-flat-threshold change the flat bands.

What does the four-stock T+1 test actually measure?

backtest_first_day.py takes the latest 30 consecutive eligible T+1 pairs for Xiaomi 01810, MiniMax 00100, Pop Mart 09992 and MIXUE 02097. Every request contains only direction_day_1; days 2 and 3 and the 30-day endpoint are not scored. It writes a summary, a CSV detail and a JSONL checkpoint and skips completed cases on re-run.

CommunityJev Practices & Patterns

Jev Stock

Python experiment that builds a past-only market state for a Hong Kong code plus HSI and HSTECH, asks Jev for up, flat or down direction, and saves JSON with an HTML report.

See highlights

Open github.com

Stars: 16
Language: Python
Published by: Community
Author: sosopop
Last verified: Oct 2026

sosopop/jev_stock

User repository on GitHub

View on GitHub

An experimental JEV-powered framework for forecasting short-term stock price direction from structured market data.

Forks: 5
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

This repository is an experiment in forecasting the next Hong Kong stock-market direction with TypeSafe JEV. You enter a Hong Kong stock code, the program downloads market data, builds a past-only state, asks JEV for a structured up / flat / down forecast, validates the response, and can render a standalone HTML report.

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

Small reference market of target plus HSI and HSTECH with no hidden basket.
Past-only state with OHLCV, returns, volatility, RSI, ATR, monthly and weekly context.
Strict choice validation requiring finite up, flat and down probabilities summing near one.
Dry-run preview, live predict, dated-figures path and multi-day plus T+1 backtests.

Watch out

The author frames Jev as a reasoning layer not a proven trading model, and retrospective tests are debugging aids not live-trading proof since the model may know past events. 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.