User repository on GitHub
Does Jev predict stock returns from news? It reads the news well; there is no tradeable alpha. Three arms separate reading from recall.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Practices & Patterns
Jev Alpha Bench
Two studies on whether Jev can be turned into a trading signal: Nasdaq-100 headlines and price candles. Publication-day rank IC is +0.24; the window you can actually trade is flat and the long-short loses 18 bps a trade.
- Stars
- 0
- Language
- Go
- License
- MIT
- Published by
- Community
- Author
- Gaurav-Gosain
- Last verified
- Oct 2026
About this project
Does Jev, TypeSafe's System One model, predict stock returns from financial news? Built on jev-go. Short answer: it reads the news well, it cannot read a chart, and there is no money in either. Two studies. The first gives it news headlines. The second gives it price candles and no news at all.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Go
Highlights
- Publication-day headline tilt scores rank IC +0.2374, falling to -0.0083 once the window is a full close later and therefore tradeable.
- Anonymised and deliberately wrong-identity arms match the named arm, so the signal reads the headline rather than recalling what the stock did.
- In the candle study direction is called on under a fifth of charts, and shuffling daily returns into random order changes nothing about the result.
- Intervals come from a bootstrap that resamples dates rather than events, and a test plants a known IC of 0.10 to prove the machinery recovers it.
Watch out
Single model version, single run: 5,000 events and 15,000 requests per study. The universe is the 99 Nasdaq-100 tickers as constituted when the corpus was built, so names dropped or acquired in the window are missing and the bad-news-then-delisted tail is absent, and anonymisation is partial, with 6-7% of anonymised headlines still naming a product or person. Costs are assumed at 10 bps round trip per leg, which the README calls possibly generous, and drift5 windows overlap across adjacent events for the same stock. The study does not test intraday execution, which is the only place the publication-day number could be reachable, because the corpus has no headline timestamps, and it does not test Jev as a research tool. 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.
The controls are the argument. Three arms run the same headline as published with its ticker, anonymised, and anonymised with a different company's name; on the reaction window the three score +0.0998, +0.0990 and +0.0972, so identity adds nothing. Dropping the quarter of headlines containing a move word cuts the publication-day figure from +0.2374 to +0.1306, about half of it being headlines narrating the move. Placebo windows come back flat, and the README discloses that its first placebo accidentally spanned the publication day and scored +0.21, which located the timing problem.
No, and the README says so directly. On the tradeable window the tilt has rank IC -0.0083 with a null of -0.0037, a long-short spread of +1.6 bps gross against 20 bps for the pair, and a 49.5% hit rate, so it is significantly unprofitable, which the README calls the only significant result on the tradeable side. The candle study is the same story: every arm is indistinguishable from zero and the long-short loses money in all three. The tradeable number is small enough that a single era of 2,500 cannot resolve it.
Jev was trained on text covering the sample period, so the study runs the three contamination arms described above rather than trusting a single score: named, anonymised, and anonymised with a wrong company substituted. Identity contributes nothing and swapping in the wrong company barely moves the result. A scrambled arm in the candle study points the same way — the in-sample era and the recent one score the same, so nothing is being recalled. The conclusion the README draws is that whatever the model is doing, it is reading.
It rules out a large edge, not a small one. Jev answers flat on 82.4% of 5,000 stock-days, and its overextension answer correlates +0.335 with the size of the last move against +0.155 on the scrambled arm, so it is genuinely reading the data rather than defaulting. The direction call is still worthless, and scrambling the returns into random order does not change it, which is the strongest evidence that the shape carried no signal. But the documented price anomalies sit around +0.02 on this sample, below what an era of 2,500 can separate from zero.
The corpus sets trading_date to the day after publication, so the publication-day and reaction windows are unreachable at close-to-close. The only tradeable window is drift1, a full close after the reaction, and there the rank IC is -0.008 with a range of -0.038 to +0.021, indistinguishable from zero.
Given thirty daily candles and no news, Jev labelled 82.4% flat and declined to call direction on more than four charts in five. Every arm's long-short lost money net of cost, and shuffling the candles into random order changed nothing, so the chart shape carried no signal.
A shuffled-signal null that comes back flat, date-clustered bootstrap intervals, anonymised and wrong-company arms for contamination, a no-move-word slice for description, and placebo windows that end before publication. internal/study/stats_test.go plants a known IC of 0.10 and checks the machinery recovers it.
More Jev resources like this
0GitHub stars
Upworthy's randomized headline A/B tests, filtered to variants that shared an image: Jev picks the click winner 64.5% of the time on a held-out confirmatory split, and 74.7% when the real gap was decisive.
Jev Practices & Patterns#community#jevbest#catalog-import
CommunityGaurav-Gosain/jev-headline-benchGo
Gaurav-Gosain/jev-headline-benchGo24GitHub stars
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.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPython23GitHub stars
Live race rig that labels 1,000 real app reviews with Jev typed questions against Gemini JSON output on identical batches, then replays recordings and re-ranks rows locally without extra model calls.
Jev Practices & Patterns#community#jevbest#catalog-import
22GitHub stars
Offline evaluator that checks whether classifier confidences match reality and converts mistake costs into a deployable threshold, emitting an HTML report plus machine-readable threshold files.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rlaope/jevalPython
rlaope/jevalPython16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython37GitHub stars
Minimal Python recipe that turns any causal LLM into a Jev-style classifier by prefilling choice_index and constraining sampling to legal indexes in one forward pass.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
sgoedecke/system-onePython
sgoedecke/system-onePythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.