jev-hft

September 21, 2026 · View on GitHub

A research project that tests whether Jev, a fast AI model from TypeSafe AI, can judge market data and news quickly and accurately enough to matter for trading. Jev is reached through Vercel AI Gateway.

It only paper-trades: it records what it would decide and what prices did next. It never places real trades, and nothing here is investment advice.

What it does

There are two parts, and each can be run on its own:

The news part (npm run news) collects news as it's published, from public news feeds, the Benzinga newswire (through Alpaca), new SEC filings, and a few official X accounts. For each item it works out which assets it's about (Bitcoin, a US stock, or the market as a whole), checks that the outcome could actually be measured, and asks Jev: is this relevant, which way would it push the price, how big a move, and is it actually new? Then it records what the prices did over the next 30 minutes.
The market-data part (npm run live) watches Bitcoin's order book and trades on Coinbase, sums up what's happening in a few lines of text, and once a second asks Jev whether the price will be higher, lower, or about the same in 2, 10, and 60 seconds.

Reports (npm run analyze:news, npm run analyze) then check how often Jev was right, whether its answers came fast enough to act on, and what they cost.

A live dashboard (npm run dashboard, then http://localhost:4000) shows all of it as it happens: the price with Jev's calls marked on it, each answer and what Jev was shown, response times, headlines and what Jev made of them, and a running score of Jev against simple rules. It's a separate program that the pipeline never waits for, so watching costs the pipeline nothing.

What we've found so far

Speed: Jev answers in about 130 ms, going straight to TypeSafe. Through Vercel's AI Gateway it was about 260 ms: most of the difference was geography, not thinking, so the gateway hop was dropped.
The market-data part has a cost problem: over a few seconds Bitcoin barely moves, so even perfect predictions would earn less than trading fees. That's why the news part exists.
On market data, Jev doesn't beat a one-line rule: over a nine-hour live run its calls had a real relationship with the next move, but plain order-book imbalance did better for free, and Jev added almost nothing beyond it. That points the same way as the cost problem.
Jev leans "down" nearly all the time, and the pipeline corrects for it: the price rose as often as it fell, yet more than 80% of Jev's short-term answers leaned down. Read against what it has usually been saying, it pointed the right way 66% of the time at 2 seconds instead of 59%, on hours that played no part in choosing the fix (docs/model.md).
Jev understands the news questions: on test headlines it called a surprise rate cut bullish, an exchange halting withdrawals bearish, a bakery story irrelevant, and a reworded repeat "not new". Whether that makes money needs real data collected over time.

Getting started

You need Node.js 24.

npm install
cp .env.example .env && chmod 600 .env   # then fill in your keys
npm run check                            # type check and tests; needs no keys

Keys go in .env, which is private and never committed. Every npm script loads it for you.

Key	Needed for
`AI_GATEWAY_API_KEY`	reaching Jev through Vercel AI Gateway
`ALPACA_API_KEY_ID`, `ALPACA_API_SECRET_KEY`	US stock prices and the Benzinga news stream (free Alpaca account)
`X_BEARER_TOKEN`	reading posts from X's official API (paid per post)
`NEWS_USER_AGENT`	your name and email, which the SEC requires for its filing feed and company list

To run it around the clock on a Raspberry Pi, see deploy/pi/README.md: one script sets everything up as a background service.

Any source whose key is missing is simply skipped. To try everything without spending anything, use the practice model, which gives random answers:

JEV_PROVIDER=mock NEWS_MANUAL=1 NEWS_SOURCES=none npm run news   # type headlines like "$BTC ETF approved"

(On a weekend, add NEWS_ONLY_TRADABLE=0 to try stock headlines such as $AAPL beats earnings; normally stocks aren't asked about while their market is closed.)

Commands

npm run news                              # the news part, live
npm run live                              # the market-data part, live
npm run record                            # save Coinbase market data to disk
npm run backtest -- data/raw/<file>       # replay saved data and ask Jev about it
npm run analyze:news -- data/decisions/news-<file>.jsonl
npm run analyze -- data/decisions/<file>.jsonl
npm run dashboard                         # watch it all live at http://localhost:4000
npm run bench                             # measure Jev's response time
npm run check                             # type check and tests

Costs and limits to know about

Jev costs $0.042 per million input tokens, about three thousandths of a cent per decision. The news part costs cents a day; the market-data part about $3 a day. Every status line and report shows what a run actually cost.
Vercel AI Gateway without credits: about 5 Jev calls every 5 minutes. That's enough for the quieter news sources, but not for the busy Benzinga stream during market hours or the market-data part. With credits on the account we saw no limit.
Alpaca free plan: stock prices from one exchange, for up to 30 stocks at a time, and they can be unreliable outside trading hours. The pipeline handles both.
X: each post read costs $0.005; searches that find nothing are free. A daily cap (default 500 posts, at most $2.50) keeps costs down; with the default accounts it's usually cents a day.

Learn more

The docs explain how every part works and why it's built the way it is, in plain language. Good places to start: