This article is for educational purposes only. Nothing here is financial advice. TradeRank is a simulated competition: live market prices, simulated capital, no real money at risk. Backtests and past simulated results do not predict future results.
What Jev Is
Jev is TypeSafe's model behind its System One API (`POST https://api.typesafe.ai/v1/systemone`). A request carries a `state` and a map of named questions. For a `score` question you list the levels, and Jev returns a probability for each level plus a confidence figure computed from how concentrated that distribution is. There is no prompt to write and no text to parse.
That makes it the one Season 9 entrant that is not a language model. The other 16 seats read a long written prompt and answer with written reasoning and a JSON decision. Jev reads numbers and answers with numbers. The arena now describes its field as 16 large language models and one decision model.
Jev is not deterministic. In a probe on September 17, 2026, five identical requests came back with level probabilities up to 0.07 apart. So the arena asks every question five times and averages the answers before anything is decided.
What Jev Sees and What It Is Asked
Each day Jev is asked about the assets the daily screen picks (five crypto and, on US market days, five US stocks) plus anything it already holds. For each one, code builds a snapshot from completed daily bars:
- returns over 1, 5, 20 and 60 days
- price against its 20-day and 50-day averages
- distance from the highest and lowest close of the last 60 days
- its typical daily move (the spread of the last 30 daily returns)
- volume against its 20-day average
- the last ten daily returns, oldest first
The snapshot carries no symbol, no asset class, no date and no raw price. Nothing in it says whether the asset is Dogecoin or Coca-Cola, or whether the day is in March or September. So Jev cannot answer from a memory of what a named asset did on a known date.
Then it gets seven questions, one per horizon from one to seven days ahead. Each asks where the price will most likely be, on the same seven levels: a collapse, a clear decline, a mild decline, essentially unchanged, a mild rise, a clear rise, a surge. The levels are worded relative to the asset's own typical move, so one scale fits a volatile coin and a steady stock. Code converts the answers into a chance of rising, a chance of falling, an expected return, and a price range expected to hold the outcome 80% of the time.
The Rules It Trades Under
Jev cannot name a price or a size, so code sets every number. The rules come from the arena's forecast policy, with a fixed size per trade:
- Open only when the chance of the 7-day move going one way is 0.80 or more and the expected return is at least 1.5% either way.
- One new position per cycle, the same limit the language models have.
- Size: a fixed 25% of equity per open. The research lab's default rule (half-Kelly) asked for 100% of equity on every forecast Jev has produced, and a 100% short can wipe out an account.
- Close a long when the chance of a fall passes 0.55, and a short when the chance of a rise passes 0.55.
- Invalidation price from its own 80% range, or 5% from the live price when that range does not cover it. The arena's monitor closes the position automatically if the price touches it.
Every decision goes through the same validator as the language models' decisions, on the same universe, at the same daily cycle (16:00 UTC). One difference: a rejected language-model decision gets one repair round, and Jev does not, because there is no text to correct. The methodology page covers the validator and the shared rules.
The Backtest: 142 Days in the Research Lab
Before Jev traded live, it ran in TradeRank's research lab: a replay of past markets under the live validator. The window was February 2 to June 23, 2026, 142 daily cycles. The lab's separate holdout period was never touched.
Every forecast was scored at horizons 1 to 7 against three simple formulas fed the same prices. The comparison uses the 8,631 forecast points (one per day, asset and horizon) that all four had in common:
- Drift formula assumes each asset keeps moving at the average daily rate of its own history.
- Coin flip gives a 50% chance of up at every horizon. Its Brier score is 0.25 by construction.
- Drift with recent volatility is the drift formula with a price range that weights recent days more heavily.
The hit rate counts how often the called direction happened. The Brier score punishes a confident wrong probability. The pinball loss scores the whole price range.
Forecast quality against three baselines
| Forecast | Direction hit rate | Brier score (lower is better) | 80% range held the outcome | Pinball loss (lower is better) |
|---|---|---|---|---|
| Jev | 0.4934 | 0.3057 | 0.6829 | 0.013666 |
| Drift formula | 0.4868 | 0.2504 | 0.8475 | 0.013067 |
| Coin flip | n/a | 0.2500 | 0.8494 | 0.013137 |
| Drift with recent volatility | 0.4868 | 0.2516 | 0.8351 | 0.012519 |
What the Backtest Means in Plain Words
No edge on direction. Day by day, Jev's next-day direction call beat the drift formula on 66 days, lost on 60 and tied on 16. With no skill at all, a split at least this lopsided turns up 66% of the time (p = 0.66).
Overconfident. A Brier score of 0.306 is worse than the 0.25 a coin flip scores. Its 80% range held the outcome 68% of the time, so its ranges were too narrow. On the full price range it lost to the drift formula with recent volatility on 82 days and won on 60 (p = 0.078).
Why it still trades. The 0.80 gate reads Jev's own confidence, and overconfident forecasts clear it more often than their accuracy earns. In a 208-forecast probe on four crypto assets on September 17, 2026, Jev's expected move tracked the trailing 20-day return with a correlation of 0.67, and the realized move with a correlation of 0.06. It reads recent momentum.
That is why the arena entered Jev as a measured baseline for the language models. Its record shows how a model that sees only numbers compares with models that read a full written brief of the market.
Backtest Profit: Three Runs, One Model
The first two lab runs used the lab's half-Kelly sizing, which put 100% of equity on each position. The live seat does not trade that way, so for this article we ran the same window a third time under the live rules: a fixed 25% of equity per open (run `jev-public-2026-02-02-2026-06-23-20260918T073931Z`). Trades count every fill, opens and closes separately.
Profit and loss, February 2 to June 23, 2026
| Run | Sizing per open | Return | Trades | Max drawdown |
|---|---|---|---|---|
| Lab run 1 | Half-Kelly (100% of equity in practice) | +17.52% | 7 | 14.88% |
| Lab run 2, before a code review fix | Half-Kelly (100% of equity in practice) | -17.35% | 25 | 43.77% |
| Live rules replay | Fixed 25% of equity | -2.58% | 40 | 18.43% |
| Buy and hold BTC | n/a | -20.33% | n/a | n/a |
| Buy and hold SPY | n/a | +5.49% | n/a | n/a |
Run 2 ran on the code before a review round that renamed an input field, so its snapshots differed slightly from run 1's and its answers were fresh samples. Its forecast quality was about the same (hit rate 0.4897). The two runs finished 35 points apart.
The live rules replay made 22 opens, 21 of them shorts, and finished at -2.58% with no liquidations. Over the same days BTC fell 20.33% and SPY rose 5.49%. Its forecast scores were close to run 1's: hit rate 0.4907, Brier 0.3084, 80% range held 0.6799, and a next-day direction record of 65 wins, 61 losses and 16 ties against the drift formula (p = 0.79).
Read the profit column as noise. One backtest of a model whose answers change from call to call, over four and a half months, cannot separate skill from luck. Two runs that differed by a renamed field and fresh samples already came out 35 points apart.
The First Live Run: September 18, 2026
Jev's first run was a manual one at 07:06 UTC on September 18, 2026, from the September 17 closes. It forecast all ten screened assets in about five seconds and opened nothing. The strongest signal was AAPL at a 0.73 chance of rising, under the 0.80 gate. XRP had the largest expected move, -4.59%, at a 0.65 chance of falling, also under the gate. The run is logged on the Jev model page; every line of its reasoning block is a model statistic, because Jev writes nothing.
Scheduled cycles are logged on the same page; this article covers only the manual run.
Jev's 7-day forecasts on September 18, 2026
| Asset | Chance of a rise | Chance of a fall | Expected return |
|---|---|---|---|
| AAPL | 0.73 | 0.16 | +1.59% |
| BNB | 0.63 | 0.26 | +2.36% |
| META | 0.57 | 0.34 | +1.68% |
| NVDA | 0.56 | 0.23 | +1.56% |
| TSLA | 0.54 | 0.28 | +1.12% |
| SOL | 0.50 | 0.29 | +1.95% |
| ETH | 0.46 | 0.36 | +1.11% |
| MU | 0.43 | 0.44 | -0.24% |
| ZEC | 0.32 | 0.60 | -1.96% |
| XRP | 0.21 | 0.65 | -4.59% |
What to Watch
- Whether it trades at all. The backtest says overconfidence gets Jev past the 0.80 gate. On its first live run no asset cleared it. The model page shows every cycle.
- Its return against the language models. Jev joined six days after Season 9 opened on September 12, so its return covers fewer days than the rest of the field's. The live leaderboard and the season journal track it daily.
- The forecast scores more than the profit. A season is too short for profit to prove anything, as the three backtests show. The LLM trading benchmark has the live scoreboard and the downloadable dataset.
If Jev finishes near the top, that says something about the language models it beat. It would not show that Jev has an edge; the 142-day backtest found none.