PI Trading Arena
Which AI makes better trading decisions? Give each model the same market picture, apply the same rules, and measure what happens to the account.
Round 02 / Published 12 Oct 2026
28 Aug–9 Oct 2026
Final mark: 12 Oct 2026
Final dollars from $1,000 · Full context · Act at ≥50% choice probability
| Rank / model | Final account | Return | Max. drawdown | API cost |
|---|---|---|---|---|
| 01Microsoft-Decision-1microsoft/microsoft-decision-1-20261009 | $1,063.56 | +6.36% | -13.13% | $0.0343 |
| 02Jev 1.13typesafe/jev-1.13-20260917 | $1,031.32 | +3.13% | -12.27% | $0.0401 |
momentum #1 (no model)
$715.77 -28.42%
buy & hold QQQ
$1,038.06 +3.81%
cash
$1,000.00 0.00%
Account value · after trading fees
Starts at $1,000 before entry fees. Recorded marks include the final liquidation; lines connect those marks. API costs are separate.
Inside the replay
Available candidates: SOL/USD · BTC/USD · SPCX · XLE · NVDA · GLD · IWM · QQQ
Microsoft-Decision-1
Cash
Pick: SOL/USD · choice probability 46.2%
Regime score: 1.94 / 2 · cash below 0.50
Jev 1.13
SOL/USD
Pick: SOL/USD · choice probability 67.0%
Regime score: 1.94 / 2 · cash below 0.50
Allocation applies the published 50% threshold and regime filter to the recorded responses. Exact states, questions, and responses are in the replay download.
Protocol / v2026.10.12
The account evolves mechanically. Each model sees an empty portfolio at every decision, so its input does not depend on earlier picks.
At 10:15, 13:30, and 15:45 New York time, build the same state for every model: up to eight momentum candidates, technicals, news, macro indicators, FOMC timing, crypto sentiment, and funding.
Send byte-identical states and structured questions through OpenRouter’s Decisions API. Check the exact returned model version and cache every response. This runner requires compatible decision models.
Hold 100% in the chosen candidate when its choice probability is at least 50% and its regime score is at least 0.50 out of 2. Otherwise hold cash. No leverage or short positions.
Fill at the close of the 15-minute bar starting at the decision time. Charge 0.05% per side for stocks/ETFs and 0.25% for crypto. Liquidate at the final mark and compare against the same-window baselines.
Base: regime, five momentum candidates, asset news. Context: adds macro ETFs, market headlines, and momentum shifts. Wide: expands to eight candidates. Full: adds FRED macro series, FOMC timing, crypto sentiment/funding, and wider news. The replay package includes all four tracks and the alternative pick-without-threshold policy for research.
Read before comparing
The full-context, 50% configuration was fixed on 12 October, after comparing tracks and policies on this round. These historical results are exploratory; a future round can test the frozen protocol prospectively.
Microsoft’s pinned version was released on 9 October; Jev’s on 17 September. Some or all replay dates predate those releases. Training-data overlap cannot be excluded.
Thirty trading days do not establish a durable ranking. The simulation omits spread, slippage, market impact, taxes, and cash interest. Drawdown is measured at recorded marks.
Inputs are filtered by historical timestamps. That alone cannot rule out revised macro observations or historical news retrieval effects. This release does not claim an independent leakage audit.
With future prices known, the optimal path over the same candidates ends at $3,946.45. This hindsight ceiling is a reference bound, not an achievable trading strategy.
Inspect / Reproduce / Extend
Download the frozen tasks, model responses, price inputs, and scoring code. Recompute the cached results without API keys or paid calls.
Download Round 2 replayAfter extracting the replay
python -m pip install -r requirements.txt
python -m arena.simulate round2 fullAdding a model requires a pinned version compatible with the Decisions API. See the guide for setup and costs. Third-party data remains subject to its providers’ terms.
Created by Joel Stafford at PI Labs. An independent pilot benchmark; no affiliation with the evaluated model providers. Explore the Nero X and Nero DX trading bots behind the research.
Simulated research results. Not investment advice.