# PI Trading Arena — Round 2 replay release

Published by PI Labs / Joel Stafford, 12 October 2026.
Canonical page: https://pilabs.space/trading-arena

## What this measures

Starting with a simulated $1,000, compare pinned decision models on identical historical
market states. Round 2 contains 90 decisions across 30 trading days, 28 August–9 October
2026, with final liquidation marked on 12 October. This is a retrospective pilot, not a
live trading competition or an independently certified industry standard.

Official configuration: `full` track, `pick>=0.5` policy. Hold the chosen candidate only
when its choice probability is at least 0.5 and its regime score is at least 0.5;
otherwise hold cash. Allocate 100% to one asset or cash, without leverage or shorts.
Fees per side: ETFs/stocks 0.05%, crypto 0.25%. Fills use the close of the 15-minute bar
starting at the decision time (15 minutes later). Account values include trading fees;
model API costs are reported separately and are not deducted from equity. The simulator
does not model spread, slippage, market impact, taxes, or cash interest.

## Reproduce the cached results

Extract the ZIP and run from its root with Python 3.10+:

```sh
python -m pip install -r requirements.txt
python -m arena.simulate round2 full
python -m arena.analyze round2
```

No API keys or model calls are needed to score the included frozen responses. The
simulation regenerates summaries and CSVs. The analysis command compares the research
`pick` policy; it does not regenerate the official-threshold bootstrap statistics in
`official_details.json`. That file is a supplied analysis artifact.

`rounds/round2/full/decision_points.jsonl` contains the exact tasks: state and questions.
`answers_*.jsonl` contains the recorded model responses, usage, version and latency.
`data/bars_15m.csv` supplies realized prices for scoring. The hindsight ceiling is an
optimal path over the same choices with future prices known; it is not a feasible
trading strategy or an answer shown to the models. All four context tracks are included.
SHA256SUMS.json identifies the included source inputs byte for byte.

## Evaluate a new model

The runner uses OpenRouter's alpha Decisions endpoint, not a generic chat completion.
Create a local `.env` with `OPENROUTER_API_KEY=your_key`, add the exact pinned model ID
to `MODELS` in `arena/run_models.py`, and its release date to `RELEASED` in
`arena/simulate.py`. The model must support the runner's choice, score and noul question
types. Use a new model name so existing cached answers remain intact.

```sh
python -m arena.run_models round2 full
python -m arena.simulate round2 full
```

New API calls incur provider charges. The alpha API and historical pinned versions may
change or become unavailable. Regenerating tasks with `arena.data_build` additionally
requires Alpaca credentials and fresh source downloads; those may not reproduce the
frozen historical inputs. Bot database snapshots and private credentials are excluded.

## Interpretation and limitations

- Microsoft-Decision-1: $1,063.56 (+6.36%); Jev 1.13: $1,031.32 (+3.13%).
- Buy & hold QQQ: $1,038.06 (+3.81%); cash: $1,000; momentum #1: $715.77.
- The full/50% configuration was selected after exploring tracks and policies on this
  round. It was fixed on 12 October 2026, after the evaluation window; it was not
  preregistered for this historical round.
- Microsoft was released 9 October 2026 and Jev on 17 September. Historical market
  information may have appeared in training. This is not a clean prospective test.
- The supplied official analysis reports Microsoft ahead in 71.9% of paired day-block
  resamples; the 90% interval for its dollar advantage is -$52 to +$115.20. That does
  not establish a reliable winner or a probability of future outperformance.
- Timestamp filtering does not prove every historical source was available in its
  present form at the time. Revised macro data and historical news retrieval remain
  limitations. No independent leakage audit is claimed here.
- Thirty days and two models provide a limited view of one market regime. Maximum
  drawdown is measured at recorded marks, not continuously within each holding period.
- This isolates allocation choices with an empty portfolio in every prompt. It does
  not evaluate the full Nero DX strategy, execution quality, or portfolio management.

These are simulated research results, not investment advice. Third-party data remains
subject to its original providers' terms. This release makes no public-domain claim.