Quant-finance coding agents
Build a Docker agent that solves quantitative finance coding tasks under pytest and financial-invariant checks.
- Verb
- solve
- Metric
- pass@1 / pass@3
- Gate
- pytest + invariants
NeurIPS 2026 Competition Track
Agenthon 2026
Verifiable AI for quantitative finance.
A four-track competition testing whether AI agents can produce finance outputs that survive automated, leakage-controlled, cheat-resistant evaluation.
Submission
Docker agent
Admissibility
g0-g3 gates
Ranking
Track metrics
Core question
Can AI agents produce finance answers that can be checked by machine?
Agenthon extends the Alphathon program into its first NeurIPS edition. The competition keeps the four-track structure and hard finance setting, while adding sealed held-out data, automated leakage controls, reproducible reruns, and public leaderboards.
Every official submission runs offline in a sandboxed Docker container. It must pass integrity, schema, cutoff/resource, and domain-semantics gates before any leaderboard metric is computed.
Four tracks
Same competition spine, four finance problems.
Reasoning-augmented time series
Forecast future panels using time-series data plus a time-stamped text corpus, then prove information uplift over text-blind baselines.
- Verb
- forecast
- Metric
- CRPS composite
- Gate
- as-of cutoff + calibration
Accelerated market simulation
Submit an ABIDES-compatible simulator that is faster while preserving matching-engine semantics and market stylized facts.
- Verb
- simulate
- Metric
- events/sec
- Gate
- semantic regression
Evidence-grounded prediction
Predict labels, values, or rankings over tabular entities with citations from a frozen evidence corpus and confidence intervals.
- Verb
- analyze
- Metric
- quality + coverage
- Gate
- faithfulness + embargo
Protocol
Submit, check, score.
The visible rule is simple. A submission is ranked only after it passes the same admissibility sequence used by every track.
01
Submit
A Docker image implements one stable CLI verb for its track.
02
Check
g0-g3 verify integrity, schema, cutoff/resource rules, and domain semantics.
03
Score
Only admissible runs receive a track metric and bootstrap confidence interval.
Public/private firewall
Transparent grading, sealed answers.
Each track has a public practice repo and a private sealed exam repo. Public repos include practice tasks, baselines, smoke scorers, and participant docs. Private repos hold held-out tasks, oracle keys, final scoring logic, and audit material.
Public practice
Private exam
Public-dev and validation units
Private-test held-out units
Runnable baselines and smoke scorer
Oracle solutions and final scorer
Manifest and canary safety checks
Canary registry and audit logs
Leaderboard
Competition scores will appear here.
Each team row will show its total score first. Expanding a row reveals the per-track score breakdown used to compute or explain that total.
Rank
Team
Total
Breakdown
-- Team placeholder 01 --
T1 Coding--
T2 Forecasting--
T3 Simulation--
T4 Tabular--
-- Team placeholder 02 --
T1 Coding--
T2 Forecasting--
T3 Simulation--
T4 Tabular--
-- Team placeholder 03 --
T1 Coding--
T2 Forecasting--
T3 Simulation--
T4 Tabular--
2026 phases
Development, final, verification.
Upcoming
Development opens July 20
Public repos open, teams practice, and the validation leaderboard goes live.
0%
complete
-
Upcoming Jul 20 - Sep 28 Development
Public repos open. Teams practice, iterate, and use the live validation leaderboard.
-
Upcoming Sep 29 - Oct 12 Final
One submission per team is evaluated on sealed private-test units with a hidden leaderboard.
-
Upcoming Oct 13 - Oct 25 Verification
Organizers rerun top submissions with fresh seeds and review reproducibility.
Scientific outputs
Agenthon is designed to explain failures, not just rank winners.
Cross-track failure map
Gate failures are labeled and aggregated to show where finance agents break down.
Information uplift
T2 isolates whether text and reasoning beat strong text-blind forecasting baselines.
Speed-realism frontier
T3 measures throughput only after semantic fidelity and stylized facts survive checks.
Faithfulness under embargo
T4 requires evidence-backed predictions that do not cite future or unsupported facts.
Resources
