Agenthon
NeurIPS 2026 Competition Track
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.
Final deadline
Final uploads for all four tracks close on Tuesday 27 October at 23:59 AoE (28 October, 11:59 UTC). On Track 2, uploads made by 26 October, 23:59 UTC count as usual; for later uploads, see the Track 2 notice.
Awards
The first, second and third teams in each of the four tracks win cash awards: $12,000 in total.
We're delighted to welcome our sponsors NVIDIA (also our technology provider), Bloomberg and AllianceBernstein, and Nebius, our cloud infrastructure partner. Thank you for your support, for helping us organize the competition, and for collaborating with us on its science and task development. Meet our supporters
Workshop program
Saturday 12 December 2026, 8:00–10:15 AM Atlanta time (EST, UTC−5), in Atlanta, Georgia.
Welcome from the organizers.
Noyan Tokgozoglu, Global Head of AI @ Optiver · LinkedIn
Jake Goldberg, Capital Markets @ NVIDIA · LinkedIn
Speaker to be announced.
Gary Kazantsev, Head of Quant Technology Strategy, Office of the CTO at Bloomberg LP · LinkedIn
Key dates
Registration runs from 17 August to 12 October 2026. Registration and Development close together on 12 October at 23:59 AoE (13 October, 11:59 UTC), the one deadline for Development uploads on every track. Every valid Development upload made before then is run, even if its run starts after the close.
Sign up and form your team.
Public practice repositories, iteration, and the validation leaderboard.
Papers for the NeurIPS workshop were due at 23:59 AoE.
One submission per entered track is evaluated on sealed held-out units; on Track 2 a team may make up to two Final uploads, at most one per UTC day, and the later one is its entry. On Track 2, every upload made by 26 October, 23:59 UTC counts as usual; for later uploads, see the Track 2 notice. Uploads close on 27 October at 23:59 AoE; organizers verify the top entries after their runs. Results and awards are published in early November.
The Agenthon workshop in Atlanta, Georgia, with posters from the winning teams and the accepted papers.
The competition closes with the Agenthon workshop at NeurIPS in Atlanta, Georgia, on Saturday 12 December, where the winning teams present posters alongside the accepted papers. All dates are given in the official rules, which govern if anything here is out of step.
Teams
You can compete alone or with up to two teammates. You join one team and stay on it, and that one team can enter as many of the four tracks as it wants.
Registration is per team, and only registered teams can submit solutions or appear on the leaderboard. Your team uses one account on the competition platform, chosen by the team — the FAQ explains what that means for everyone else on the team.
Have a question about entering? Start with the FAQ.
Core question
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.
Agents in all four tracks run in sandboxed Docker containers without general internet access. Supported Coding, Forecasting and Explainability submissions can call House Nemotron through a restricted connection. Simulation has no network access. Build required dependencies and permitted artifacts into your image before evaluation.
Four tracks
Build a Docker agent that solves quantitative finance coding tasks under pytest and financial-invariant checks.
Forecast future panels using time-series data plus a time-stamped text corpus. The score evaluates the forecast distribution; information uplift is a research aim.
Submit an ABIDES-compatible simulator that is faster while preserving matching-engine semantics and market stylized facts. Development results are practice feedback, not controlled Final timing.
Predict labels, values, or rankings over tabular entities, supported by a frozen evidence corpus. Follow the track guide for prediction, evidence and reasoning evaluation.
Protocol
Each unit is checked for admissibility and evaluated using its track's scoring rules. Participant failures remain in the evaluation denominator; organizer faults are handled separately. If two Final submissions finish a track with the same ranking score, the tie is broken in favour of the one uploaded earlier.
Upload the toolkit-generated ZIP identifying your agent image and proving team registration.
g0-g3 verify integrity, schema, cutoff/resource rules, and domain semantics.
Track-specific scoring produces Development feedback. Track 1 uses pass@1 without a confidence interval.
Public/private firewall
Each track has a public practice repo and a private sealed exam repo. Public repos include practice tasks, examples, local checks, available baselines and participant docs. Private repos hold evaluation tasks, reference answers, scoring material and audit records.
Leaderboard
Development leaderboards on agenthon.net are public and visible to everyone. Final results for all four tracks are published here after verification. Each track uses its own scale. Higher is better for the displayed leaderboard scores; the raw Forecasting loss is converted for this display. Use CodaBench for uploads and processing status. Follow the logged-in website for submission opening status. From 5 October 2026, 00:00 AoE (12:00 UTC), the Track 1 board ranks only runs uploaded after 25 September 2026 at 05:00 UTC, for every team (Track 1 README, rules 8 and 9).
Final results, published after verification.
Uploaded: when each team made its Final upload, in UTC. Equal scores are ranked by upload time, the earlier upload first, as the published tie-break says. An entry the organizers ran again keeps its original upload time.
T1 Coding: teams 1–25 of 61
T1 Coding: teams 26–50 of 61
T1 Coding: teams 51–61 of 61
No team matches that search.
Final results, published after verification.
T2 Forecasting: teams 1–25 of 87
T2 Forecasting: teams 26–50 of 87
T2 Forecasting: teams 51–75 of 87
T2 Forecasting: teams 76–87 of 87
No team matches that search.
Final results, published after verification.
T3 Simulation: teams 1–25 of 55
T3 Simulation: teams 26–50 of 55
T3 Simulation: teams 51–55 of 55
No team matches that search.
Final results, published after verification.
T4 Explainability: teams 1–25 of 63
T4 Explainability: teams 26–50 of 63
T4 Explainability: teams 51–63 of 63
No team matches that search.
2026 phases
Development runs through 12 October 2026. The joint Final + Verification phase runs from 13 to 27 October 2026. Registration and Development close on 12 October at 23:59 Anywhere on Earth (AoE, UTC−12). Final uploads close on 27 October at 23:59 AoE (28 October, 11:59 UTC), extended from 25 October because CodaBench was unreachable. An upload made by then runs and counts even if its run finishes after the close; organizers verify the top entries after their runs, so verification can end after 27 October. Keep your image pullable until the results are published. On Track 2, every upload made by 26 October, 23:59 UTC counts as usual; for later uploads, see the Track 2 notice.
Upcoming
Build and test with the public starter packages. Check the logged-in website for submission opening status.
Teams practice and iterate with the public starter packages. Submission opening is announced separately.
One submission per entered track is evaluated on sealed private-test units; on Track 2 a team may make up to two Final uploads, at most one per UTC day, and the later one is its entry. Uploads close on 27 October at 23:59 AoE; organizers verify the top entries after their runs. There is no separate verification submission. Run outputs, logs, and score files stay hidden. Results and awards are published in early November.
Awards
The top three teams in each track win a cash award, decided by the final standings after verification. That is $3,000 per track and $12,000 in total.
In each of the four tracks.
In each of the four tracks.
In each of the four tracks.
Amounts are in US dollars. A team's award is split equally among its registered members. Eligibility, verification, tax and payment requirements are set by the official rules, and a confirmed winner who accepts an award publishes the code needed to reproduce the method.
Scientific outputs
Gate failures are labeled and aggregated to show where finance agents break down.
T2 investigates whether text and reasoning improve forecasts; this is not a separate scored component.
T3 measures throughput only after semantic fidelity and stylized facts survive checks.
T4 requires evidence-backed predictions that do not cite future or unsupported facts.
Call for papers
Accepted papers will be presented as posters at the Agenthon workshop at NeurIPS in Atlanta on Saturday 12 December. See the workshop program.
Our Supporters
A number of sponsorship tiers and opportunities are available, including monetary and awards sponsorship, event space, data, infrastructure, and compute.
Agenthon Questions and Tasks are of real relevance to hedge funds and asset managers, spanning alpha forecasting, portfolio optimization, computational statistics, machine learning, and AI. Questions are provided jointly by the SQA, CEWIT and our Question Partners.
See Questions 2025 for last year's questions and a feel for Agenthon priorities.
Organizing Committee
Assistant Professor, Department of Applied Mathematics and Statistics, Stony Brook University
Vice President, Society of Quantitative Analysts
Chief Investment Officer, Atlas Ridge Capital
Adjunct Professor, NYU Courant
Executive Advisory Board, Columbia Business School, Program for Financial Studies
Head of Machine Learning Strategy, CTO Office
Bloomberg, Toronto, Canada
Head of Quant Technology Strategy, Office of the CTO
Bloomberg, New York, USA
Global Head of Capital Markets Strategy
NVIDIA Corporation, USA
T1 · Coding
Zhikang Dong
Track Lead T1
Independent Researcher
T2 · Forecasting
Ruolan Sun
Track Lead T2
Ph.D. Student, Stony Brook University
T3 · Simulation
Haohan Xu
Track Lead T3
Ph.D. Student, Stony Brook University
T4 · Explainability
Mathew Thiel
Track Lead T4
Quant Research Analyst, validityBase
