Agentic Financial Reasoning, Hedging & Audit
IEEE Big Data Cup 2026 · Phoenix, Arizona · 14–17 December 2026
Official website · Submit challenge paper · Submit Task 1 predictions · Submit Task 3 predictions · Terms · Privacy · IEEE Cup overview
FinReason Cup is Challenge 03 of the IEEE Big Data Cup 2026. It asks a central question for financial AI: can a system's reasoning be executed, checked, and reproduced—not merely presented as a plausible answer?
The competition brings together three complementary tasks spanning symbolic reasoning, sequential financial decisions, and structured verification.
Solve multi-step financial problems and return both a final answer and values for the published intermediate-step slots. The frozen scorer reports Final answer and Reasoning steps separately. The released Task 1 package freezes the prediction schema, evaluator contract, numerical tolerance policy, and participant data. The hub publishes the validator and current submission guide.
Select an asset pair and manage a zero-net-dollar position over time using point-in-time prices, news, and corporate filings. The task is designed to reward relative-value reasoning rather than unhedged directional exposure.
Training data is available from HERCULEAN.
Download prices.parquet, news.parquet, and filings.parquet from the
Task 2 training data hub.
Development and private evaluation splits, exact market windows, eligible assets,
execution assumptions, transaction costs, and position-validity rules will be
published with the scorer.
Perform targeted numeric-fact verification on organizer-packaged SEC EDGAR XBRL filing materials by comparing reported values with values derived from their calculation context. This task is not a full financial-statement audit.
The starter kit, validator, and scorer are released in finreason_task3/, and
the practice phase is open: upload predictions over the 332 public FinMR cases
and the scoring workspace returns accuracy and the three error rates
immediately. Practice answers are public, so practice results are never ranked.
The development and test phases combine separately constructed held-out cases
subject to leakage review and open when those datasets are published.
| Task | Current evaluation |
|---|---|
| Task 1 · Reason | Final answer and Reasoning steps (live) |
| Task 2 · Hedge | Cumulative return, Sharpe ratio, and maximum drawdown (provisional) |
| Task 3 · Verify | Accuracy and structural, extraction, and calculation error rates (live) |
Task 1 and Task 3 each use a frozen public schema, validator, and scorer linked from their participant hubs. Task 2 formulas, tolerances, tie-break procedures, submission contracts, and validity rules will be published with its public scorer.
Schedule update, 7 September 2026: the previously published combined 15 November cutoff has been advanced. Final system submissions and required solution materials for all three tasks are due on 15 October 2026, 23:59 Anywhere on Earth. Working Notes are due separately on 23 October 2026, 23:59 Anywhere on Earth.
Teams seeking final ranking and awards must submit Working Notes (a challenge paper) through the official FinReason Cup SC03 track in CyberChair.
- Length: up to 10 pages total, including references
- Format: IEEE two-column conference template
- Review: single-blind, following the conference policy; include author names and affiliations in the PDF.
- Task 1 Working Notes must cite FinChain, ACL 2026, using the complete official BibTeX displayed in the homepage citation block, with a copy button and downloadable .bib file. The displayed and copied text comes from the same
public/references/finchain.bibfile. - All Working Notes must cite the FinReason Cup overview and the overview for every task entered. Multi-task papers cite the Cup overview once plus all applicable task overviews; Task 1 also requires FinChain. The homepage citation guidance provides four provisional organizer BibTeX entries with copy buttons, individual downloads, and a combined .bib download. Their author fields are
{{Pending}}until names and order are finalized. Refresh the entries before final paper submission; citation keys remain stable. - Working Notes deadline: 23 October 2026, 23:59 Anywhere on Earth
- Paper review feedback and organizer acceptance decisions: 6 November 2026
- Camera-ready deadline for accepted papers: 13 November 2026, 23:59 Anywhere on Earth
CyberChair currently displays its deadline as to be announced and allows up to 10 pages including references. Follow the FinReason organizer schedule above while the portal display is being updated.
The paper portal is separate from the competition submission path. Task 1 does not require pre-registration, organizer approval, an access code, or an account. Teams upload predictions through the verified development and test pages linked from the participant hub. Development immediately returns two scores, a receipt, and a current rank, and the team's best eligible result appears on the public leaderboard. Test shows format feedback and an acceptance receipt. Task 3 follows the same shape, and all three of its phases are open. Task 2 solution materials and submission routes will be published after organizer testing. Competition submissions close on 15 October; the later Working Notes deadline does not extend the competition cutoff.
The organizers review the Working Notes and decide paper acceptance. Selected accepted papers may be included in the conference proceedings, subject to conference publication, camera-ready, registration, and presentation requirements. Submission, ranking, or a certificate does not guarantee paper acceptance or proceedings publication.
Updated 16 September 2026: Test submissions are unlimited before 15 October 2026, 23:59 AoE. There is no daily or total Test submission-count limit. Keep the same Team Name and Contact Email. The latest accepted, valid submission before the deadline supersedes earlier submissions for final evaluation. Invalid or rejected uploads leave the last accepted submission unchanged. An identical previously accepted file returns its original receipt and acceptance time; it does not become a new latest revision. Earlier records remain private for audit purposes. Test scores and ranks remain hidden until final results. This supersedes previous Test submission-count limits; Development limits are unchanged.
FinReason does not offer cash prizes. Registration support is not confirmed at this time, and the organizers do not promise registration funding. A team will receive a participation certificate if it completes both:
- at least one valid final solution submission under the applicable task rules by 15 October 2026, 23:59 Anywhere on Earth; and
- a Working Notes submission through CyberChair SC03 by 23 October 2026, 23:59 Anywhere on Earth.
Winning teams will receive a winner certificate. Additional award categories remain provisional until published. Certificates do not imply challenge-paper acceptance or publication.
A quota applies per phase, and a submission is counted only once it reaches scoring — a file rejected during validation costs nothing, so you can fix a format problem without spending an attempt.
| Task 3 phase | Limit | Counted per | Resets |
|---|---|---|---|
| Practice | 20 per hour | Uploader | Hourly |
| Development | 3 per day | Team | 00:00 UTC |
| Test | 3 in total | Team | Never |
Practice is counted per uploader rather than per team because it asks for no Contact Email and so has no team to count against. The test quota does not reset: three accepted test submissions is all a team gets for the competition.
Task 1 and all three Task 3 phases are open for direct participation. Choose one consistent Team Name and follow the step-by-step submission guide for Task 1 or Task 3. In Task 1, development requires Team Name and a canonical ZIP. Test requires the same Team Name, a private Contact Email, and a canonical ZIP. Contact Email is not a login and is used only for submission identification, submission-related support, matching final results to the related challenge paper, and selecting the latest accepted test submission and handling repeat uploads through a non-public pseudonymous identifier.
| Resource | Status |
|---|---|
| Official challenge website | Live |
| Task 1 training, development, and test files | Live as 13 development files and 3 test files |
| Task 1 direct web upload | See the participant hub for current verified availability |
| Task 1 step-by-step submission guide | Live |
| Task 1 validator, sample B0, and B1 baseline | Live in this repository |
| Task 1 organizer baseline scores | Lite Baseline, Standard Baseline, and Advanced Baseline ranked on the 580-question development set |
| Task 2 training data | Live: prices, news, and filings (Parquet) |
| Task 3 starter kit, validator, and scorer | Live in this repository |
| Task 3 practice phase (332 public FinMR cases) | Live — scored on upload, never ranked |
| Task 3 development and test questions | Live as 680 cases each, answers withheld |
| Task 3 development phase | Live — scored, ranked, public leaderboard |
| Task 3 test phase | Live — acceptance receipt only |
| Task 2 starter kit and baselines | Coming soon |
| Participant support | zhuohan.xie@mbzuai.ac.ae |
| Terms of Participation | Live |
| Privacy Notice | Live |
| Challenge paper submission | Open |
| Task 1 participant hub | Public downloads live; direct-upload status published here |
| Task 1 leaderboard | Development only |
| Task 3 participant hub | Live; per-phase status published here |
| Task 3 step-by-step submission guide | Live |
| Task 3 leaderboard | Development only |
The participant hub publishes only organizer-verified competition-platform and submission links and records their current availability.
| Milestone | Date |
|---|---|
| Competition closes for Tasks 1, 2, and 3: final system submissions and required solution materials | 15 October 2026, 23:59 AoE |
| Final competition results | 16 October 2026, after 12:00 UTC (16:00 Abu Dhabi time) |
| Team Working Notes through CyberChair SC03, up to 10 pages including references | 23 October 2026, 23:59 AoE |
| Paper review feedback and organizer acceptance decisions | 6 November 2026 |
| Camera-ready papers | 13 November 2026, 23:59 AoE |
| Organizer overview due to the conference | 20 November 2026 |
| Winning teams announced (awards) | 25 November 2026 |
| IEEE Big Data 2026, Phoenix, Arizona | 14–17 December 2026 |
AoE means Anywhere on Earth (UTC−12). The 15 October competition cutoff passes at 12:00 UTC on 16 October, so final results will not be released before that time. Task 2 release dates will be published after organizer testing; the final competition cutoff remains 15 October for every task.
This repository contains the organizer-maintained FinReason Cup website and the released Task 1 and Task 3 participant data and tools. Task 2 participant resources will be linked here after organizer verification.
Please use only links marked as verified on the official challenge website.
The public /task1/ route is the stable participant entry point:
- exactly 13 development files and three public test files are downloadable from Pages with their frozen manifests;
/task1/submit/publishes two distinct organizer-verified direct-upload links after both workspaces pass deployment checks;- the development page accepts Team Name and the 580-row predictions ZIP, then immediately returns Final answer, Reasoning steps, a receipt, and current rank;
- each team's best eligible development result is ranked alongside Lite Baseline, Standard Baseline, and Advanced Baseline on the same 580 development questions; Financial Rules remains a reference in the expandable score guide; baseline rows are marked as references, while participant-only team ranks continue to match submission receipts;
- the separate test workspace accepts the 928-row test predictions ZIP and requires Team Name and Contact Email, then shows public-format validation feedback and an acceptance receipt when accepted, with no score, rank, answer-correctness feedback, or score-derived signal;
- the leaderboard hub has Dev and Test submission status tabs;
- Contact Email, submission files, gold answers, and private evaluation records stay inside the verified private submission boundary and are never published.
The step-by-step submission guide
identifies the exact development and test input files, the five required JSONL
fields, the validator and packaging commands, and each phase's receipt behavior.
Every uploaded ZIP must contain exactly one root-level file named
predictions.jsonl; the paper PDF is submitted separately through CyberChair.
The public JSON feed uses a frozen aggregate-only development leaderboard contract. It publishes only Team Name, rank, the two aggregate scores, and acceptance time. It must not expose predictions, attachments, Contact Email, receipts, private identifiers, test results, or additional metadata. The response must be no larger than 1 MiB.
The Pages workflow builds the website, frozen public downloads, guarded links,
and the aggregate development leaderboard view. It never receives a Contact
Email, submission archive, gold answer, or private evaluation record. Live mode
requires two distinct verified root *.hf.space URLs and the verified public
development leaderboard endpoint. Until then, the public site shows the upload
links as pending rather than publishing an unverified route.
The public /task3/ route follows the same boundary with three phases instead
of two:
- practice runs on the 332 public FinMR cases whose answers ship with the
starter kit. Upload
predictions.jsonl— or a ZIP containing exactly one root-level file of that name — with a Team Name only, and the scoring workspace returns accuracy, the structural, extraction, and calculation error rates, and a receipt. Because the answers are public, practice results are never ranked and never reach the leaderboard: a practice ranking would measure who read them; - development runs on organizer-held gold and is the phase that ranks. It additionally requires a Contact Email, which is stored only as a salted hash;
- test runs on a separate receipt-only workspace that returns an acceptance receipt and nothing else — no score, rank, diagnostic, or score-derived signal. That workspace serves no scoring or leaderboard route at all, and refuses to start if it is configured in a way that could return one.
Each phase reports its own status on the hub, and a phase whose dataset does not
exist reports itself as pending rather than accepting work it cannot score. The
step-by-step submission guide
covers preparing the public cases, the three required JSONL fields, and the
validator and local scorer in finreason_task3/.
The public JSON feed uses the frozen aggregate-only development leaderboard
contract finreason.task3.development-leaderboard/1.0.0. It publishes only Team
Name, rank, the four aggregate rates as two-decimal percentages, and acceptance
time. Rows are ordered by accuracy descending, then by calculation, extraction,
and structural error rates ascending, then by Team Name.
Requirements: Node.js 22.13 or newer.
npm ci
npm run devOpen http://localhost:3000.
Run the same checks used by the GitHub Pages workflow:
npm run lint
npm testnpm test builds the static GitHub Pages export and validates its rendered HTML
contract. Pushes to main deploy through
.github/workflows/deploy-pages.yml.
Participant support, privacy questions, and correction or deletion requests can be sent to zhuohan.xie@mbzuai.ac.ae. See the Privacy Notice for the public website boundary, public leaderboard fields, external services, and the Task 1 retention policy. Readable Contact Email is kept only while needed for the stated purposes, then deleted or minimized. Encrypted submission archives and non-public operational event records may remain in restricted organizer-private repository history for challenge administration and audit; the Privacy Notice explains this boundary and its exceptions.
The organizer team is led by The Fin AI, with contributors affiliated with MBZUAI, McGill University, Stevens Institute of Technology, Yale University, and the University of Manchester. Affiliations do not imply institutional sponsorship.
Task-specific dates, platform settings, award categories, and resource licenses are published only after organizer verification. The Terms of Participation describe the current organizer-maintained participation rules.
Last reviewed: 7 September 2026.