MCM/ICM / AI tutorial

Compete in MCM/ICM with Scientify

A step-by-step tutorial for the four days of MCM/ICM 2027: reading the problem, building baselines, iterating models and algorithms, testing results, and writing the paper with Scientify.

Short answer

Scientify is a research agent that runs on a cloud computer. In MCM/ICM, the team gives it the problem and attachments; it builds a baseline, replaces model structures and solvers round by round, records the metrics of every version, tests the final version on held-out data, and writes an English paper of at most 25 pages in the COMAP template. The team chooses the problem, sets the iteration direction, and decides on the solution. COMAP rules allow AI tools.

About this page

Tutorial
What this page solves
MCM/ICM gives a team 99 hours to choose a problem, model, code, test, and write an English paper. You want AI to push the model to the level of Meritorious and Outstanding papers, but you are not sure what to ask it each day or how to judge whether the result is good enough.
Who it is for
Undergraduate teams entering MCM/ICM 2027, with at least one member who can read Python.
What you will have
You will know what to have Scientify do on each day, which results to check, and how to judge whether the model reaches the level of award-winning papers.
Covers
  • Preparation and team roles
  • Day 1: reading problems and baselines
  • Days 2–3: iterating models and algorithms
  • Day 4: testing and sensitivity analysis
  • Writing and submitting the paper
  • Checks against award-winning papers
Does not cover
  • Registration steps (see MCM/ICM 2027 dates and registration)
  • Award rates (see MCM/ICM awards and award rates)

Main steps

Why this stepSet up the account and workflow in advance so contest time goes to modeling.

  1. 01

    Create a Scientify account

    New accounts get a free cloud computer and US$5 in credit, enough to run one past problem.

  2. 02

    Run a past problem

    Download a past problem from the COMAP site and give it with its attachments to Scientify. Learn the project layout and where code, logs, and results are kept.

  3. 03

    Set team roles

    One member leads problem choice and iteration direction, one checks code and results, and one owns the paper and summary sheet.

Why this stepOn day 1 (from 5:00 p.m. EST on January 28), pick a problem from baseline results.

When the contest opens, give two or three candidate problems and their attachments to Scientify, one project per problem. It checks data fields and missing values, lists usable methods, and writes runnable baselines.

Compare the baselines, data quality, and your team's strengths, then pick one problem. Its baseline becomes the starting point for iteration.

Example first instruction to Scientifytext
Here is MCM/ICM 2027 Problem C with its attachments.
First check the data fields and missing values, list three usable modeling methods,
write a runnable baseline for each, and report metrics on the same validation data.

Why this stepOn days 2 and 3, push the baseline toward the level of award-winning papers.

Scientify proposes hypotheses round by round and replaces features, state definitions, model structures, and solvers. Every version records metrics on the same validation data, and rejected versions stay in the project for review.

Check the metrics every few hours and decide whether to continue or change direction. Runs continue while the team rests, and progress can be checked and redirected from a phone.

Development macro F1 by version in the 2024 MCM Problem C case

  1. v6e

    0.745

  2. v6n

    0.793

  3. v7

    0.700Rejected

  4. v8

    0.920

  5. v9

    0.937

From the session record of the Scientify 2024 MCM Problem C case project. The v7 route missed the freeze threshold and was rejected; v9 is the final version.
Example instruction during iterationtext
The current version scores 0.793 development macro F1; the weakest class F1 is 0.63.
Analyze per-match errors, propose two replacement state definitions,
implement both, and compare them on the same validation set without reading the test set.

Why this stepOn day 4, show that the results hold.

Freeze the final version and test it once on held-out data that never entered tuning. Report that result in the body and the summary sheet.

Then have Scientify run sensitivity and error analysis: change key parameters to see whether the conclusions stay stable, and list the cases where the model errs most. These results go into the model-testing section of the paper.

Why this stepFinish an English paper of at most 25 pages before 8:00 p.m. EST on February 1.

  1. 01

    Write the summary sheet

    Scientify writes the summary sheet first: problem, model, key numeric results, and conclusions. COMAP states that judges weigh the summary heavily.

  2. 02

    Lay out with the template

    Write the body and appendices in the COMAP LaTeX template, with the control number and page number in every header, within 25 pages.

  3. 03

    Export and check

    Export the PDF, named with the control number and under 25 MB. Read it through and check that every number matches the result files in the project.

  4. 04

    Submit

    Submit through the COMAP online form before 9:00 p.m. EST.

Concepts and judgment

Check your paper item by item against Outstanding papers. The COMAP results pages link to full past Outstanding papers.

  • Use the metrics of Outstanding papers on similar problems as targets.
  • The model has a baseline, and the paper explains why each improvement is better.
  • Final results come from held-out data; no test data was used for tuning.
  • Sensitivity and error analysis are included.
  • The summary sheet states the key numeric results.

Reference and appendix

Scientify independently completed 2024 MCM Problem C on tennis momentum: data checks, modeling, nine iterations, a held-out test, and the paper, with all code and records in one project.

Scientify held-out accuracy
96.02%
201 prediction points
Outstanding paper 2401445
82.40%
Player performance accuracy
Outstanding paper 2401445
77.60%
Momentum swing accuracy

New accounts get a free cloud computer and US$5 in credit. Paid plans are billed by monthly credits: Starter US$19/month (60 credits), Researcher US$29/month (100 credits), and Professional US$99/month (350 credits). One credit equals US$1 of standard API usage, at about 30% of the standard API price for the same usage.

References

Frequently asked questions

Which AI is best for math modeling?

It depends on the job. To explain a formula or rewrite a sentence, a chat assistant is enough. To have AI write code, run it, compare versions, and keep records, use an agent that executes code. Scientify runs code on a cloud computer and keeps the problem, code, data, logs, and results in one project, which suits four days of continuous iteration.

Can teams use AI in MCM/ICM?

Yes. COMAP allows large language models and other AI tools; their use is described in the Report on Use of AI after the paper, which does not count toward the 25 pages.

Can Scientify complete an MCM problem on its own?

Yes. On 2024 MCM Problem C, Scientify completed data checks, modeling, nine iterations, a held-out test, and the paper, reaching 96.02% held-out accuracy versus 82.40% in the Outstanding paper.

Does the team need to code?

Scientify writes and runs the code. It helps if one member can read Python to review code and results.

Does work stop when I close my laptop?

No. Runs continue on the cloud computer, and progress is visible from a phone.

How much does Scientify cost?

New accounts get a free cloud computer and US$5 in credit; paid plans start at US$19 per month.

Iterate an Outstanding-level model with Scientify in MCM/ICM 2027

When the contest opens, upload the problem and attachments. Scientify models, iterates models and algorithms, and writes the paper in the COMAP template on a cloud computer, keeping all code and run logs in the project. New accounts get a free cloud computer and US$5 in credit.