Early October 2026 · Venue TBA

Gaussathon

A hackathon for people who estimate things. Two days, a handful of genuinely difficult datasets, and no way to hide behind a chat window. Bring priors, bring derivations, bring a model you can defend at a whiteboard.

No language models. No agents. No prompt engineering.

Register interest Read the charter
days
—
hrs
—
min
—
sec
—
§1

The charter

What this event is, stated plainly, so nobody wastes a weekend.

Most hackathons have collapsed into the same event: wire four APIs together, put a chat box on top, demo it before it breaks. That is a fine skill. It is not this skill.

Gaussathon is a room full of statisticians, econometricians, quants, signal processing people, epidemiologists, Kagglers and applied mathematicians, all pointed at problems where the data is scarce, contaminated, censored, confounded or simply strange — the kind of problem where a bigger model does not help and a better estimator does.

You will be judged not only on whether your number is good, but on whether your reasoning is beautiful.

  1. No language models

    No LLMs in your submitted pipeline — not as a featuriser, not as an embedder, not as a labeller, not as a judge. If the artefact ships with a tokenizer, it does not ship.

  2. No agents, no copilots

    Write your own code. Autocomplete off, assistants closed. We are not moralising about tools in general; we are protecting the one weekend a year where the modelling is unambiguously yours.

  3. Method over horsepower

    Compute is capped and identical for everyone. A gradient-boosted kitchen sink will not out-run a well-specified likelihood on these datasets, and it will lose on the other two tiers regardless.

  4. Show the derivation

    Every submission carries a short technical note: model, assumptions, identification strategy, and where it breaks. Uncertainty quantification is not optional. A point estimate with no interval is an unfinished sentence.

  5. Everything open at the close

    Code and notes are released publicly when scoring ends. The archive is the point: a permanent, citable record of how strong people attacked the same hard problem five different ways.

§2

Three tiers of judgement

Lexicographic, not a weighted sum. Tier 1 gates. Tier 2 separates. Tier 3 breaks ties and wins the room.

01

Performance

A single pre-announced metric on a held-out set you never see. Sealed until the deadline; scored automatically.

  • Public leaderboard on a subsample, private leaderboard decides.
  • Proper scoring rules where the target is a distribution — CRPS, log score, pinball.
  • Bootstrap intervals reported on every rank, so nobody wins by a rounding error.

Gates entry to the finals. Being fastest is not enough to win.

02

Elegance

Does the approach explain the data-generating process, or merely fit the file? Judged from your technical note by a panel.

  • Parsimony: parameters earned, not spent.
  • Assumptions stated, tested, and honestly reported when violated.
  • An idea a reader can carry home and reuse on a different problem.

The tier most teams underestimate, and the one the judges argue about.

03

Complexity

Depth and ambition of the model — the structure you were willing to take on, and actually made work.

  • Hierarchy, latent structure, measurement error, non-stationarity, censoring.
  • Inference machinery you built rather than imported.
  • Credit for hard things that converged. Ambition without diagnostics scores zero.

Complexity in service of the problem. Never complexity as decoration.

Panel composition and the exact rubric are published with the problem statements, before the clock starts. Judges recuse themselves from teams they have co-authored with.

§3

The problems

Released at kickoff. Below is the flavour — real datasets, sourced from working scientists and industry partners, all of them nasty in a specific and interesting way.

Track A

Estimation under scarcity

A few hundred observations, thirty candidate covariates, and a coefficient somebody will make a decision with. Shrinkage, partial pooling, and the discipline to say how little you know.

Track B

Signal in a hostile process

Non-stationary, heavy-tailed, irregularly sampled, and contaminated by an instrument that changed halfway through the record. Find the thing that is actually there.

Track C

Many rows, few labels

Millions of numeric and time-series measurements, of which a few hundred carry a label, because each one cost an expert an hour. Make the unlabelled mass earn its keep, and show what it bought you over a labelled-only baseline.

Track D · data partner

Structure without supervision

Twenty-eight instruments across equities, FX, rates and metals, updating hourly, with no labels and no agreed definition of what a regime even is. Learn a representation that makes the cross-asset structure legible — then defend it on probes you did not get to choose, against a production embedding as the baseline.

Open

The wildcard

One problem nobody sees in advance, including most of the judges. Announced on the morning of day two, scored on the same three tiers, weighted for the shorter clock.

Track D data is supplied by Heston Labs, whose Latent API streams the series and the 128-dimensional embedding that serves as the baseline. Gaussathon's organiser founded Heston Labs and is recused from judging that track.

§4

Format

Two days. Teams of one to four. Times are indicative until the venue is confirmed.

  1. Day 1 · 09:00

    Problems released

    Statements, data dictionaries, baselines and the scoring harness go live simultaneously. No head start for anyone.

  2. Day 1 · 10:00

    Open modelling

    Work in the hall or wherever you think best. Judges and invited statisticians float for consultation — asking a good question of them is encouraged, not penalised.

  3. Day 1 · 19:00

    Chalk talks

    Short, optional, whiteboard-only. Somebody explains an estimator they love. Historically the best part of the weekend.

  4. Day 2 · 09:00

    Wildcard drops

    The unseen problem opens for anyone who wants it.

  5. Day 2 · 16:00

    Submissions close

    Code, technical note, and a reproducible run. Private leaderboard is revealed at the deadline, not before.

  6. Day 2 · 17:00

    Defence & results

    Shortlisted teams present for eight minutes and take questions from the panel. Prizes, then the archive goes public.

Teams
1–4 people
Cost
Free to attend
Compute
Identical, capped, provided
Languages
Anything — R, Python, Julia, Stan
Format
In person
Places
Limited; applications reviewed
§5

Questions

Is the no-LLM rule actually enforceable?

Not perfectly, and we are not pretending otherwise. We enforce it three ways: submitted environments are inspected, every shortlisted team defends its derivation live in front of a panel that will ask why you chose that link function, and the whole archive is public afterwards. It is an honour system with teeth. People who want to cheat at a statistics hackathon have already lost the thing worth winning.

Can I use scikit-learn, XGBoost, Stan, PyMC, brms?

Yes. Standard statistical and ML libraries are entirely welcome — this is a ban on language models and coding agents, not on established software. Neural networks are allowed where they genuinely suit the problem, though on these datasets they rarely do, and tiers 2 and 3 will ask you to justify the choice.

Do I need a PhD?

No. We care about how you think about uncertainty, not your credentials. Strong self-taught modellers, working quants, and undergraduates who read too much are all welcome. Applications are reviewed because places are limited, not because we are screening for titles.

What do I actually submit?

Three things: predictions in the specified format, a reproducible code bundle that runs end to end inside the provided environment, and a technical note of at most four pages covering the model, assumptions, inference, diagnostics and failure modes. The note is what tiers 2 and 3 are scored from.

Can I come without a team?

Yes, and many do. There is a team-forming session before the problems drop, and solo entries are competitive — a single person with the right prior has won things like this before.

Who owns the work?

You do. Publishing openly at the close is a condition of entry, under a permissive licence of your choosing, but the work and any paper that comes out of it are yours.

Come and estimate something hard.

Registration is by application and places are limited. It takes two minutes: who you are, what you have modelled, and which track you are drawn to.

Apply to attend Sponsor the event

Early October 2026 · Venue TBA · Free to attend