Since 2016, researchers at The Alan Turing Institute have run more than 80 datathon challenges with more than 60 partner organisations. When they wrote down what they had learned for the NeurIPS 2023 conference, they did not start with prizes, venues or team building. They started with the data: whether it fits the question, whether it is ready, how biased it is, how sensitive it is and whether there is enough of it (Mougan et al., 2023). For anyone planning a first datathon, that order is the real lesson.
If you run an accelerator, a university programme or a corporate data team, your hardest conversation will not be with sponsors. It will be with the data owner: the head of a department, a hospital’s IT lead or a data protection officer who has to agree that outsiders may work with the data at all. This guide shows how to organise a datathon around that conversation. It covers the formats, the data checks, an eight-week plan, teams and mentors, judging and the weeks after the event.
Our view in one sentence: a datathon is a data release project with an event at the end. If the dataset is not ready, not cleared and not tied to a question someone owns, no team can rescue it over a weekend.
What is a datathon?
The term is often used as a synonym for hackathon, so a clear definition helps. A datathon is a time-boxed event in which teams analyse a prepared dataset to answer a set question. It can last one day, a weekend or a full week, on site or online. The result is usually a model, an analysis, a visualisation or a short report, not an app.
The difference to a hackathon lies in what teams start with. At a hackathon, teams build something new, and our guide on how to host a hackathon argues that matching people before and after the weekend decides which prototypes survive. At a datathon, the raw material already exists. Every team works on the same data, so the quality of that data sets the ceiling for every result. A strong team with a broken dataset produces a careful explanation of why the data is broken. Nobody needed a weekend for that.
Datathons come in three main formats. The right one depends less on your ambition than on what the data owner can release.
| Format | How it works | Typical length | Typical output |
|---|---|---|---|
| Competitive datathon | Many teams work on the same task and are scored against a hidden test set or by a jury | One day to several weeks, often online | A ranking and winning models |
| Collaborative datathon | One team per challenge works with the challenge owner instead of against other teams | Often five days | A technical report and code |
| Open data datathon | Teams choose a question within a theme, using public datasets | One day to one week | Analyses and visualisations |
Each format has a well-known example. The WiDS Datathon run by Women in Data Science is competitive and global: its first 2024 challenge had more than 4,100 participants from 100 countries (WiDS Worldwide, 2024). The Turing’s Data Study Groups are collaborative: three to four challenges run in parallel for a week, with eight to twelve participants per challenge (Mougan et al., 2023). And the Virginia Office of Data Governance and Analytics publishes a short guide for schools and agencies that build datathons around the state’s open data portal.
For you, the choice follows from the data. Public data supports any format and any number of participants. Sensitive data points to a small collaborative group in a controlled environment.
Why the data decides everything
Whichever format you choose, the same questions about the data come first. The Turing team rates every dataset on five dimensions, each on a scale from “insufficient” to “optimal”. The table translates them into the questions you should put to the data owner before you set a date.
| Dimension | What it asks | Question for the data owner | Red flag |
|---|---|---|---|
| Appropriateness | Is the data relevant to the question? | Does it contain the variable we want teams to predict or explain? | The target variable is missing |
| Readiness | Is the data ready for analysis? | Is it clean, complete and documented in a data dictionary? | Teams would spend the event merging files |
| Reliability | How biased is the data? | Which groups, periods or sites are missing or overrepresented? | Nobody knows where the bias comes from |
| Sensitivity | How private or confidential is it? | Can it be shared openly, anonymised, synthesised or only in a secure environment? | Highly sensitive data with no safe route |
| Sufficiency | Is there enough of it? | Have similar volumes produced useful results before? | Too few records for any firm conclusion |
Source: based on the data assessment framework in Mougan et al., “How to Data in Datathons”, NeurIPS 2023.
Two of the five cause most of the trouble. The first is readiness. The data platform company Hex, writing from its own internal datathons, advises against letting teams clean raw data during the event, because time spent on preparation leads to weaker demos (Hex blog, 2022). Cleaning is not a team-building exercise.
The second is sensitivity. The Turing framework uses data tiers: open or synthetic data at the low end, data that needs safeguards against re-identification in the middle, and data so sensitive that a datathon is not supported at the top. The data owner decides which tier applies, and the tier decides almost everything else: who can take part, where they sit and which tools they may use.
How to organise a datathon in eight weeks
Because the data sets the pace, the plan starts with the data owner, not the venue. Eight weeks work for open data or data that has been cleared before. For sensitive data, allow far more time: the organisers of the Technion and Rambam healthcare datathon in Israel started planning about six months ahead (Sobel et al., 2022).
Planning timeline
A datathon in eight weeks
- Week 8 Challenge owner and question Name the person who will use the answer and write one question per challenge.
- Week 7 Assess the data Rate it on the five dimensions with the data owner. Drop challenges with a red flag.
- Week 6 Clear privacy and terms Agree the data form, access, ownership of results and deletion on one page.
- Weeks 5 to 3 Recruit with skills data Ask applicants for role, methods and domain knowledge, then select a mix.
- Week 2 Test the environment Every team logs in, runs a test script and confirms access to sample data.
- Week 1 Brief the teams Send the data dictionary, seed questions, rules and judging criteria.
- Event Analyse, present, decide Mentor check-ins, short presentations and a jury that includes the challenge owner.
- Days 1 to 30 Follow-up Booked meetings, a written report, data deletion and a check-in on day 30.
Write a question the data can answer
The question decides what kind of work you get. “Find insights in our customer data” invites twenty unrelated dashboards. “Which features of last year’s support tickets predict a cancellation within 90 days?” invites comparable answers that a challenge owner can act on. The second question also passes the appropriateness test, because it names a target variable the data must contain.
Write one question per challenge and add a handful of “seed questions” that help teams get started, as Hex recommends. Then check that someone on the business side wants the answer. In the Turing’s Data Study Groups, the challenge owner sets the challenge and stays for the whole week, but does not direct the work (Digital Catapult call for challenge owners). That balance keeps the question honest.
Clear the data before you invite anyone
This is the step that most first-time organisers underestimate. The Turing team describes its data protection and ethics review as a separate step “comparable in size to data assessment”. In other words, plan as much time for the legal side as for the data itself.
Under the European General Data Protection Regulation (GDPR), the key distinction is between pseudonymised and anonymous data. Recital 26 states that pseudonymised data, which can be linked back to a person with additional information, is still personal data. Only information that no longer relates to an identifiable person falls outside the data protection principles. Removing names is therefore rarely enough. If real anonymisation is not possible, three routes remain: use comparable open data, use synthetic data that mirrors the original population (the Turing team did this for a government partner), or keep the real data inside a secure environment that participants cannot leave with copies.
The data owner signs once, so give them one page. It should answer five points:
- Which data, in which form: the exact extract, its anonymisation or synthesis and who prepared it.
- Who may access it, and where: registered participants only, on site or in a named environment.
- What participants may take home: usually code and slides, never the data.
- Who owns the results: models, code and reports, and whether they may be published.
- When the data is deleted: a date, and who confirms it.
The Technion and Rambam organisers combined several of these measures: anonymised data prepared by the hospital’s teams, an updated ethics approval and a non-disclosure agreement for every participant (Sobel et al., 2022). This is general information, not legal advice. Check your own case with your data protection officer.
Test the environment two weeks before
The second most common failure is technical: forty people arrive and half of them cannot open the data. The Technion and Rambam datathon gave each team its own virtual machine in a secure cloud, with at least eight cores and 16 GB of memory. Teams tested the setup two weeks before the event, and the full databases were loaded only the day before.
For open data you do not need that level of control. A shared repository and a clear data dictionary may be enough. Either way, the dictionary matters. The Turing framework rates data as fully ready only when it is clean, complete and documented with a data dictionary or data card. Ten minutes of reading saves hours of guessing.
Build mixed teams and give each a mentor
A datathon rewards teams that combine methods with knowledge of the field. In the healthcare datathon in Haifa, the most successful teams mixed clinicians, data managers and engineers. Every team also had two mentors, a clinician and a senior data scientist, and the organisers concluded that mentors who follow a team through the whole event are necessary.
One datathon, up close
The Technion and Rambam healthcare datathon
Team sizes vary with the format. The Virginia guide suggests three to five people per team for open data events, while the Turing’s collaborative groups put eight to twelve people on one challenge. More important than the number is the mix. Ask every applicant for their role, their methods and their domain knowledge, and form teams before the event rather than during the kick-off. The registration form is where that mix is won or lost, and our guide to registration questions for matchmaking shows which questions to ask before and after confirmation.
Judge the insight, not only the score
Competitive datathons often score submissions automatically against a test set the teams never see. That is fair and fast, but a leaderboard rewards the best number, not the most useful answer. Teams can also tune their models to the visible part of the scores, which is why many organisers keep part of the test set back for the final ranking.
For most organiser goals, a jury with a published rubric works better. The Virginia guide scores creativity, impact, use of data and presentation. A version that keeps the challenge owner’s question at the centre could look like this:
| Criterion | What judges ask | Suggested weight |
|---|---|---|
| Answer to the question | Does the result answer the question the challenge owner asked? | 30% |
| Method and validity | Is the approach sound, and would it hold on new data? | 25% |
| Use of the data and its limits | Did the team spot gaps and bias, and say what they mean? | 20% |
| Usefulness | Could the challenge owner act on it within six months? | 15% |
| Communication | Can a non-specialist follow the result in five minutes? | 10% |
The third criterion is deliberate. A team that says honestly what the data cannot show has done the challenge owner a favour. Put the challenge owner, a data scientist and a domain expert on the jury. Keep presentations to three to five minutes, as Hex suggests, and with many teams, let judges pick finalists from short videos first.
After the datathon: where the data owner meets the teams
The data owner who signed your one page will ask one question afterwards: was it worth releasing the data? The answer depends on what happens in the next 30 days, and that part is easy to forget once the room is cleared.
The Turing model builds the follow-up into the format. After the week, the challenge owner receives a report on the approaches explored and possible next steps, plus the code; the report may be published, with redactions where commercial interests require them. Hex calls dedicated follow-up time for the best projects “the ultimate prize”. And the Technion and Rambam datathon produced something beyond its models: a roadmap for safe data sharing between a university and a hospital.
Plan four things before the event ends:
- Book a meeting between every finalist team and the relevant challenge owner, on site, before people leave.
- Agree a date for the written report and name the person who edits it.
- Confirm the data deletion, or a new access agreement for teams that continue.
- Hold a check-in on day 30 to see which results are in use.
A datathon is the wrong tool when the business already knows its problem and needs a supplier with a working product. In that case, our guide to the open innovation challenge shows how a public call and clarification meetings lead to a paid pilot.
Solution: Converve supports the people side of a datathon. Registration questions capture each participant’s role, methods and field, which you can use when forming teams. Challenge owners, mentors and teams then request and confirm one-to-one meetings in time slots you define, based on a rule-based meeting matrix that your team controls. Our solution for startup and investor events shows how this works in practice.
Frequently Asked Questions
What is a datathon?
A datathon is a time-boxed event in which teams analyse a prepared dataset to answer a set question. It usually lasts between one day and one week. The result is a model, an analysis, a visualisation or a short report rather than a software product.
How do datathons work?
The organiser and a challenge owner prepare a dataset and one question per challenge. Teams of three to twelve people, depending on the format, analyse the data with support from mentors. At the end, a jury or an automatic score on a hidden test set ranks the results, and the best teams present them to the challenge owner.
What is the difference between a datathon and a hackathon?
At a hackathon, teams build a new prototype, such as an app or a tool. At a datathon, every team works on the same existing data, so data preparation and data protection matter far more. A hackathon tests what teams can build; a datathon tests what the data can show.
How long does a datathon last?
Anything from one day to a week. The Alan Turing Institute’s Data Study Groups run for five days, many university and company datathons take a weekend or two days, and online competitions such as the WiDS Datathon run for several weeks. Preparation takes longer: we recommend about eight weeks for open data, and the Technion and Rambam healthcare datathon was planned about six months ahead.
Can you use personal data in a datathon?
Only with great care. Under the GDPR, pseudonymised data is still personal data (Recital 26), so removing names is not enough. Organisers use truly anonymised extracts, synthetic data or secure environments that participants cannot leave with copies, combined with an ethics review and a non-disclosure agreement. This is general information, not legal advice.
What are good datathon ideas?
Good datathon ideas start with a dataset and a decision someone wants to make. Examples include predicting which support tickets lead to cancellations, finding patterns in hospital admissions with anonymised data, or using a city’s open transport data to show where services are missing. Avoid open questions such as “find insights”, which produce answers nobody can compare.
Conclusion: clear the data first
A datathon looks like a weekend of analysis. In practice, it is a data release project with an event at the end. Check the data on five dimensions, clear it with the data owner on one page, test the environment two weeks ahead and write a question someone actually wants answered. Then build mixed teams, judge the insight rather than only the score, and book the meetings before everyone goes home.
Do that, and the data owner’s question has a good answer. If you want to plan the team formation and the meetings around your next datathon, get in touch with Converve.
