16 Data Science Projects With Real Datasets (Beginner to Advanced)

Last updated: November 2026. Written by Josh Hutcheson, OnlineCourseing editor. See our review methodology.

QUICK VERDICT

Data science projects are how you show you can do the job before anyone has hired you to do it. This list gives you 16 of them, grouped by level and by career goal. Each one names a public dataset we loaded on 6 November 2026, its size, what the project proves, our estimate of the time it takes and the question we would expect a hiring manager to ask about it.

Time estimates and interview questions are our own judgment, and so are the opinions on what impresses reviewers. We quote no hiring statistics here because we did not find a source strong enough to use. For three of the datasets we also ran the first steps ourselves, so you can see what you are walking into.

If you want SQL, spreadsheet and dashboard work, see our 15 data analytics projects. That page covers fraud detection, movie recommendations, review sentiment, fake news and weather forecasting, so none of those are repeated here. If you need the skills first, start with our guide to the best data science courses.

Data science projects at a glance

Why Titanic-style projects no longer impress

These projects teach well and persuade poorly. That is our opinion, and it comes down to three facts about them.

• They are small and solved. UCI lists iris at 150 rows and describes it as “an exceedingly simple domain”. The OpenML copy of Titanic has 1,309 passengers. Neither one forces a hard decision about messy data.
• The answers are a search away. Titanic is a well-known Kaggle competition with public solutions all over the web, so a reviewer cannot tell your thinking from a tutorial you followed.
• Nothing depends on the result. No one acts on a prediction of who survived in 1912, so you never have to weigh the cost of being wrong. Real work starts with that question.

Use them in your first weeks if they help. Leave them out of the portfolio you send to employers. The same goes for any dataset you meet in a course, unless you add your own question.

Beginner data science projects

These run on a laptop and need Python, pandas and basic statistics. Each has an obvious baseline, so you can show improvement in plain numbers.

1. Ames house price model

• Dataset: Ames Housing documentation: 2,930 Iowa home sales from 2006 to 2010, 82 columns, published as an alternative to the Boston housing data.
• Level and time: Beginner. About 15-25 hours, by our estimate.
• What to do: Predict sale price and report mean absolute error in dollars. When we split the data 80/20 and tried it, predicting the median price every time missed by $54,236 on average. Living area alone cut that to $38,191. Adding overall quality and year built cut it to $26,398. Your job is to beat that and explain which features earned their place.
• What it proves: Regression, feature selection and comparison against a baseline.
• What a hiring manager is likely to ask: “Which homes does your model get most wrong, and why?”

2. MineThatData email test analysis

• Dataset: MineThatData E-Mail Analytics Challenge: 64,000 customers, randomly split three ways into a men’s email, a women’s email and no email. The CSV is about 4 MB.
• Level and time: Beginner. About 10-20 hours, by our estimate.
• What to do: Decide which email to send and to whom. Our first pass: visit rates were 18.3% (men’s), 15.1% (women’s) and 10.6% (no email). Mean spend per customer was $1.42, $1.08 and $0.65. The men’s email beat no email by $0.77 per customer, with a 95% interval of about $0.49 to $1.05. Next, test whether the winner changes across customer segments.
• What it proves: Experiment analysis, confidence intervals and turning a result into a decision.
• What a hiring manager is likely to ask: “If each email cost five cents to send, would you still send it to everyone?”

3. Income prediction with a fairness audit

• Dataset: UCI Adult: 48,842 rows and 14 features from the 1994 US Census database, with the task of predicting income over $50,000.
• Level and time: Beginner. About 15-25 hours, by our estimate.
• What to do: The classifier is the easy part, and many tutorials stop there. Compare error rates across sex and race groups, then write up what you would change before anyone used it for a real decision. Note in your write-up that $50,000 in 1994 is a dated threshold.
• What it proves: Classification plus the judgment to test a model for uneven errors.
• What a hiring manager is likely to ask: “Which fairness measure did you pick, and what does it trade away?”

4. SMS spam classifier

• Dataset: UCI SMS Spam Collection: 5,574 labeled text messages.
• Level and time: Beginner. About 10-15 hours, by our estimate.
• What to do: Build a bag-of-words model with logistic regression. Report precision and recall, then pick a threshold and defend it. A real user would hate losing one genuine message far more than seeing one spam message.
• What it proves: Text cleaning, thresholds and metrics chosen to match a cost.
• What a hiring manager is likely to ask: “Why not simply report accuracy?”

Intermediate data science projects

These add a trap that a tutorial would normally hide: leakage, repeated records, time order or a business cost you have to estimate.

5. Hotel cancellation model

• Dataset: Hotel booking demand via TidyTuesday: 119,390 bookings by our row count, from Antonio, Almeida and Nunes (2019).
• Level and time: Intermediate. About 20-30 hours, by our estimate.
• What to do: Predict cancellations and design an overbooking or deposit policy from the result. Two traps we found on loading it: 37.0% of bookings cancel, so a model that always says “no” is already 63% accurate. And the reservation_status column matches the target exactly (every “Canceled” or “No-Show” row is a cancellation), so leaving it in produces a perfect and useless model. We also counted 31,994 exact duplicate rows, which you need to investigate before splitting.
• What it proves: Spotting leakage and tying a model to a policy with a dollar value.
• What a hiring manager is likely to ask: “Which columns would not exist at the moment you need the prediction?”

6. Developer salary model

• Dataset: Stack Overflow Developer Survey 2025: over 49,000 responses from 177 countries. Earlier years and the 2026 survey are also published, under an open license.
• Level and time: Intermediate. About 15-25 hours, by our estimate.
• What to do: Model pay by role, experience, country and technology, then compare years. Most of the work is cleaning and deciding what to do with currencies and outliers. Be explicit that these are people who chose to answer a Stack Overflow survey, so the results describe that group and not all developers.
• What it proves: Messy survey data, sampling bias and clear communication of limits.
• What a hiring manager is likely to ask: “Who is missing from this data, and how does that change your conclusion?”

7. Credit card default risk model

• Dataset: UCI Default of Credit Card Clients: 30,000 Taiwanese card customers with 23 features and payment records from April to September 2005.
• Level and time: Intermediate. About 20-30 hours, by our estimate.
• What to do: Predict default as a probability and test whether those probabilities are honest, using a calibration curve. Then pick a credit limit or review policy and estimate what it would save. Note that the data is two decades old and from a single market.
• What it proves: Calibrated risk estimates, which finance teams care about more than a ranking score.
• What a hiring manager is likely to ask: “A customer asks why they were declined. What do you tell them?”

8. Hospital readmission model

• Dataset: UCI Diabetes 130-US Hospitals: 101,766 encounters and 47 features covering 1999 to 2008.
• Level and time: Intermediate. About 25-40 hours, by our estimate.
• What to do: Predict readmission within 30 days. The dataset has a patient_nbr field, and encounters are the row level, so split by patient or the same person can land in both training and test data. Finish with how a care team could use the score and what it would cost them in false alarms.
• What it proves: Grouped data splits, healthcare context and careful language around a sensitive outcome.
• What a hiring manager is likely to ask: “Your model flags 20% of patients. How many can a care team actually call?”

9. Household electricity forecast

• Dataset: UCI Individual Household Electric Power Consumption: 2,075,259 one-minute readings from one home near Paris over 47 months. UCI says nearly 1.25% of rows have missing values.
• Level and time: Intermediate. About 25-35 hours, by our estimate.
• What to do: Forecast next-day usage. Start with a seasonal naive baseline (same hour last week) and only keep a model that beats it. Validate with scikit-learn’s TimeSeriesSplit, because a random split lets the model see the future.
• What it proves: Time series validation and handling gaps in sensor data.
• What a hiring manager is likely to ask: “How much better than last week’s value is your forecast, and is the gain worth the complexity?”

Advanced data science projects

These are for people applying to machine learning or senior analytics roles. The modeling matters less here than the engineering, causal reasoning or evaluation around it.

10. Jet engine remaining useful life

• Dataset: NASA CMAPSS Jet Engine Simulated Data: four simulated datasets totaling 708 training engines (100, 260, 100 and 248), with 26 columns per cycle.
• Level and time: Advanced. About 40-60 hours, by our estimate.
• What to do: Predict how many cycles each engine has left. Start with the simplest set, FD001, then move to FD004, which has six operating conditions and two fault modes. Translate the output into a maintenance rule. Say clearly that the data is simulated.
• What it proves: Sensor data, sequence modeling and a metric where predicting late costs more than predicting early.
• What a hiring manager is likely to ask: “Which error is worse here, too early or too late, and did your loss function reflect that?”

11. Advertising uplift model

• Dataset: Criteo Uplift Prediction Dataset: 13,979,592 rows in the corrected version, 297 MB compressed, with 85% of users treated.
• Level and time: Advanced. About 40-60 hours, by our estimate.
• What to do: Estimate who buys because of an ad, as opposed to who would have bought anyway. Criteo’s own erratum says the first version of the data had a leak, so use the un-biased version and mention why. Conversions are rare at about 0.29%, which makes evaluation the hard part.
• What it proves: Causal thinking, which many portfolios never touch.
• What a hiring manager is likely to ask: “How is uplift different from predicting conversion, and how did you evaluate it with no ground truth for each user?”

12. Satellite land-use classifier as a live app

• Dataset: EuroSAT on Zenodo: 27,000 labeled Sentinel-2 images in 10 classes. The authors report 98.57% accuracy.
• Level and time: Advanced. About 40-60 hours, by our estimate.
• What to do: Use transfer learning, then deploy it so anyone can upload an image. Do not treat the accuracy as the goal, since the benchmark is already high. Test it on images from a region or season outside the training data and report what breaks.
• What it proves: Deep learning, deployment and honest testing outside the training distribution.
• What a hiring manager is likely to ask: “Where does this fail, and how would a user know?”

13. Taxi demand pipeline with drift monitoring

• Dataset: NYC TLC Trip Record Data: monthly Parquet files for yellow, green and for-hire trips, published with a typical two-month delay.
• Level and time: Advanced. About 60-80 hours, by our estimate.
• What to do: Build a pipeline that ingests a new month, retrains or scores a zone-level demand model, and flags when input data or accuracy shifts. The monthly release gives you a built-in test of whether your monitoring works. TLC warns that it does not vouch for the accuracy of the data and that the schema may change, so your pipeline needs checks for both.
• What it proves: Production habits: scheduling, data validation, tests and monitoring.
• What a hiring manager is likely to ask: “What happens to your pipeline if a column changes type next month?”

LLM and RAG project ideas, with honest caveats

RAG (retrieval-augmented generation) means fetching relevant documents and giving them to a language model to answer from. It is easy to make a demo and hard to make it reliable, so a project that measures reliability stands out more than one more chatbot, in our opinion.

Four caveats before you start:

• You will likely pay for API calls. Set a spending cap, sample your data and report what a run cost.
• Results age quickly. Name the model and date in your write-up, since a rerun on a newer model may give different numbers.
• Never commit an API key. A key in a public repository tells a reviewer more than the project does.
• This is not proven to beat other projects. We found no reliable hiring data on it. Prioritize it if the jobs you want mention LLMs.

14. RAG over SEC 10-K filings

• Dataset: FinanceBench: 150 open annotated questions with evidence, drawn from a larger set of 10,231. Filings are available through EDGAR’s APIs.
• Level and time: Advanced. About 40-60 hours, by our estimate.
• What to do: Build retrieval over filings and score it against the benchmark answers. Measure retrieval and answer quality separately, using metrics such as context recall and faithfulness. EDGAR asks you to declare a user agent and stay under 10 requests per second. For a reality check, the FinanceBench authors reported in 2023 that GPT-4-Turbo with a retrieval system incorrectly answered or refused 81% of their sample. Newer models will differ, and measuring how much is the project.
• What it proves: Evaluation discipline, the skill most LLM demos skip.
• What a hiring manager is likely to ask: “When your system got an answer wrong, was the failure in retrieval or generation?”

15. LLM vs. baseline on contract clause extraction

• Dataset: Contract Understanding Atticus Dataset (CUAD): more than 13,000 expert labels across 510 commercial contracts and 41 clause types, under a CC BY 4.0 license. The Zenodo download is 105.9 MB.
• Level and time: Advanced. About 20-40 hours, by our estimate.
• What to do: Pick five of the labeled clause types and find them in held-out contracts. Compare a TF-IDF and logistic regression baseline with an LLM on accuracy, cost per contract and speed. Missing a clause is usually worse than flagging an extra one, so report recall for each type. Say clearly that this is a data project and that the output is not legal advice.
• What it proves: The judgment to ask whether an LLM is worth it, including when the answer is no.
• What a hiring manager is likely to ask: “At what volume would you pick the cheaper model, and how much recall would you give up?”

16. Text-to-SQL evaluation

• Dataset: BIRD: 12,751 question-SQL pairs across 95 databases totaling 33.4 GB. Its Mini-Dev set has 500 questions and was updated in October 2026.
• Level and time: Advanced. About 40-60 hours, by our estimate.
• What to do: Use Mini-Dev to test whether an LLM can answer business questions in SQL. Run the generated queries and compare results, because SQL that looks right can return the wrong rows. Group failures by type, such as wrong joins or misread columns, then fix one with better schema context and re-measure.
• What it proves: SQL strength plus a useful, measurable LLM application.
• What a hiring manager is likely to ask: “Would you let this run against production data, and what guardrails would you add?”

What kind of projects should you do for a data science job?

Choose by the job you want. Read five real postings for that title first and note the skills they repeat. Here is how we would match this list to common goals:

When two projects look equally useful, pick the one you can talk about with interest for 20 minutes. If you have worked in an industry, use its data, because that background is hard for other candidates to copy. Our guide to getting an entry-level data science job covers the rest of the job search, and our interview prep guide covers how to discuss your work in a project round.

How many data science projects do you need?

Our answer is three finished projects, with a fourth only if it covers something the first three do not. One strong project is enough to begin applying, so do not wait for a perfect set. This is our judgment. We found no credible survey that sets a number.

The Bureau of Labor Statistics projects data scientist employment to grow 35% from 2025 to 2035, with about 24,800 openings a year. Those figures say nothing about how competitive entry-level roles are, and many openings replace people leaving the field. We cover that gap in our data science career guide. A smaller number of finished projects with clear write-ups will serve you better there than a long list of unfinished notebooks.

How to use statistics in data science projects

Statistics is what lets you say a result is real and how large it is. Five habits cover most projects:

• State a baseline first. Our Ames test shows why: a $26,398 error means little until you see that guessing the median misses by $54,236.
• Report intervals as well as point estimates. “$0.77 more per customer, between $0.49 and $1.05” supports a decision. “$0.77” alone does not.
• Check sample size before you run an experiment. statsmodels has power and sample size calculations for common tests. A small test can miss a real effect.
• Split data the way it will be used. Respect time order and keep repeated people together, as in projects #8 and #9.
• Separate practical from statistical significance. With millions of rows, tiny effects become significant. Say whether the effect is large enough to act on.

Testing many segments also raises the odds of a false win, so tell readers how many comparisons you ran.

Where to get data for data science projects

• UCI Machine Learning Repository: well-documented classics with a stated task, used for several projects above.
• Data.gov: 535,095 US government datasets listed when we checked.
• Hugging Face Datasets: 1,119,306 listed when we checked, and the natural place for text and LLM work. Quality varies, so read each dataset card.
• Google Dataset Search: a search engine for datasets across the web.
• Kaggle: a large free source of datasets and notebooks, and useful for reading how others approached a problem. Its pages returned a CAPTCHA to our test browser, so we could not confirm current details or load individual Kaggle datasets, and none of the 16 projects above depend on it.
• Agencies directly: the SEC and NYC TLC publish real operational data with documentation. The CFPB Consumer Complaint Database returned 18,965,764 complaints through its API when we checked, and is useful for trend analysis. The CFPB says it is not a statistical sample, so do not generalize from it to all consumers.

Before you commit, check the license, the date, how the data was collected and whether it has real errors you can write about. Data you collect yourself through a public API can also set you apart, as long as you follow the source’s terms.

How to write up and publish a project on GitHub

GitHub says a README is often the first item a visitor sees, so write it for a reviewer with two minutes. Our suggested structure:

• Result first. One sentence with the finding and the number, then the recommendation.
• Question and why it matters. Who would use this, and what decision would it change?
• Data. Source link, date, size, license and known problems.
• Method. Your baseline, what you tried and why you kept or dropped it.
• Limits. Where it fails and what you would do next.
• How to run it. Exact commands and a pinned list of dependencies.

For model projects, add a short model card covering intended use and performance across groups.

How to organize notebooks and files

Use a standard layout such as Cookiecutter Data Science and keep raw data separate from cleaned data. Number your notebooks in order, move reusable code into scripts and rerun everything from a clean start before you publish. Then a stranger can reproduce it, which is what you are claiming they can do.

Watch file sizes. GitHub blocks files larger than 100 MiB and warns above 50 MiB, so link to source data or include a script that downloads it. Add a .gitignore for data and credentials before your first commit.

Where to host and run data science projects online

For most people, a pinned GitHub repository with a clear README and one live app is plenty. You do not need a GPU workstation for this list, and we would not buy one for a portfolio. A free AWS account can work too, but cloud setup and billing risk are a separate project unless the jobs you want ask for them.

How to write about a project on your resume

Use one line per project: the question, the data, the method and the result. For example: “Modeled hotel cancellations on 119,390 bookings, removed a leaking status column and proposed a deposit policy with an estimated revenue effect.” Only quote numbers you can defend, and link the repository. Put projects above education if you have no related job history, and make sure you can explain every line in an interview.

Where to get real-world data science experience

• DrivenData: social-impact competitions, plus practice competitions, including an LLM document summarization one when we checked.
• DataKind: a data-for-good organization that invites volunteers. In our view this suits people who already have some skills.
• Your current job: a small analysis that changes one real decision is often the best project you can have. Check what you may share publicly, and leave out company data.
• A local nonprofit or small business: offer a scoped analysis with a clear deliverable. A real user will give you feedback that no dataset can.

How to get your data science portfolio reviewed

Do a self-check first. Ask someone outside data to read your README for two minutes, then tell you what you found and why it matters. If they cannot, fix that before asking an expert. Then pick a route:

You will get better feedback with a narrow question. Send one project and one target job posting, and ask: “Would this make you want to interview me for this role? What is the first thing you would doubt?” Avoid asking whether your portfolio is good. Treat advice from any one person as a single opinion, and change something only if you hear it twice or the reason convinces you.

Courses that help if you need skills or guided projects first

You do not need a course to do these projects if you can already work in Python and SQL. If you cannot, a structured course saves time. Here are the options we would look at, with full comparisons in our best data science courses guide:

• IBM Data Science Professional Certificate (Coursera): our pick for beginners. It is a 12-course series with an applied capstone, rated 4.6 from 151,401 reviews, and Coursera estimates 4 months at 10 hours a week. See our IBM certificate review.
• DataCamp Projects: 135 guided projects listed when we checked, which is useful practice before you work alone. Treat them as training, since other learners complete the same ones. See DataCamp pricing for what plan you would need.
• Udacity Data Scientist Nanodegree: the advanced option, listed at 61 hours, for people who want a person to review their project work.
• Generative AI with Large Language Models (Coursera): a short grounding before projects #14 to #16, rated 4.8 from 3,660 reviews. Our generative AI course guide compares more options.

Frequently asked questions

What are data science projects?

A data science project uses data and statistics or machine learning to answer a question or make a prediction, then explains the result and what to do about it. For a job seeker, it is evidence that you can do the work before anyone has hired you to do it.

What are good data science projects for beginners?

Start with a project where a simple method can give a defensible answer. Our picks are the Ames house price model, the MineThatData email test and the SMS spam classifier. All three load on a laptop, and each has a clear baseline to beat.

How many data science projects do you need for a portfolio?

Our view is three finished projects, and one is enough to start applying if it is strong. Cover one statistical analysis, one model compared against a baseline and one that someone can use or run. This is our judgment from reading the role requirements. We have no hiring survey that sets a number.

What kind of projects should I do for a data science job?

Match the job title. For analyst-leaning roles, do experiments and business analysis. For machine learning engineer roles, do something deployed and monitored. For a specific industry, use that industry’s data and its real cost of being wrong.

Where can I get data for data science projects?

The UCI Machine Learning Repository, Data.gov, Hugging Face Datasets, Google Dataset Search and Kaggle are the usual starting points. Government agencies such as the SEC, CFPB and NYC TLC publish large, real datasets.

Do Titanic and iris projects still count?

They are good practice and weak portfolio pieces, in our opinion. Iris has 150 rows and UCI itself calls it an exceedingly simple domain. Learn on them if they help, then show something else.

How do I list data science projects on my resume?

Give each one a line with the question, the data, the method and the measured result, with a link to the repository. Put them under a Projects heading, and place that heading above education if you have no related job history.

Where should I host a data science project?

Keep the code on GitHub. Use Streamlit Community Cloud to share a simple app free, or Hugging Face Spaces for a demo, checking the current Space plans first. GitHub Pages works for a static write-up.

Do I need a GPU for data science projects?

Not for most of this list. The tabular projects run on an ordinary laptop. Google Colab offers free GPU access for the image project, though Google says its resources are not guaranteed and usage limits change.

Can I use a guided course project in my portfolio?

Yes, as long as you say it was guided and add your own work, such as a second dataset, a different question or a deployed version. Reviewers tend to recognize popular capstones, in our view, so the copy alone tells them little.

Should my portfolio include an LLM or RAG project?

Include one if you are applying to roles that mention LLMs, and make its evaluation the focus. For other data science roles, treat it as optional. We have not found reliable hiring data showing it beats a strong conventional project.

How long does a data science project take?

Our estimates run from about 10 hours for a focused beginner project to 80 hours for a deployed pipeline with monitoring. Budget a good share of that for the write-up, because it is the part most people read.

Related reading

• 15 data analytics projects with free datasets
• Best data science courses
• How to become a data scientist without a degree
• How to get an entry-level data science job
• Data science interview prep