My Kaggle Playground competitions in 2025

Kaggle runs a series of Playground competitions designed to help people learn and practice machine learning in a fun, low-stakes environment.

Each competition lasts one month and involves a simple predictive modeling task. This month’s challenge is about predicting the risk of road accidents, while past ones have covered topics like rainfall prediction, introvert vs. extrovert classification, and heart rate estimation.

This year, I’ve been active on Kaggle for the first time since joining the platform. These Playground challenges have been an amazing way to learn, experiment, and grow - both as a machine learning scientist and as a “Kaggler.” Here’s what I’ve learned from taking part.

The first competition is the one that you never forget.

For me, it was the March 2025 competition, and it involved predicting the probability of rain in a synthetic dataset based on weather measurements in Hong Kong.

This is the playground competition where I spent the most time. The one that made me suffer the most, tweaking code on the sofa after a long day of work, feature engineering every variable, and researching data. It is also the competition where I failed most spectacularly, and where my model overfit big time.

My first model was a simple regression, and, ironically, it performed better than many of the fancy models I built later. As the competition went on, I kept expanding my notebook with ideas from others - feature engineering, model tuning, even some ensembling.

Towards the end, panic set in. Everyone seemed to be improving their scores, so I decided to copy a public notebook and submit it instead of my own. That turned out to be a huge mistake - the final score was worse than the one I’d achieved myself.

One of the biggest lessons from these competitions is how much feature engineering matters.

In the Rainfall challenge, for example, converting dates into months and seasons helped, since rainfall patterns vary across the year. It’s also important to encode that day 1 and day 366 represent the same season — something raw numeric features can’t capture.

Adding lagged features (e.g., previous-day rainfall) also improved results, because weather patterns are often correlated across time.

In short: good feature engineering helps the model understand the world better — and can make a huge difference.

If you follow the Playground competitions long enough, you’ll notice a trend: AutoML frameworks like AutoGluon and H2O.ai dominate the leaderboards.

In many cases, you can run AutoGluon for nine hours and land near the top. Traditional machine learning problems - tabular data with well-defined features - are largely solved.

Of course, thoughtful feature engineering and robust validation still help fine-tune your performance. But when in doubt, AutoML will get you a strong baseline.

Another recurring theme is ensembling and stacking - combining multiple models to boost accuracy. Ensembling merges predictions from several models, often weighting them by performance. Stacking uses the outputs of one model as inputs for another.

In the final days of a competition, you’ll often see public notebooks full of creative ensemble blends - as participants scramble for that extra 0.0001 leaderboard improvement.

It’s funny to see how people on Kaggle try all sorts of things to win the challenge - even when people cheat, they do it in an intellectually stimulating way.

For example, in my first challenge, somebody managed to reverse-engineer the solution. In this challenge, people were allowed to make 5 submissions per day. For every submission, Kaggle would return a score, based on a portion of the test data (the final score of the challenge was based on the full test dataset).

Somebody managed to use these 5 submissions per day to reconstruct the expected values of the test dataset, using a technique called hitch-hiking. They submitted a variety of predictions, then used combinatorial statistics to infer the correct answer. It’s a way of cheating, but it is clever.

So, this was my experience with Playground competitions, this year.

I would recommend them to anyone to at least try a challenge - it’s a good way to practice Machine Learning skills, and it is interesting to see how different people approach a problem.