This is a synthetic dataset: a numeric target and 254 anonymous features, generated from one underlying data model for both the training and test rows. Nothing is known about what the features mean, so every modeling decision has to come from the data itself.
Your goal is to predict the target for the test rows. Predictions would be judged by mean squared error on the withheld targets. There is no answer key in the sandbox, so your own estimate of out-of-sample error is part of the answer.
The data
train— 5,000 rows:id,targetand featuresf_0…f_253. Some feature values are missing.test— 1,000 rows:idand the same features, withtargetwithheld.
Your task
- Explore the data and report what shapes your modeling choices: the distribution of the target, how much is missing and whether missingness looks random or informative, which features behave like categories rather than continuous values, and how strongly features relate to the target and to each other. Check whether
testlooks liketrain. - Build a model to predict
target. With 254 features and 5,000 rows, compare at least two approaches that make different trade-offs (for example, a regularized linear model and a tree ensemble, or using every feature versus a selected subset), and explain why you chose the final one. - Estimate the mean squared error you expect on
testand how confident you are in that number. Describe how you kept imputation, feature selection and tuning inside the validation loop, and how your error compares with a naive baseline such as predicting the mean. - Identify the features that matter most. How stable is that list across folds and across models, and what can and can’t you conclude about the underlying data model from it?
- If you had more time, or if the test rows turned out to come from a different distribution, what would you try next and why?
What to deliver
Predictions for all 1,000 rows of test, keyed by id, and a brief write-up (1–2 pages) for a modeling lead: the techniques you used, the most important features and your estimate of predictive performance. Include the code that reproduces the predictions.