Nivaan Homes, a listings platform, has 12,000 rental listings with size, location, age, floor and days on market. Design a model to predict monthly rent: which target, which features, how do you validate it, and which feature is a leakage trap?

Predict the log of monthly rent from size, locality, age and floor, validate with folds grouped by building plus a final test on the most recent months, and leave days on market out. Log rent makes errors proportional, so a Rs 5,000 miss counts for more on a cheap flat than an expensive one. Days on market is recorded after the rent is set and rises when a flat is overpriced, so using it leaks the answer.

Predicting raw rupees lets the expensive flats dominate. A Rs 5,000 miss on a Rs 20,000 flat is 25% out; the same miss on a Rs 2 lakh flat is 2.5%. Predicting the logarithm of rent makes the model care about percentage errors, which is what a landlord feels, and it tames the long tail of very expensive flats. Size enters as log square feet for the same reason: rent roughly scales with size, so on the log scale the relationship is close to a straight line. One detail to state: turning a log prediction back into rupees gives something close to the median rent, not the mean, which is usually what a suggested rent should be anyway.

Think of an examiner who can see how long each student took before marking the paper: the time taken tells him something about the answer, but he cannot use it before the exam is written. The test for every feature is whether it exists at the moment the prediction is made; days on market fails, because it is recorded when the listing closes, after the rent is decided. It also carries the answer directly: a flat listed above the market rent sits unlet for longer. Including it is target leakageWhen a model is given information that would not be available at prediction time, often information caused by the thing being predicted, so it looks accurate in testing and fails in use.. Size, locality, age and floor are all known the day the flat is listed. Locality needs care: encode it using average rents computed only from the training folds, never from the whole data set, or the leak comes back by another door.

A random split of rows is the natural first idea and it flatters the model. The same building, and often the same flat, appears many times, so a random split puts near copies in both training and test, and the model is rewarded for remembering buildings rather than pricing flats. Group the folds by building so that each building sits entirely in training or entirely in test. Then hold back the most recent two months as a final test, because rents drift over time and the model will always be used on the next listing, never on a past one. Report the median absolute percentage error, which a landlord understands.

Read the table from the bottom. The 6% model is the one a careless candidate reports, and it would fail the day it goes live because the feature does not exist yet. An honest 11% that holds up on new buildings is worth more than a 6% that cannot be run, and the interviewer is checking whether you can tell them apart. Say the limitation too: listings carry asking rents, and if the data recorded the rent actually agreed, that would be the better target.