Predicting Restaurant Tips: A Regression Pipeline That Beat the Baseline 15–30%
A small, honest modelling problem done properly: feature engineering, cross-validation, and a baseline that had to be beaten before anything shipped.
- Client
- Hospitality analytics
- Industry
- Hospitality
15–30%
Lower error vs baseline
Cross-validated
Model selection
Reusable
EDA → train → evaluate
Deployable
Into analytics tooling
The problem
Tip income drives real operational decisions in hospitality — shift allocation, staffing levels, service model — and it was being estimated from rules of thumb. The question was whether the dining variables already being captured could predict it well enough to plan against.
The available signals were the ordinary ones: bill amount, party size, day, time of service, and smoker status in the source data.
How we approached it
Exploratory analysis first, and it earned its place: the relationship between bill and tip was strong but not linear across party sizes, and time of service interacted with day in a way a naive model would smear together.
Feature engineering followed from that — encoding categoricals properly rather than treating them as ordinal numbers, and deriving the interaction terms the exploration had exposed.
Model selection ran across candidates with cross-validation rather than a single train-test split, because on a dataset this size a favourable split can flatter a model by several points and a single split is how people convince themselves of results that do not hold.
Results
Prediction error fell 15–30% against a naive baseline — the baseline being the obvious percentage-of-bill assumption, which is what the model actually has to beat to be worth deploying.
The deliverable was a reusable Python pipeline from exploratory analysis through training to evaluation, packaged so it can be retrained on new data and dropped into lightweight analytics or decision-support tooling rather than living in a notebook.
What we'd flag
This is a small-data problem and it should be treated as one. The gain over baseline is real and modest, and anyone promising a transformative result from these variables is selling something.
Tipping behaviour is also culturally and regionally specific. A model fit on one venue's data should be refit rather than transplanted, and we would say the same about any similar behavioural target.
