Cheat sheet

Mercedes Hackathon — Cheat Sheet

268,000 rows of UK used-car listings, seven messy tables, and the case for domain knowledge in data cleaning. You don't clean what you don't recognise.

← Read the full projectUpdated June 2026
1

The brief

Mercedes hackathon, sponsored by IE. 24 hours.

  • Seven tables about used-car listings in the UK — vehicle specs, prices, locations, options, condition reports.
  • 268,000 rows in the main table after the first join.
  • No documentation. Column names like transmission, engine_size, electrics, trim — and you have to guess what each one means.

The goal was predicting price, but most of the day went into cleaning.

2

Domain reads

Things that look like garbage but aren't, once you know cars:

  • engine_size = 1.6 is litres, not arbitrary numbers. 0.0 means "electric, no combustion engine".
  • transmission strings: "Auto", "Automatic", "Semi-Auto", "CVT", "Tiptronic" → all need normalising to auto / manual / semi-auto.
  • electrics is a feature pack column ("heated seats, sat-nav"), not "electric vehicle".
  • fuel_type "Electric" means EV. "Hybrid" is partial. "Plug-in" is a third thing.
  • trim is a marketing-spec label ("Sport Edition", "AMG Line") — high-cardinality, but groupable.

None of this is in the data dictionary. All of it is obvious if you've ever looked at a car ad.

3

Cleaning moves

  • Normalise transmission strings with a domain-informed lookup table.
  • Encode engine_size: keep numeric, but add an is_electric flag for engine_size = 0.
  • Collapse fuel_type into petrol / diesel / hybrid / EV / other.
  • Group trim by brand into standard / premium / luxury tiers.
  • Drop electrics for the price model (it was a noisy multi-value field) OR explode it into individual binary feature flags.
  • Outlier prices — UK used-car prices follow a long right tail. Log-transform target.
4

Feature engineering

After cleaning:

  • age = current_year - manufacture_year
  • mileage_per_year = mileage / age (handles odd "1995 with 30k miles" cases)
  • price_per_litre = price / engine_size (premium markers)
  • One-hot for make, fuel_type, transmission tier
  • Target-encode model (high cardinality) inside CV folds
5

The model

  • Random Forest for a quick robust baseline.
  • XGBoost with target-encoding of model for the leaderboard submission.
  • Log-transformed target (log1p(price)) to handle skew. Exponentiate at predict time.

Honest CV with GroupKFold by make to prevent leakage — otherwise the model learns "BMW always £30k" and looks falsely strong on random splits.

6

What I learned

  • Domain knowledge is a force multiplier in data cleaning. Anyone can normalise strings; not everyone knows which strings mean the same thing.
  • Compound-meaning columns (electrics, trim) need a decision: drop, simplify, or explode.
  • Group-aware CV matters when a categorical feature dominates (here, make).
  • Log-transform skewed targets. R² improves, residuals stop fanning out, predictions stop crashing on luxury cars.
  • The hackathon judges noticed the cleaning narrative more than the model accuracy.