AutoTrader Price Regression

Predicting second‑hand car prices with classic tabular ML.
Data  402 006 rows · 12 columns (target =price).
Models  Linear Regression · Random Forest · Gradient Boosting · Voting Ensemble.

---

Table of contents

1. Project motivation
2. Data
3. Quick start
4. Notebook & code guide
5. Results at a glance
6. Model interpretation
7. Directory layout

---

Project motivation

Buying a used car is a price‑sensitive decision.
The goal is to build transparent, reproducible baselines that predict price given mileage, age, fuel type and a handful of categorical descriptors.
Grades in the coursework are not the focus; clean code and solid discussion are.

---

Data

• Source AutoTrader extract supplied by Manchester Metropolitan University.
  The licence prohibits redistribution, so the CSV is not committed to this repository.
• Rows 402 006 Columns 12 (all except price used as predictors).
• Cleaning steps
  • Trim outliers in mileage & price via 1.5 × IQR.
  • Drop cars registered before 1975.
  • Mode‑impute gaps in fuel_type, body_type, standard_colour.
• Engineered features
  • vehicle_age = 2024 – year_of_registration
  • mileage_to_age_ratio = mileage / vehicle_age

See notebooks/01_autotrader_walkthrough.ipynb for the exact code.

---

Quick start

# clone repo
git clone https://github.com/hamzahassan9320/autotrader-price-regression.git
cd autotrader-price-regression

# place the CSV in the expected location
mkdir -p data
cp /path/to/Adverts.csv data/

# set up environment
conda create -n autotrader-price python=3.10
conda activate autotrader-price
pip install -r requirements.txt

# full pipeline
python -m src.train --csv data/Adverts.csv

# run the Streamlit app locally
streamlit run app.py

Tested with Python 3.10 and scikit‑learn 1.3.2.

4 · Notebook & code guide

file	purpose
notebooks/01_autotrader_walkthrough.ipynb	data snapshot, EDA, demos
src/data.py	load + cleanse CSV
src/features.py	feature engineering & preprocessing
src/models.py	pipelines · param grids · grid‑search helper
src/train.py	one‑shot CLI training run; saves models & plots
src/visualise.py	regenerates figures in docs/images/

---

5 · Results at a glance

model	CV MAE ↓	Test R²
Linear Regression	1 642 ± 394	0.79
Random Forest	1 831 ± 51	0.90
Gradient Boosting	2 742 ± 95	0.87
Voting Ensemble	1 894 ± 44	0.89

Random Forest brings the best MAE and R² without visible over‑fit.

---

6 · Model interpretation

• SHAP beeswarm → global drivers (top features: vehicle_age, mileage).
• SHAP waterfall → why a single advert (row 39) is priced ± £9 k.
• Partial dependence → price drops near‑linearly with age; flattening after ~15 yrs hints at a market floor.

All figures live in docs/images/, regenerated by src/visualise.py.

---

7 · Directory layout

.
├── data/                # <empty> – you add Adverts.csv locally
├── notebooks/           # single exploratory notebook
├── src/                 # reusable code
├── configs/             # YAML config(s)
├── docs/images/         # plots for README
└── requirements.txt