Autotrader Machine Learning Predictive Modelling

Overview

This project implements an end-to-end Machine Learning pipeline to predict the selling price of used cars. The target variable is price, a continuous numerical value in GBP, making this a regression task. The primary goal is to accurately model the non-linear depreciation of vehicles and the complex dynamics of the used car market.

Data Exploration and Processing

Thorough exploratory data analysis and processing were conducted to prepare the data for machine learning:

• Handling Skewness: The car prices and mileage displayed heavy positive right-skewness. A log transformation (np.log1p) was applied to the target variable to compress the scale and allow the models to focus on percentage errors rather than absolute differences.

• Feature Engineering: A vehicle_age feature was derived from the registration year to create a more intuitive, ratio-scale metric for vehicle depreciation. Erroneous registration years (e.g., years < 1900 or > 2025) were dropped to prevent the model from learning irrational rules from corrupted data entries.

• Imputation: Missing continuous values (like mileage) were filled using the median due to skewness, while missing categorical features utilized the mode.

• Categorical Encoding: A hybrid encoding strategy was used to handle varying cardinality. Low-cardinality features utilized One-Hot Encoding. High-cardinality features (like standard_model) utilized Target Encoding (mean encoding) to prevent the "curse of dimensionality" and maintain a compact feature space.

• Scaling: Standard Scaling (Z-score normalization) was applied to continuous numerical features (mileage, vehicle_age) to ensure algorithms relying on Euclidean distances (like kNN) performed optimally.

• Data Splitting: The dataset was split 80/20 into training and test sets using a fixed random seed. All preprocessing was fitted on the training set only and then applied to the test set to prevent data leakage and ensure realistic evaluation.

Modelling Strategy

All data processing steps were encapsulated within Scikit-Learn Pipeline objects to strictly prevent data leakage. Four different regression models were built, trained, and optimized using GridSearchCV with k-Fold Cross-Validation:

• Linear Regression: Served as a simple baseline model to judge whether more complex models add value.
• K-Nearest Neighbours (kNN): A non-parametric model tuned for the optimal number of neighbors (k=7) and uniform weight functions.
• Decision Tree Regressor: Captured non-linear patterns, optimized with a maximum depth of 15 and a min_samples_leaf of 4 to prevent overfitting.
• Random Forest Regressor: An ensemble method used to reduce the high variance of individual decision trees and improve generalization. The best parameters were a maximum depth of 20, minimum samples split of 5, and 100 estimators.

Results and Evaluation

The models were evaluated on the unseen test set using R-Squared (R²), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE).

Model	R-Squared (R²)	RMSE (£)	MAE (£)	MAPE (%)
Random Forest	0.8997	3780.01	1963.86	15.58%
Decision Tree	0.8541	4558.10	2259.61	17.07%
k-Nearest Neighbours	0.8198	5064.08	2884.28	23.19%
Linear Regression	0.7711	5705.80	3450.48	35.47%

Key Findings:

• The Random Forest Regressor is unequivocally the best-performing model, capable of explaining 90% of the price variability.
• The model's MAPE indicates that, on average, the automated prediction is within 15.58% of the actual list price, which is highly accurate for the unpredictable used car market.
• Residual analysis demonstrated that the ensemble Random Forest approach successfully smoothed out individual tree errors and maintained consistent accuracy across different segments of the market.

Feature Importance

Analysis of the Random Forest model identified the top three drivers of car prices:

1. Vehicle Age (Year of Registration): The dominant predictor that establishes the baseline value tier for a vehicle.
2. Mileage: The primary depreciation factor that modifies the baseline value set by the car's age.
3. Standard Model: The specific brand and model type, which is critical for distinguishing between market segments (e.g., Luxury vs. Economy).