This week’s assignment was hands-down one of the most practical and eye-opening tasks so far: build and compare multiple regression models to predict used car prices from a clean 200-record dataset.

Quick Project Overview

Goal → Create a reliable pricing tool for an automotive business Dataset → 200 vehicles with features like Year, Mileage, Horsepower, Engine Size, Owners, Brand, Fuel Type, Transmission, etc.

Models I Built & Compared

• Multiple Linear Regression

• Polynomial Regression (Degree 3)

• SVR (with and without scaling)

• Decision Tree (default + tuned with GridSearchCV)

• Random Forest

• Gradient Boosting

I also did:

• Full EDA + preprocessing (One-Hot Encoding + StandardScaler)

• 5-fold cross-validation

• Outlier detection (IQR method)

• Model export with joblib for “production” simulation

The Big Surprise

• Year → 39% importance (newer = much more valuable)

• Mileage & Previous Owners → ~13% each (wear & tear kills value)

• Horsepower + Engine Size → ~12% combined

• Brand prestige (BMW, Mercedes, etc.) → secondary effect

Business-Ready Recommendations I Wrote in the Report

• Focus inventory on 2020+ models, low mileage, ≤1 owner

• List cars at Predicted Price + $1,500 buffer (covers typical error)

• Collect more data on color, interior condition, location to capture remaining ~6–7% variance

• Retrain every 3–6 months (especially with rising EV/hybrid values)

Key Takeaway from Week 15

Simpler is often better — especially with small-to-medium datasets. Linear Regression was interpretable, fast, stable across cross-validation folds, and beat every fancy ensemble. Trees and boosting added complexity without real gain here.

Grateful to the DataraFlow team (shoutout to Sir Winner and everyone keeping us on track 🙌). Next weeks are getting even more intense — stay tuned!

Would love your thoughts: When was the last time a simple linear model surprised you by outperforming everything else?

#DataScience #MachineLearning #Regression #Python

#DataraFlow #UsedCars #AjiboyeLearnsData