ML Pipeline — Resale Car Price Prediction

📦

Data collection

CarDekho dataset — 15,411 used car records scraped from cardekho.com with 13 features, target variable: selling_price (INR)

cardekho_imputated.csv 15,411 rows 13 features

🧹

Data cleaning

Pre-imputed dataset — zero null values found · Drop car_name and brand (redundant with model column)

No nulls Drop car_name Drop brand

🔍

Exploratory data analysis

Price heavily right-skewed (₹2–15L range) · max_power strongest predictor (0.75) · engine second (0.59) · vehicle_age & km_driven negatively correlated

Price distribution Correlation heatmap sns / plotly

⚙

Feature identification

7 numerical features · 4 categorical features · Label Encoding applied to model column (120 unique car models)

7 numerical 4 categorical LabelEncoder on model

🔄

Preprocessing — ColumnTransformer

80/20 train-test split (random_state=42) · OneHotEncoder (drop=first) on seller_type, fuel_type, transmission_type · StandardScaler on numerical features

OneHotEncoder StandardScaler 80/20 split

MODEL COMPARISON

Lower performers

Linear Regression · Lasso · Ridge — Test R² 0.66 · KNN — Test R² 0.92 · Decision Tree — overfit (Train 0.99 → Test 0.88)

not selected

Random Forest Regressor

Best test R² 0.93 · Minimal overfitting · Strongest generalisation across all metrics

selected ✓

🎯

Hyperparameter tuning

RandomizedSearchCV · n_iter=100 · cv=3 · Best: n_estimators=1000, max_features=7, max_depth=None, min_samples_split=2

RandomizedSearchCV n_iter=100 cv=3

📊

Evaluation

RMSE ₹2,12,015 · MAE ₹98,050 · R² 0.9403 on test set · Top predictors: max_power, engine, vehicle_age, km_driven

R² Score RMSE / MAE Feature importance

💾

Model export

Saved with joblib · car_price_predictor.pkl + preprocessor.pkl · Sample prediction: Ford Ecosport (6yr) → ₹6,08,500

joblib .pkl deploy ready