Insurance Charges Prediction

An end-to-end regression pipeline predicting individual medical insurance charges from demographic and lifestyle features — covering cleaning, encoding, domain-driven feature creation, scaling, and model evaluation.

Final model: Linear Regression — Adjusted R² = 0.796 on a held-out 20% test set.

---

Problem

Insurance pricing depends on which personal attributes genuinely drive cost. The goal here was not just to maximise accuracy, but to build a pipeline where every feature transformation is deliberate and every retained variable is justified.

---

Dataset

1,338 policyholder records (1,337 after de-duplication), 7 columns:

Feature	Type	Description
age	numeric	Policyholder age (18–64)
sex	categorical	Gender
bmi	numeric	Body mass index (15.96–53.13)
children	numeric	Number of dependents (0–5)
smoker	categorical	Smoking status
region	categorical	US region (4 levels)
charges	numeric	Target — annual medical cost billed

Mean charge: $13,270 · Std: $12,110 — a strongly right-skewed target, confirmed in the distribution plots.

---

Pipeline

1. Exploratory analysis Distribution plots (histogram + KDE) for all numeric features, count plot for dependents, null-value audit (dataset is complete — zero nulls).

2. Cleaning Removed 1 exact duplicate row (1,338 → 1,337).

3. Encoding

• Binary label encoding for sex → is_female and smoker → is_smoker, with explicit renaming so the column name states what 1 means
• One-hot encoding for region with drop_first=True to avoid the dummy variable trap

4. Feature engineering Created a bmi_category feature from clinical thresholds rather than arbitrary bins — underweight (<18.5), normal (18.5–24.9), overweight (25–29.9), obese (30+) — then one-hot encoded it. This lets the model capture the non-linear jump in cost at the obesity threshold that raw BMI alone would smooth over.

5. Scaling StandardScaler applied to age, bmi, and children so no single feature dominates by magnitude.

6. Modelling 80/20 train-test split (random_state=42), Linear Regression baseline, evaluated with Adjusted R² rather than plain R² — the adjusted metric penalises the added dummy columns and gives an honest read on whether the engineered features earned their place.

---

Result

Metric	Value
Adjusted R² (test)	0.7963
Model	Linear Regression (OLS)
Test size	268 records

The model explains roughly 80% of variance in insurance charges. Smoking status is the dominant cost driver — the separation is visible in the exploratory plots well before modelling.

---

Tech Stack

Python · Pandas · NumPy · Scikit-learn · Matplotlib · Seaborn

---

Run Locally

git clone https://github.com/Himanshi252005/Insurance-Charge-Prediction.git
cd Insurance-Charge-Prediction
pip install pandas numpy scikit-learn matplotlib seaborn jupyter
jupyter notebook insurance.ipynb

---

Next Steps

• Compare against Ridge, Random Forest and Gradient Boosting to test whether the relationship is genuinely linear
• Add RMSE and MAE alongside R² for an error figure in dollars
• Formally quantify the smoker effect with a hypothesis test rather than relying on visual separation

---

Author: Himanshi Rathore · LinkedIn · GitHub