results

Clustering of Insurance Customers – Results (PCA + KMeans)

Project: Unsupervised customer segmentation
Goal: Identify risk profiles and optimize pricing
Dataset: data_synthetic.csv Method: PCA 2D + KMeans (4 clusters)
Code: clustering_pipeline.py (with random_state=42)

---

Exploratory Data Analysis (EDA)

The following histograms and scatter plot were generated from the simulated claims data.
They provide the raw risk distribution before clustering and help validate the four customer profiles.

Plot	Description
	Claim Frequency (Sim_Frequency) – Poisson-like, heavily right-skewed. ~8k k customers have 0 claims, ~12 k have 1 claim, ~11 k have 2 claim then a rapid drop. A tiny tail reaches ≥ 10 claims.
	Expected Loss (Expected_Loss) – Highly skewed, with a sharp peak at 0 € (most safe drivers) and a long tail up to ~2 000 €.
	Total Severity (Sim_Total_Loss) – Similar shape to Expected Loss (log-normal behavior). Most policies have low total payout; a few extreme cases exceed 2 000 €.
	Frequency vs Total Severity, coloured by Risk Profile (0–3) – Strong positive correlation (higher frequency → higher total loss). Higher Risk Profile values systematically sit on the upper-right part of the cloud, confirming the simulation logic.

How the EDA connects to the clusters

Cluster	Frequency (mean)	Expected Loss (mean)	Interpretation from EDA
0 – Gold Safe	0.83	19.71 €	Dominates the left-most bars (0–1 claims) and the near-zero loss region.
1 – High Frequency Loser	5.91	425 €	Located in the far-right tail of the frequency histogram and the high-loss tail.
2 – Balanced Premium	2.94	147 €	Middle of the frequency distribution (2–4 claims) with moderate loss.
3 – Claim Heavy, Low Premium	2.55	47 €	Surprisingly low loss despite a high claim history (3.71) – explains the under-pricing (only 1.9 k premium).

These plots confirm the clustering: the four groups naturally separate along the two risk dimensions (frequency & severity) that dominate the PCA projection.

Key Findings

Cluster	Name	Risk	Avg. Profit	Action
0	Gold Safe	Low	+3,203 €	Retain
1	High Frequency Loser	Very High	+3,114 € (high cost)	Renegotiate / Exclude
2	Balanced Premium	Medium-High	+3,643 €	Maintain
3	Claim Heavy, Low Premium	High	+1,892 €	Increase premium

Key insight: Cluster 3 pays only 1.9k but has Claim History = 3.71 → underpriced by 60%+

---

PCA 2D Visualization

(Generated with random_state=42 – fully reproducible)

• X-axis (PCA1): Premium, Income, Expected Loss
• Y-axis (PCA2): Claim frequency, Claim History
• Cluster 0 (blue): bottom-left → low risk, high value
• Cluster 1 (orange): top → high cost
• Cluster 2 (green): center → balanced risk
• Cluster 3 (red): center-left → high claims, low premium

---

Cluster Summary (mean values)

Cluster	Sim_Freq	Exp_Loss	Premium	Income	Credit	Prev_Claims	Claim_Hist
0	0.83	19.71	3223	84.7k	671	1.53	1.07
1	5.91	425.28	3539	81.4k	671	2.23	3.84
2	2.94	147.51	3790	83.4k	665	2.10	2.62
3	2.55	47.91	1939	80.4k	683	1.27	3.71

---

Customer Profiles – Description

Cluster	Profile
0 – Gold Safe	Very few claims, high premium, good credit score. Ideal customer.
1 – High Frequency Loser	5.9 claims/year, expected loss 425€. Not sustainable.
2 – Balanced Premium	2.9 claims, pays 3.8k → best margin.
3 – Claim Heavy	Claim History very high (3.71), premium only 1.9k → underpriced!

---

Recommended Actions

Cluster	Strategy
0	Loyalty discount, upselling (home, life)
1	Increase premium to 5.5k+ or do not renew
2	Offer telematics → discount for safe driving
3	Increase premium to 3.2k+ (mandatory)

---

Technical Pipeline

Pipeline([
    ("prep", ColumnTransformer([
        ("num", StandardScaler(), num_features),
        ("cat", OneHotEncoder(drop="first"), cat_features)
    ])),
    ("pca", PCA(n_components=2, random_state=42)),
    ("kmeans", KMeans(n_clusters=4, random_state=42))
])