How is K-means clustering used to group assets for portfolio construction, and what are its limitations with financial return data?

K-means clustering partitions assets into K groups by minimizing within-cluster variance. Instead of relying on subjective sector labels (which mix unrelated companies), clustering uses statistical return patterns to reveal natural groupings.\n\nAlgorithm Steps:\n1. Choose K (number of clusters)\n2. Initialize K centroids randomly\n3. Assign each asset to the nearest centroid (Euclidean distance in feature space)\n4. Recompute centroids as the mean of assigned assets\n5. Repeat steps 3-4 until assignments stabilize\n\nWorked Example:\nLakefront Asset Management wants to diversify a 50-stock portfolio beyond traditional GICS sectors. They compute 8 features for each stock: 12-month return, 60-day volatility, beta, dividend yield, P/E ratio, debt-to-equity, revenue growth, and earnings stability.\n\nRunning K-means with K=6 produces:\n\n| Cluster | Profile | Stocks | Traditional Sectors Mixed |\n|---|---|---|---|\n| 1 | High-growth, high-vol | 8 | Tech + Biotech + Consumer Discretionary |\n| 2 | Stable dividend payers | 11 | Utilities + REITs + Consumer Staples |\n| 3 | Cyclical value | 9 | Industrials + Materials + Energy |\n| 4 | Defensive low-beta | 7 | Healthcare + Telecom + Utilities |\n| 5 | Leveraged growth | 8 | Financials + Tech + Real Estate |\n| 6 | Quality compounders | 7 | Tech + Healthcare + Consumer |\n\nClusters 2 and 4 both contain Utilities stocks — but cluster 2 groups them with REITs based on yield characteristics while cluster 4 groups others with Healthcare based on low-beta behavior. This captures economically meaningful distinctions that sector labels miss.\n\nChoosing K (Elbow Method):\nPlot within-cluster sum of squares (WCSS) against K. Lakefront tested K=3 to K=12:\n\n- K=3: WCSS=284 (too few, heterogeneous clusters)\n- K=6: WCSS=121 (elbow point, clear inflection)\n- K=10: WCSS=89 (marginal improvement, overly granular)\n\nK=6 provided the best balance between granularity and statistical stability.\n\nLimitations with Financial Data:\n- K-means assumes spherical clusters — financial return distributions are often elongated or asymmetric\n- Sensitive to outliers (extreme returns distort centroids)\n- Clusters may be unstable across time periods — quarterly reclustering often reassigns 20-30% of assets\n- Euclidean distance in high dimensions suffers from the curse of dimensionality\n\nAlternatives: Hierarchical clustering handles non-spherical shapes; DBSCAN automatically determines K and handles outliers.\n\nExplore clustering applications in our CFA Quantitative Methods question bank.