Pharma Company Segmentation (R)

This repository contains an end‑to‑end workflow to segment pharmaceutical companies using unsupervised learning in R. The goal is to turn raw fundamentals into actionable business segments that help guide investment or product strategy.

Summary of results: Using hierarchical clustering and K‑Means on 21 companies × 9 financial metrics, identified 3 clear segments with 7/9/5 firms. Cross‑method agreement was ~81% (17/21), and K‑Means explained ~45% of between‑company variation (between_SS / total_SS = 44.7%). A 4×4 Self‑Organizing Map (SOM) was used to visualize topology and neighborhood distances.

---

Table of Contents

1. Project Structure
2. Dataset
3. Environment
4. How to Run
5. Methodology
6. Key Findings
7. Visualizations
8. Business Insights & Strategy
9. Limitations
10. Next Steps
11. Citations

---

Project Structure

.
├─ data/
│  └─ Pharmaceuticals.csv                # input dataset (21 rows x 14 columns)
├─ scripts/
│  └─ analysis.R                         # main R script (see snippets below)
└─ README.md

If you don’t use this exact structure, just update the file paths in your R session (e.g., read.csv("data/Pharmaceuticals.csv")).

---

Dataset

Rows: 21 companies
Core features used (9): Market_Cap, Beta, PE_Ratio, ROE, ROA, Asset_Turnover, Leverage, Rev_Growth, Net_Profit_Margin
Other columns (excluded from clustering): identifiers and metadata such as Symbol, Name, Median_Recommendation, Location, Exchange.

Example schema (first few):

• Market_Cap (numeric, $B)
• Beta (numeric)
• PE_Ratio (numeric)
• ROE (%)
• ROA (%)
• Asset_Turnover
• Leverage
• Rev_Growth (%)
• Net_Profit_Margin (%)

---

Environment

• R: 4.2.2 (aarch64-apple-darwin20)
• Key packages:
  • Base/Recommended: stats, graphics, utils
  • Clustering: cluster
  • SOM: kohonen (and its dependency Rcpp)

Install packages as needed:

install.packages(c("cluster", "kohonen"))

---

How to Run

Minimal end‑to‑end script (place as scripts/analysis.R or run in an R console):

# 1) Load data
data <- read.csv("data/Pharmaceuticals.csv")

# 2) Select numeric features and scale
z <- data[, c("Market_Cap","Beta","PE_Ratio","ROE","ROA",
              "Asset_Turnover","Leverage","Rev_Growth","Net_Profit_Margin")]
X <- scale(z)  # z-score

# 3) Distance + hierarchical clustering (complete & average)
dist_mat <- dist(X)
hc_complete <- hclust(dist_mat, method = "complete")
hc_average  <- hclust(dist_mat, method = "average")

# 4) Choose K via elbow (WSS) and run K-Means (k = 3)
wss <- (nrow(X) - 1) * sum(apply(X, 2, var))
for (k in 2:20) wss[k] <- sum(kmeans(X, centers = k, nstart = 25)$withinss)

set.seed(123)
km <- kmeans(X, centers = 3, nstart = 50)

# 5) Cross-method agreement (complete vs average)
m_complete <- cutree(hc_complete, 3)
m_average  <- cutree(hc_average,  3)
agreement  <- mean(m_complete == m_average)  # ~0.81 on this data

# 6) Summary
list(
  kmeans_sizes = km$size,                           # 7, 9, 5 (order may vary)
  between_ratio = km$betweenss / km$totss,         # ~0.447
  cross_method_agreement = agreement
)

(Optional) Self‑Organizing Map visualization:

library(kohonen)
set.seed(222)
g <- somgrid(xdim = 4, ydim = 4, topo = "rectangular")
map <- som(X, grid = g, alpha = c(0.05, 0.01), radius = 1)

# Basic plots
plot(map)                               # default overview
plot(map, type = "dist.neighbours")     # U-matrix (neighbour distances)
plot(map, type = "count")               # data density per node
plot(map, type = "codes")               # codebook vectors

---

Methodology

1. EDA & Preprocessing

   • Inspected distributions and pairwise relationships.
   • Standardized all 9 numeric features (z‑scores).

2. Clustering

   • Hierarchical clustering with Euclidean distance, complete & average linkage.
   • K‑Means (k=3), selected via qualitative elbow (WSS) and interpretability.
   • Stability check: Compared cluster membership between complete vs. average linkage → ~81% agreement (17/21 firms).

3. Evaluation

   • K‑Means between_SS / total_SS = 44.7%.
   • Silhouette inspected qualitatively (not reported here as a single number).
   • SOM used to inspect topological separation and neighborhood structure.

---

Key Findings

K‑Means (k=3) cluster sizes: 7, 9, 5 (order varies by seed, but stable).
Between-cluster variance explained: ~44.7%.
Cross-method agreement (HC complete vs average): ~81% (17/21).

Cluster profiles (means in original units, from hierarchical characterization):

Segment	Market Cap	Beta	P/E	ROE (%)	ROA (%)	Asset Turnover	Leverage	Rev Growth (%)	Net Margin (%)
A (High-profit incumbents)	97.11	0.43	20.95	35.70	14.95	0.80	0.33	10.16	20.17
B (High-price / low-profit)	26.91	0.64	55.63	10.10	4.20	0.70	0.32	6.99	5.13
C (Smaller, high-growth, more debt)	8.82	0.62	19.61	16.96	6.24	0.54	1.11	21.14	13.19

(Dataset means for reference: Market Cap 57.65, P/E 25.46, ROE 25.80, ROA 10.51, Leverage 0.59, Rev Growth 13.37, Net Margin 15.70.)

---

Visualizations

Core visuals used in the analysis:

• Pairs plot / Scatter plots (e.g., Revenue Growth vs Net Margin with labels)
• Dendrograms (complete & average linkage)
• Elbow plot (WSS vs. k)
• Silhouette plot (qualitative check)
• SOM U‑Matrix / Counts / Codes to inspect topology

Example snippets:

# Pairs plot (quick EDA)
pairs(z)

# Labeled scatter
plot(Rev_Growth ~ Net_Profit_Margin, data = data)
with(data, text(Rev_Growth ~ Net_Profit_Margin, labels = Symbol, pos = 2, cex = 0.7))

# Dendrograms
plot(hc_complete, hang = -1, labels = data$Symbol, main = "HC (Complete)")
plot(hc_average,  hang = -1, labels = data$Symbol, main = "HC (Average)")

# Elbow plot
plot(1:20, wss, type = "b", xlab = "Number of Clusters (k)", ylab = "Within-Cluster Sum of Squares")

# Silhouette (hierarchical example)
library(cluster)
plot(silhouette(cutree(hc_complete, 3), dist_mat))

---

Business Insights & Strategy

• Segment A — Profitable incumbents: Strong profitability and returns (ROE ~35.7%, margin ~20.2%), moderate growth, lower leverage.
  Strategy: Core allocation / priority partners; focus on scale features and reliability.

• Segment B — High‑price, low‑profit: Elevated P/E (~55.6×) with weak margins (~5.1%).
  Strategy: Watchlist until margin expansion; alerts around profitability milestones.

• Segment C — Small, fast‑growing, more debt: Higher growth (~21.1%) with higher leverage (~1.11).
  Strategy: Selective pilots; monitor balance‑sheet risk; emphasize ROI tracking.

---

Limitations

• Small N (21): Findings are directional; statistical confidence is limited.
• Single snapshot: No time series; clusters may shift with new data.
• Feature set: Fundamental metrics only; adding pipeline/therapeutic area/region could refine segments.

---

Next Steps

• Add time‑series features (TTM, YoY deltas) and reassess stability.
• Try Gaussian Mixture Models and PAM; report silhouette and ARI.
• Enrich with qualitative variables (therapeutic focus, R&D spend, patent cliffs).
• Package the workflow as an R Markdown report for reproducibility.

---

Citations

• R Core Team (2022). R: A Language and Environment for Statistical Computing.
• Wehrens, R., & Buydens, L. M. C. (2007). Self‑ and Super‑Organizing Maps in R: The kohonen Package. Journal of Statistical Software.

---

Author

Pranav Manjunath Bhat — Arizona State University