What it means
The starting point is a table of records with numeric attributes: customers with their spend, order frequency and tenure, for example. The method measures how far apart any two records are on those attributes and then builds groups of records that sit close together.
The most common approach in business analytics is k-means, where you choose how many clusters you want, the software places that many centre points at random, assigns every record to its nearest centre and then moves each centre to the average of its members. That loop repeats until the groups stop changing, which usually takes only a few passes.
Scale matters enormously, because a distance calculation treats a difference of 1,000 in dollars as far bigger than a difference of 10 in order count even if the second is more meaningful. Analysts therefore normalise each attribute first, typically by converting it to a standard score so every attribute contributes on equal footing.
Finance teams use clustering for customer segmentation that drives pricing and credit policy, for grouping stores or branches into comparable peer sets, and for anomaly detection where a small, odd cluster of transactions is worth a closer look. It also appears in portfolio work, grouping holdings that behave similarly so that concentration risk becomes visible.
The main nuance is that clustering describes rather than predicts, and it will always produce groups even when none genuinely exist. Choosing the number of clusters is a judgement call informed by tools such as an elbow chart, and the real test is whether the resulting groups are stable over time and mean something to the people who must act on them.
In practice
Real-world examples.
Example
A retail bank clusters 400,000 current account holders on balance, transaction count and overdraft use, and finds four natural groups. The smallest group, high balance and low activity, becomes the target list for the wealth advisory team.
Example
An equity research team clusters 300 listed companies on margin, growth rate and capital intensity rather than by official sector labels. Two software names land in the same cluster as several industrial services firms, which changes the peer group used for valuation multiples.
Example
A payments processor runs clustering on merchant behaviour and isolates a tiny cluster with unusually high average ticket size and refund rates. The risk team reviews those accounts and finds two merchants breaching their contract terms.
Formula
Calculation
Straight-line (Euclidean) distance between a record and a cluster centre with two attributes: distance = square root of ((x1 - x2) squared + (y1 - y2) squared). Each record joins the cluster whose centre is nearest.
Take a subscription business that has scaled two attributes onto comparable ranges: average monthly spend in hundreds of dollars, and orders placed per year. One customer sits at spend 12 (that is $1,200 a month) and 40 orders. Cluster centre A sits at spend 9 and 36 orders, and cluster centre B sits at spend 12 and 52 orders.
Distance to centre A = square root of ((12 - 9) squared + (40 - 36) squared) = square root of (9 + 16) = square root of 25 = 5.0. Distance to centre B = square root of ((12 - 12) squared + (40 - 52) squared) = square root of (0 + 144) = 12.0. Because 5.0 is less than 12.0, the customer is assigned to cluster A, the moderate-spend, moderate-frequency group, rather than the high-frequency group.Case study
Seen in the real world.
This is an illustrative, fictional scenario. Marrowfield Coffee Roasters, a wholesale roaster, had been treating all 1,800 of its cafe customers with one price list and one delivery schedule. Its finance director suspected the flat approach was subsidising some accounts at the expense of others, but had no evidence.
The analytics team normalised four attributes for each account, weekly kilograms ordered, order frequency, average days to pay and support tickets raised, then ran a k-means clustering with four groups. One cluster of about 200 accounts ordered small volumes very frequently, paid late and generated most of the support load, while another cluster of 150 accounts ordered large volumes on a predictable fortnightly cycle and paid on time.
Marrowfield introduced a minimum order value for the first group and a small volume rebate for the second. Because the exercise was descriptive rather than predictive, the team re-ran the clustering every quarter to check the groups were still stable, and treated the segments as a management tool rather than a permanent truth about its customers.
Watch out
Common mistakes.
- Running clustering on raw values without normalising them first, which lets whichever attribute has the largest numbers dominate every distance calculation.
- Believing the clusters are real categories that exist in nature, when the method will always split data into however many groups you asked for.
- Choosing the number of clusters purely by which chart looks neatest, instead of checking whether the groups are stable across time periods and useful to the business.
Questions
People also ask.
How is cluster analysis different from regression?
Regression predicts a specific outcome you already care about, while clustering has no outcome variable and simply groups records that resemble one another.
How many clusters should I ask for?
Most business segmentations land between three and eight, because fewer than three rarely tells you anything new and more than eight becomes impossible to act on.
Can cluster analysis handle non-numeric data such as country or product type?
Yes, but only with methods designed for it, such as k-modes or converting categories into indicator variables, because straight-line distance is not meaningful for text labels.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%