Finding the Bank Customers Who Can Safely Spend More
Aug 2026 · 5 min read
A bank wanted more credit card activity from the customers it already had. By clustering 1,938 customers into four personas, I found its best spenders are likely held back by their credit limits, not by their willingness to spend.
- Role
- Data Analyst, solo project
- Industry
- Banking
- Tools
- Python (pandas, scikit-learn), Google Colab, Google Slides
- Timeline
- 1 week in Aug 2026
- Customers clustered
- 1,938
- Power Users' avg spend
- Rp795M
- Their avg credit limit
- Rp98M
- Their avg DTI
- 0.22
#TL;DR
- Problem
- A bank wanted to increase credit card transaction activity among its existing customers within 6 months.
- What I did
- Cleaned and merged card and customer data, analyzed card performance and fraud, then clustered customers with k-means into four personas.
- What I found
- Power Users (8.6% of customers) spend the most, about Rp795M each in 6 months, but on a limit of only about Rp98M and with a healthy debt-to-income ratio (0.22). Their spending looks capped by their limit, which makes a limit increase the most promising low-risk lever, tested on a small group first.
In 6 months, Power Users spend 8 times their credit limit. VIPs have room they don't use.
#1.My role & scope
- Role: Data Analyst, solo project
- What I owned: data cleaning, merging, card performance and fraud analysis, feature selection, scaling, choosing k, modeling, persona naming, recommendation, slides
- Timeline: 1 week in Aug 2026
- Deliverables: Colab notebook + Google Slides deck with card performance insights, personas and recommendation
#2.Background
- Context: A bank with 2,000 credit card customers holding 5,531 cards across 4 card brands.
- Problem: Card transaction activity among existing customers wasn't growing, and the bank didn't know which customers could realistically be encouraged to spend more.
- Business question: How can the bank increase credit card transaction activity among its existing customers over the next 6 months, using insights from card performance and customer profiles?
- Decision maker: Performance Management Lead
#3.Method / Strategy
- Cleaning the card table:
- Removed 31 exact duplicate rows
- Removed 37 cards that were expired, had a zero limit or no limit as of 31 May 2025
- Filled blank fraud and non-fraud counts and amounts with 0, because blank meant no activity
- Fixed data types: "Rp" money text to numbers, date text to real dates
- Fixed card brand typos ("Visa " with a trailing space, "Jcb" to "JCB") to get 4 clean brands
- Result: 5,531 cards, 0 missing values
- Enriching the customer table: added age, a retired flag and debt-to-income ratio (DTI).
- Merging: combined both into one row per customer: 1,938 customers with recency, 6-month transaction count and amount, credit limit, number of cards, credit score, age and DTI.
- Card performance analysis: looked at profit, fraud and card brands before segmenting.
- Feature selection: clustered on 3 features: transaction count (how often they use the card), transaction amount (how much they spend) and credit limit (how much room the bank gives them). These directly answer "who can spend more?" Three features keep the clusters easy to read.
- Scaling: used StandardScaler. All three features were strongly right-skewed with high-value outliers, and MinMaxScaler would have squashed most customers into a tiny range.
- Choosing k: tested k = 2 to 6. The elbow bent around 3 to 4, and the silhouette score was highest at k = 4 (0.467), so I picked 4 clusters.
- Naming: gave each cluster a persona name based on its profile. I used names, not cluster numbers, because the numbers can change on every re-run.
#4.Metrics
| Metric | Why it matters |
|---|---|
| Transaction count (6 months, non-fraud) | How often customers actually use their card |
| Transaction amount (6 months, non-fraud) | How much value each customer brings |
| Credit limit | How much room the bank gives them to spend |
| Debt-to-income ratio (DTI) | Checks whether giving more credit is safe |
| Fraud rate and fraud share of profit | Shows how much risk eats into profit |
| Silhouette score | Checks that 4 clusters is a sensible split (0.467, the highest tested) |
Transaction count (6 months, non-fraud)
- Why it matters
- How often customers actually use their card
Transaction amount (6 months, non-fraud)
- Why it matters
- How much value each customer brings
Credit limit
- Why it matters
- How much room the bank gives them to spend
Debt-to-income ratio (DTI)
- Why it matters
- Checks whether giving more credit is safe
Fraud rate and fraud share of profit
- Why it matters
- Shows how much risk eats into profit
Silhouette score
- Why it matters
- Checks that 4 clusters is a sensible split (0.467, the highest tested)
#5.Solution
- Model: k-means, 4 clusters, on standardized transaction count, transaction amount and credit limit
- Persona profiles: size, behavior, average spend, limit, DTI and age for each persona
- Colab notebook: View notebook (opens in a new tab)
- Deck: View deck (opens in a new tab)
#6.Result / Output
The four personas
1,938 customers, averages over the last 6 months.
| Persona | Customers | Avg transactions | Avg spend | Avg limit | Avg DTI | Avg age | Recommended move |
|---|---|---|---|---|---|---|---|
| Inactive | 888 (45.8%) | 37 | Rp22M | Rp44M | 0.29 | 35 | Re-activation campaigns |
| Reliable Core | 786 (40.6%) | 501 | Rp326M | Rp74M | 0.25 | 52 | Keep them steady with light loyalty perks |
| Power Users | 167 (8.6%) | 1,037 | Rp795M | Rp98M | 0.22 | 54 | Raise credit limits |
| VIP High-Limit | 97 (5.0%) | 552 | Rp510M | Rp261M | 0.19 | 56 | Retention perks |
Inactive
- Customers
- 888 (45.8%)
- Avg transactions
- 37
- Avg spend
- Rp22M
- Avg limit
- Rp44M
- Avg DTI
- 0.29
- Avg age
- 35
- Recommended move
- Re-activation campaigns
Reliable Core
- Customers
- 786 (40.6%)
- Avg transactions
- 501
- Avg spend
- Rp326M
- Avg limit
- Rp74M
- Avg DTI
- 0.25
- Avg age
- 52
- Recommended move
- Keep them steady with light loyalty perks
Power Users
- Customers
- 167 (8.6%)
- Avg transactions
- 1,037
- Avg spend
- Rp795M
- Avg limit
- Rp98M
- Avg DTI
- 0.22
- Avg age
- 54
- Recommended move
- Raise credit limits
VIP High-Limit
- Customers
- 97 (5.0%)
- Avg transactions
- 552
- Avg spend
- Rp510M
- Avg limit
- Rp261M
- Avg DTI
- 0.19
- Avg age
- 56
- Recommended move
- Retention perks
Personas
Almost half the customers barely use their card.
888 of the 1,938 customers (45.8%) are Inactive. Power Users are a small group: 167 customers, 8.6%.
Show data
| Persona | Customers | Share |
|---|---|---|
| Inactive | 888 | 45.8% |
| Reliable Core | 786 | 40.6% |
| Power Users | 167 | 8.6% |
| VIP High-Limit | 97 | 5.0% |
Key insights
Insight 1 · Fraud is small per transaction but eats into profit
Fraud takes about 15% of the profit.
Net profit over 6 months was about Rp5.83B. Fraud takes about 15% of it, cutting the real margin from 1.5% to 1.28%.
Show data
| Step | Rp |
|---|---|
| Fee income (1.5% of Rp456.06B non-fraud spend) | Rp6,840,847,479 |
| Lost to fraud | −Rp1,013,378,300 |
| Net profit, 6 months | Rp5,827,469,179 |
Insight 2
Fraud is widespread but shallow
Only 0.22% of transaction value is fraud, but it touches about 10% of cards. Each hit is small (about Rp1.85M on average), so it's under control but worth watching.
Insight 3
Mastercard wins on volume, Amex on value per card
Mastercard drives nearly half (49.7%) of total transaction value, while Amex customers spend the most per card but are a small premium niche.
Insight 4 · Power Users look held back by their limits
In 6 months, Power Users spend 8 times their credit limit. VIPs have room they don't use.
They transact the most and spend the most, but their limits are small compared to their spending, and their DTI is healthy (0.22). That pattern points to the limit, not their willingness, as what caps their spending.
Show data
| Persona | Avg spend (Rp M) | Avg limit (Rp M) | Avg DTI |
|---|---|---|---|
| Inactive | 22 | 44 | 0.29 |
| Reliable Core | 326 | 74 | 0.25 |
| Power Users | 795 | 98 | 0.22 |
| VIP High-Limit | 510 | 261 | 0.19 |
Recommendation (from the deck)
Focus on the Power Users. They spend the most but have low credit limits compared to their spending and a healthy DTI (0.22), so a limit increase is the most promising low-risk lever. I'd start with a test: raise limits for a random half of the Power Users, keep the other half as a control, and compare spending and repayments before rolling it out. Re-activate the Inactive segment with campaigns, and keep VIPs with retention perks. Keep the Reliable Core steady with light loyalty perks. This grows transaction volume where it's safest, which matches the bank's goal of getting customers to spend more.
#7.Reflection
- Spend vs limit points to the cap, but doesn't prove it. Total spend over 6 months can exceed a limit when customers pay their balance down during the month. Utilization (balance ÷ limit) and payment history would confirm the limit is really what holds Power Users back, so I'd check those first. That's also why the recommendation starts with a test.
- Personas are a 6-month snapshot. I'd re-run the clustering on newer data to check that the four groups, and who sits in each, stay stable.
Data note: Built on a banking dataset: 2,000 users, 5,531 credit cards after cleaning, the last 6 months of transactions, snapshot as of 31 May 2025.
View deck (opens in a new tab)View notebook (opens in a new tab)