Transaction Data Classification
Introduction to Classification Models
Once transaction data has been cleaned and structured, the next step is to classify transactions into meaningful categories — such as food, transportation, entertainment, or utilities. Classification is a core task in financial analytics that helps banks and fintechs understand user behavior, detect fraud, and generate personalized insights.
Supervised vs. Unsupervised Learning
There are two main approaches to transaction classification:
- Supervised Learning
In supervised learning, the model is trained on labeled data, where each transaction is already assigned to a known category. The algorithm learns patterns from these examples and applies them to classify new, unseen transactions.
Example: A model trained on labeled transaction data (“Starbucks → Food & Beverage”) can automatically categorize similar new entries. - Unsupervised Learning
When labeled data isn’t available, unsupervised algorithms group transactions based on similarity patterns in spending amount, merchant names, or descriptions.
Example: Clustering techniques can discover new or emerging spending patterns — such as grouping payments related to ride-hailing apps under “transportation.”
Popular Machine Learning Models for Classification
Several powerful models are commonly used in transaction classification tasks:
- Random Forest: An ensemble of decision trees that performs well with structured numerical and categorical data. It’s often used as a baseline model for tabular financial data.
- Support Vector Machine (SVM): Effective for separating spending patterns with clear decision boundaries, such as distinguishing between online and in-store transactions.
- XGBoost: A gradient boosting algorithm known for its high accuracy and efficiency, especially in large-scale financial datasets.
- Neural Networks: Deep learning architectures that capture complex, nonlinear relationships — particularly useful when combining text and numerical features.
Each model has its strengths, and the best choice depends on data volume, complexity, and labeling quality.
Using NLP for Transaction Descriptions
Natural Language Processing (NLP) plays a vital role in enhancing classification accuracy. Many transaction records include free-text descriptions like merchant names, product details, or payment notes.
By applying tokenization, embedding, and Named Entity Recognition (NER), NLP techniques can extract meaningful features (e.g., “Uber,” “Starbucks,” “Electric Bill”) that help models correctly assign categories.
Data Splitting in Machine Learning
Training Set
Used to train the model by adjusting its parameters.
Makes up 60–80% of the dataset.
The model learns key patterns and relationships from labeled data.
Validation Set
Used to fine-tune the model and prevent overfitting.
Represents about 10–20% of the dataset.
Helps optimize hyperparameters to improve generalization.
Test Set
Used to evaluate the model’s final performance on unseen data.
Also about 10–20% of the dataset.
Applied only after final model selection to measure real-world accuracy.
Validated Indicators
When evaluating a machine learning classification model, it’s essential to measure how accurately and effectively the model performs across different scenarios.
To do this, we rely on several key performance metrics, each capturing a different aspect of model behavior.
The most commonly used metrics include:
AUC (Area Under the Curve) – Quantifies the overall ability of the model to distinguish between classes; a higher AUC indicates better performance.
Accuracy – Measures the overall proportion of correctly classified transactions.
Precision – Indicates how many of the transactions predicted as “positive” are actually correct.
Recall (Sensitivity) – Measures how well the model identifies all actual positive cases.
F1-Score – A balanced metric that combines Precision and Recall into a single measure.
ROC (Receiver Operating Characteristic) Curve – Visualizes the trade-off between true positive and false positive rates.
TP = True Positives，TN = True Negatives ，FP = False Positives ，FN = False Negatives
Credit Scoring with Machine Learning
Constructing a Credit Scoring Model Using Transaction Data
Building a credit scoring model with machine learning involves transforming raw financial data into predictive insights about borrower reliability. Below is a step-by-step outline of how this process works in practice.
Step 1: Preprocess the Data
Before modeling, transaction data must be cleaned and standardized to ensure consistency and accuracy.
- Handle missing values through imputation or removal.
- Normalize transaction amounts to reduce scale imbalance between small and large spenders.
- Encode categorical features such as merchant types, payment methods, and user demographics into numeric form suitable for machine learning algorithms.
Step 2: Train and Tune Models
Once the data is prepared, various machine learning models are trained and optimized.
- Start with Logistic Regression for interpretability and baseline performance.
- Experiment with XGBoost, which captures complex nonlinear relationships and performs well on tabular financial data.
- Test Neural Networks for deeper feature learning, especially when combining numeric and text-based transaction attributes.
- Use cross-validation and hyperparameter tuning to improve model robustness.
Step 3: Evaluate Performance
Model performance is assessed using multiple evaluation metrics to ensure fairness and predictive power.
- AUC-ROC evaluates the model’s ability to distinguish between good and bad credit risks.
- F1-Score balances precision and recall for imbalanced datasets.
- Accuracy provides an overall success rate but should not be used alone for skewed classes.
Step 4: Deploy and Monitor
After selecting the best-performing model, it’s deployed into production within banking or fintech systems.
- Integrate the model into credit decision pipelines or loan approval systems.
- Continuously monitor performance to detect data drift or changing borrower behavior.
- Retrain periodically using updated transaction data to maintain accuracy and compliance.
Case Study & Hands-on Application
Step 1: Understanding the Data
We use bank transaction data that contains historical customer behavior. The dataset includes:
| Feature Name | Description |
|---|---|
| Account Id | Unique identifier for each customer |
| Transaction Date | Date of the transaction |
| Transaction Year, Transaction Month | Time-based transaction features |
| Transaction Type | Credit (income) or Debit (expense) |
| Transaction Amount | Amount spent or received |
| Transaction Description | Text describing the transaction |
| Transaction Classification | Category of the transaction (e.g., “internal transfer”, “income – pos”) |
| Industry Class | Type of business where the transaction took place |
Step 2: Data Preprocessing & Feature Engineering
To make transaction data suitable for credit scoring, we transform raw records into meaningful features that capture a customer’s financial behavior. This stage ensures the dataset is clean, consistent, and ready for modeling.
1. Data Cleaning
a. Handle Missing Values
- Fill in missing TransactionType by analyzing text patterns in TransactionDescription (e.g., keywords like “salary” → Credit, “payment” → Debit).
- Fill in missing TransactionClassification using the IndustryClass reference (e.g., “supermarket” → Food & Groceries).
b. Remove Duplicates and Outliers
- Identify and delete duplicate transactions that may arise from system retries or synchronization issues.
- Detect outliers in transaction amounts using Z-score filtering to remove extreme values that could distort the model.
2. Normalize Numerical Data
- Apply Min–Max Scaling to the TransactionAmount field so that all values fall within a standard range (typically 0 to 1).
- Normalization ensures that large transactions do not dominate smaller ones during model training, allowing balanced learning across all features.
2. Feature Engineering
We derive customer-level aggregated features for credit risk assessment:
| Feature Name | Calculation | Purpose / Reason for Inclusion |
|---|---|---|
| Total Transactions | Count of all transactions per month | Measures the customer’s overall financial activity and engagement. |
| Average Transaction Amount | Mean value of TransactionAmount | Reflects general spending behavior and transaction consistency. |
| Maximum Transaction Amount | Highest transaction value within each period | Highlights large purchases or unusual spending spikes. |
| Spending-to-Income Ratio | Total Debit Transactions ÷ Total Credit Transactions | Indicates financial balance; a higher ratio may signal higher credit risk. |
| Late Payment Indicator | Count of transactions containing “DEBIT INTEREST CHARGED” | Detects repayment issues or late fee occurrences. |
| Credit Utilization Rate | Total Credit Transactions ÷ Credit Limit | Measures financial leverage and credit dependency. |
| Income Variability | Standard deviation of credit transaction amounts | Distinguishes stable income earners from those with fluctuating income patterns. |
Step 3: Splitting Data into Training, Validation, and Test Sets
Before building the model, the dataset must be divided into separate subsets to ensure that performance evaluation is unbiased and reliable.
| Dataset Split | Proportion | Purpose |
|---|---|---|
| Training Set | 70% | Used to train the machine learning model and help it learn patterns and relationships from the data. |
| Validation Set | 15% | Used to fine-tune model parameters and prevent overfitting through hyperparameter tuning. |
| Test Set | 15% | Reserved for final model evaluation to measure real-world performance on unseen data. |
Step 4: Training the SVM Model
Why Use Support Vector Machine (SVM)?
SVM is a robust classification algorithm particularly effective for financial data analysis. It performs well in high-dimensional spaces, handles imbalanced datasets, and can capture non-linear relationships using kernel functions.
Key Advantages:
- Performs effectively with structured numerical and categorical features.
- Uses kernel tricks (e.g., RBF) to separate complex financial data patterns.
- Suitable for credit scoring tasks where risk categories overlap.
Model Training Workflow
| Stage | Description |
|---|---|
| Label Encoding | Convert categorical credit scores into numerical labels — for example: Low Risk → 0, Medium Risk → 1, High Risk → 2. |
| Kernel Selection | Choose the RBF (Radial Basis Function) kernel to model non-linear decision boundaries. |
| Hyperparameter Tuning | Optimize C (regularization parameter) and gamma (kernel coefficient) using Grid Search with Cross-Validation (CV). |
| Model Training | Train the SVM on the training dataset, validate performance on the validation set, and refine based on feedback. |
Step 5: Model Evaluation
Key Performance Metrics
| Metric | Purpose / Description |
|---|---|
| Accuracy | Measures the overall proportion of correct classifications made by the model. |
| Precision | Evaluates how many of the customers predicted as high-risk are truly high-risk, minimizing false positives. |
| Recall | Measures the model’s ability to correctly identify all actual high-risk customers, reducing missed detections. |
| F1-Score | Provides a balanced metric that combines Precision and Recall, useful for imbalanced datasets. |
| AUC-ROC | Assesses the model’s ability to distinguish between different risk levels by analyzing the ranking quality across thresholds. |
Step 6: Model Deployment & Real-World Application
| Action | Description |
|---|---|
| Integrate into Risk Systems | Deploy the trained credit scoring model within the bank’s risk assessment platform to dynamically score customers based on their latest transaction data. |
| Monitor Model Drift | Continuously track model performance to detect data drift or shifts in customer behavior, retraining the model with updated data every six months. |
| Adjust Risk Thresholds | Regularly fine-tune credit-risk cut-off values in response to changing economic and market conditions to maintain accurate risk predictions. |
Further study
ALTERNATIVE DATA
| Category | Data Type | Description & Relevance |
|---|---|---|
| Bank Transactions | Spending behavior | Total transactions, average monthly spending, and largest purchase — indicate overall financial activity. |
| Income consistency | Frequency and stability of salary deposits — reflects income reliability. | |
| Credit card utilization | Percentage of credit limit used — measures debt management. | |
| Loan repayments | Timeliness of loan and credit card payments — reveals repayment discipline. | |
| Transaction classification | Spending categories (e.g., essentials vs. discretionary) — show lifestyle and financial priorities. | |
| Mobile & Telecom | Call & messaging behavior | Frequency of communication and international calls — used to infer stability and engagement. |
| Data usage patterns | Regular vs. erratic data usage — potential proxy for consistency and lifestyle. | |
| Bill payment history | Record of late vs. on-time payments — indicates payment reliability. | |
| E-commerce & Digital Payments | Online shopping frequency | Activity on platforms like Amazon or Alibaba — signals spending habits. |
| Buy Now Pay Later (BNPL) | Use of installment payments and repayment behavior — reflects credit discipline. | |
| Digital wallet transactions | Activity on PayPal, Alipay, or Apple Pay — shows digital finance adoption. | |
| Employment & Income | Salary deposits | Stability and frequency of income inflows. |
| Job tenure | Duration with current employer — longer tenure implies stability. | |
| Industry stability | Job security level varies by industry type. | |
| Social & Online Behavior | Social media activity | Engagement level on LinkedIn, Facebook, etc. — hints at social credibility and stability. |
| Professional network | Links to verified professionals — signals employment authenticity. | |
| Marketplace activity | Buying or selling activity on eBay, Taobao — indicates entrepreneurship or side income. | |
| Utility & Subscriptions | Utility bill history | Consistency in paying electricity, water, and internet bills on time. |
| Subscription services | Regular payments for Netflix, Spotify, or gym — shows financial discipline. | |
| Government & Public Data | Tax history | Timeliness of tax filing and payment. |
| Legal records | Presence of bankruptcies, lawsuits, or court rulings — reflects financial risk. | |
| Property ownership | Ownership or rental history — indicates asset stability. | |
| Alternative Financial Data | Microloan history | Borrowing and repayment patterns in microfinance platforms. |
| Peer-to-peer lending | Participation and repayment behavior in P2P lending platforms. | |
| Behavioral & Psychological Traits | Risk-taking patterns | Frequency of speculative investments or gambling — indicates financial risk appetite. |
| Lifestyle indicators | Travel frequency, luxury spending — reveals lifestyle-driven financial behavior. |
ML & LLM
Comparison of Machine Learning Approaches for Credit Scoring
| Approach | Model Type | Best Use Case | Advantages | Challenges |
|---|---|---|---|---|
| Logistic Regression | Traditional ML | Simple credit scoring | Easy to interpret, meets regulatory requirements | Assumes linear relationships |
| Decision Trees | Traditional ML | Risk classification | Captures non-linear patterns, easy to explain | Can overfit small datasets |
| Random Forest | Ensemble ML | Transaction classification | High accuracy, robust to noise and outliers | More computationally demanding |
| XGBoost | Ensemble ML | Default prediction | Excellent performance, handles missing data | Requires careful parameter tuning |
| Support Vector Machine (SVM) | Traditional ML | Fraud detection | Works well with high-dimensional data | Difficult to interpret results |
| Neural Networks | Deep Learning | Complex transaction behavior | Learns deep and abstract patterns | Needs large data and training time |
| LSTMs (Long Short-Term Memory) | Deep Learning | Time-series financial data | Captures temporal and sequential dependencies | Hard to train and fine-tune |
| BERT-based NLP Models | LLM | Transaction text classification | Extracts meaning from transaction descriptions | Needs fine-tuning and large resources |
| GPT-based Models | LLM | Alternative credit scoring | Handles unstructured and contextual financial data | Highly computationally expensive |
Review Questions
- In a machine learning–based credit scoring system, the dataset is divided into three subsets.
Identify these subsets and explain the purpose of each one.
2. What is transaction data, and what are its key attributes?
Describe the main components typically included in transaction datasets.
3. What are the main types of missing data encountered in transaction data processing?
Define each type and provide an example illustrating how it may occur.
4. What methods can be used to handle missing data?
Briefly explain each method and how they differ in application.
5. NLTK and spaCy are both tools for Natural Language Processing (NLP).
Compare their main characteristics and typical use cases.
6. List some commonly used machine learning models for transaction classification.
Briefly describe at least two models and highlight their key features.
7. What are the key evaluation metrics used in credit scoring models based on machine learning?
Explain what role each metric plays in assessing model performance.
8. Define credit scoring.
How do financial institutions apply machine learning techniques to improve credit scoring accuracy?
Conclusion
Throughout this module, we explored how machine learning transforms raw transaction data into meaningful insights for credit scoring and financial decision-making.
Students learned how to:
- Preprocess and clean transaction data, addressing missing values, noise, and normalization challenges.
- Apply Natural Language Processing (NLP) techniques using tools such as NLTK and spaCy to analyze text-based transaction descriptions.
- Build and evaluate classification models like Logistic Regression, Random Forest, XGBoost, and SVM to categorize transactions and predict credit risk.
- Measure performance using validated indicators such as Accuracy, Precision, Recall, F1-Score, and AUC-ROC.
- Implement the full credit scoring workflow, from data preparation and model training to deployment and monitoring in real-world banking systems.
In conclusion, machine learning enables smarter, data-driven lending decisions. By leveraging transaction-level data and continuous model updates, financial institutions can improve credit assessment accuracy, expand credit access responsibly, and adapt to changing market and customer behaviors.
THANK YOU