Detecting fraud is tough because fraud cases are less than 0.2% of all transactions. Recently several of our customers have been working on similar use cases, and I wanted to share how they are doing it at a massive scale with an agentic workflow.

To demonstrate this, I built a detailed anomaly detection system to analyze the Credit Card Fraud Detection dataset from Kaggle. The goals were to test different anomaly detection methods, find the best one, and set up an easy process for retraining and deployment with little manual work.

You can follow along in this Zerve Canvas.

The Challenge: Extreme Class Imbalance in Anonymized Data

The dataset has 284,807 transactions from European cardholders. Only 492 of these are fraud, which is just 0.172%. Features V1 to V28 are hard to understand because they come from PCA projections. The dataset also includes real values for Time and Amount, and a Class label where 1 means fraud and 0 means a normal transaction. This data is high dimensional, anonymized, and very unbalanced, making it ideal for testing anomaly detection methods.

Why Automate the Workflow?

Doing this project manually means repeating preprocessing steps, adjusting thresholds by hand, collecting metrics inconsistently, and risking mistakes.

Using Zerve's agent for Data Science to build an automated workflow, it allowed me to:

• Apply the same preprocessing for both training and inference

• Test multiple models in a consistent way

• Record every change, threshold, and metric

• Retrain and redeploy with just one pipeline run

Data Understanding & Preprocessing

We used z-score scaling at the beginning to standardize all numerical features for consistent results. The data was split into training and testing sets using stratified sampling, 80% for training and 20% for testing, to keep the rare fraud cases balanced. These steps were included in a reusable Zerve pipeline that processes all new batches of transactions in the same way.

Model Portfolio and Training Approach

We tested different unsupervised and supervised models to detect anomalies:

• Isolation Forest – Uses trees to score anomalies based on path length

• One-Class SVM – Separates normal transactions by defining boundaries

• Autoencoder – Detects anomalies by measuring reconstruction errors

• Deep SVDD-style AE – Learns a tight feature space around normal data

• Elliptic Envelope – Assumes normal data follows a multivariate Gaussian distribution

• HBOS – Scores outliers using histograms

We trained the models using only Class=0 (normal) transactions so they could learn typical patterns. For neural network models, we used a validation set to choose anomaly thresholds dynamically.

Automated Benchmarking and Selection

The pipeline used the stratified test set to make predictions and calculated:

• Precision

• Recall

• F1-score (main metric due to class imbalance)

• Area Under the Precision-Recall Curve (AUPRC)

• Confusion matrix

Zerve created a versioned report that gathered all results into one document, making it easier to compare models. The selection started by looking at the F1-score, then used AUPRC or recall to decide if F1-scores were tied. The best model was saved as a versioned .pkl file for production.

Automated Inference and Retraining

The production system loads the selected model and applies the same preprocessing steps before scoring new data. Scoring can be done either daily in batches or whenever needed.

Here's how it works:

• Load the chosen model.

• Preprocess data just like during training.

• Score incoming data using the preprocessed information.

When new labels are available, the system calculates and records performance metrics with timestamps. If it detects a drop in performance due to changes or feedback, it automatically retrains itself. This helps the system keep up with evolving fraud patterns.

Productivity Gains

Switching from manual coding to automated orchestration offers many benefits, such as:

• Time Saving – No need to repeat setup for each retrain

• Reproducibility – Complete logs of transformations and metrics for every run

• Consistency – It uses the same rules and thresholds during training and inference

• Scalability – It is easy to add more models or new data sources

Thoughts on Why This Matters for Fraud Detection

Detecting fraud well means spotting fake activities accurately, acting fast to stop losses, and doing this regularly. Using Zerve's automation tools, we built a system that finds unusual activity quickly with little need for people to step in. This mix of accuracy and speed is key in high risk areas like stopping fraud.

Manual vs. Automated Fraud Detection

Experience Zerve Now

Join our free tier and explore what is possible.