Loading...
A practitioner course on building fraud-detection systems that actually work in production. Fraud is the biggest employer of machine-learning talent in South African financial services, and it looks nothing like a textbook classification problem: 0.17% positive rate, labels that arrive 60 days late, an adversary that adapts every week and a millisecond decision budget.
This course teaches how to handle every part of that reality. You will engineer velocity features on real transaction data, defeat the metric traps that make imbalanced problems look easier than they are, calibrate a gradient boosting model with a proper cost matrix, layer in unsupervised anomaly detection where labels do not exist, use graph analytics to surface fraud rings and anti-money-laundering patterns, and design the production system around it: real-time inference, rules alongside ML, analyst feedback and drift monitoring.
SA-specific throughout: SIM-swap and card-not-present as recurring worked examples, POPIA constraints on features, FIC Act obligations, SARB Directive 3/2020. Prerequisites: Python for AI and Machine Learning Fundamentals. Capstone builds an end-to-end pipeline on real card-fraud data.
JP Instructor
Already have an account? Sign in
What fraud actually looks like across cards, account takeover, SIM-swap, APP scams, synthetic identity and money laundering. The fraud lifecycle from attempt to chargeback, the four costs (fraud loss, false-positive friction, ops, opportunity), the SA context (SABRIC stats, SARB Directive 3/2020, FIC Act) and where AI fits versus rules. 5 lessons.
10 minutes
12 minutes
12 minutes
12 minutes
12 minutes
10 minutes
What signals actually predict fraud: transaction attributes, device and network, historical velocities, behavioural baselines, merchant risk. Building time-window aggregations (last 1h, 24h, 7d), peer-group features, SIM-swap indicators, and handling PII under POPIA. Practical: build velocity features on a real transaction dataset. 6 lessons.
10 minutes
12 minutes
12 minutes
12 minutes
12 minutes
10 minutes
Why accuracy is meaningless at a 0.17% fraud rate, and what to use instead. AUPRC vs AUROC and when each lies. Precision at K, recall at fixed false-positive rate, cost-weighted metrics. Resampling (SMOTE, undersampling) and their traps. Threshold selection driven by a cost matrix. Practical: baseline logistic and gradient boosting on imbalanced data, compared honestly. 6 lessons.
10 minutes
12 minutes
10 minutes
12 minutes
12 minutes
10 minutes
Gradient boosting (XGBoost, LightGBM) as the workhorse and why. Feature importance and SHAP for per-decision explanations. Calibration so probabilities mean what they say. Ensembling rules + model + secondary model. Model monitoring with PSI and drift detection. Practical: train, calibrate and explain a gradient boosting fraud model. 6 lessons.
10 minutes
12 minutes
12 minutes
10 minutes
12 minutes
10 minutes
When labels are absent, delayed by chargeback horizons, or wrong. Isolation Forest, Local Outlier Factor, autoencoders for reconstruction-based anomaly. Peer-group analysis. When to use unsupervised versus supervised, and how to combine them. Practical: Isolation Forest on a new merchant category with no labels. 5 lessons.
10 minutes
12 minutes
12 minutes
10 minutes
10 minutes
Why relationships matter in fraud detection. Building a transaction graph. Community detection for fraud rings. Connected components, k-core and PageRank for AML. The role of graph databases in a modern fraud stack. Practical: build a small transaction graph in Python and surface a suspicious cluster. 5 lessons.
10 minutes
12 minutes
12 minutes
12 minutes
10 minutes
Real-time inference architecture with sub-100ms budgets. Feature stores and streaming features. Why rules never die and how they compose with ML. Case management, analyst feedback loops and champion/challenger testing. Model retraining cadence and adversarial drift. Explainability for regulators and disputes. Practical: sketch a real-time scoring pipeline. 6 lessons.
12 minutes
12 minutes
10 minutes
10 minutes
12 minutes
10 minutes
A 6-step PROJECT capstone that builds a full fraud detection pipeline on a real imbalanced dataset (IEEE-CIS or Kaggle credit card). EDA and honest baseline, velocity feature engineering, gradient boosting model with proper AUPRC and cost-matrix evaluation, threshold calibration from a cost function, SHAP explanation for a flagged decision and a drift check with a re-train recommendation.
180 minutes
Join thousands of students mastering AI & Machine Learning with HardCoded
Sign Up to Get Started