Credit Risk Scoring with LSTM Networks
"I had no prior experience with sequence models. The programme forced me to understand why architecture choices matter before touching any code. The feedback on my first draft was blunt - that was exactly what I needed."
Thabo, on the review process
Applied LSTM networks to predict default probability on retail lending portfolios using South African credit bureau data. Final model was evaluated against a logistic regression baseline across three threshold settings.
Sentiment-Driven Portfolio Rebalancing
NLP pipeline extracting sentiment from JSE earnings call transcripts, mapped to monthly portfolio weight adjustments.
FinBERT fine-tuned on a manually labelled subset of 340 transcript segments. Rebalancing logic kept deliberately simple to isolate the signal contribution.
Fraud Detection Using Autoencoders
Unsupervised anomaly detection on transaction sequences to flag unusual payment patterns without requiring labelled fraud examples.
Reconstruction error threshold tuned against a held-out set of known fraud cases provided by a Nigerian fintech partner under NDA.
Volatility Forecasting with Temporal Convolutional Networks
TCN architecture applied to realised volatility estimation across emerging market equity indices, comparing against GARCH benchmarks.
Dilated convolutions allowed the model to capture dependencies at multiple lags simultaneously - a design choice justified in the write-up with ablation results.
What makes these projects worth reading
Each submission goes through two rounds of structured critique - one from a peer cohort member, one from a programme mentor. The goal is a defensible argument, not just working code.
Students are required to document failed experiments alongside successful ones. That requirement produces write-ups that reflect how financial modelling actually works in practice.