Extra Projects
Machine Learning and Decision Analysis
Loan Default Prediction and Lending Decision Analysis
A credit-risk modelling project using the HMEQ home-equity loan dataset to predict borrower default and assess lending decisions under asymmetric costs.
The analysis compares logistic regression, tree-based models, ensemble methods, and XGBoost, then evaluates not only classification performance but also the expected financial loss from false approvals and false rejections. The project highlights why lending models should be assessed using decision-relevant costs rather than accuracy alone.
Tools and methods: Python, scikit-learn, XGBoost, model evaluation, threshold optimisation, expected-loss analysis, SHAP.
Agent-Based Simulation and Network Economics
Network Segregation Model
This NetLogo project explores how segregation can emerge in a social network. The model is inspired by Schelling’s classic segregation model, but replaces spatial movement with network link formation: agents do not move across locations, they choose whom to connect to.
Agents belong to one of two groups and evaluate their immediate network neighbourhood according to the share of same-type neighbours. The version implemented here uses a threshold-flat utility function: agents want at least a chosen share of similar neighbours, but once this threshold is reached they receive no additional benefit from further homogeneity. In other words, agents do not directly seek full segregation; they only require “enough” similarity.
Links are also costly to maintain. Agents therefore balance the benefit of satisfying their similarity threshold against the cost of holding connections. When agents are unhappy, they may cut, create, or rewire links if doing so improves their own utility.
The model shows how decentralised individual decisions can generate unintended aggregate segregation. Even moderate similarity thresholds can gradually reduce cross-group links and produce clustered networks, because agents do not fully internalise how their link choices affect others.
How to use the model: choose the number of agents with the buttons slider, set the desired similarity threshold with %-similar-wanted, and adjust the link cost and payoff parameters. Press Setup to create the initial random network, then press Go to let unhappy agents update their links until the network reaches a stable outcome.
Key outputs include the percentage of happy agents, the number of links, the share of same-type links, the clustering coefficient, the ghetto rate, and average welfare.