Research

Machine Learning Theory

A Framework for Optimal High-Dimensional Regularisation

Develops a framework for formally identifying which regularisation method is optimal according to dataset properties and the researcher’s objective. Introduces a novel estimator, the Manifold-adaptive Entropic Shrinkage Algorithm (MESA), an $\ell_p$ estimator that exhibits both stability and sparsity in challenging environments. MESA provably outperforms LASSO, Ridge, and Elastic Net in high-dimensional settings with significant correlation between variables, which is demonstrated with an empirical application to large-scale political text data.

Neural Networks as Optimal Continuous Piecewise-Linear Approximators (with Jakob Foerster)

Demonstrates that ReLU neural networks can be understood as near-optimal continuous piecewise-linear approximators. Introduces a computable measure of the inherent complexity of a learning problem, and applies optimisation theory to explain when and why neural networks predict well. A meta-learner approach allows for principled architecture selection by using observable dataset features to predict depth and width requirements given a performance target.

Political Economy

Measuring Democratic Responsiveness across Multiple Principals

Studies how elected representatives respond to competing pressures from voters, donors, and political elites. Develops a unified Bayesian framework, with emergent dimension discovery via sparse factor analysis and multilevel regression with post-stratification to measure influences nonparametrically and at scale across the UK, US, and Australia.

Narratives and Markov Equivalence in Political Economy

Provides a new formalisation of the idea that political polarisation can be traced to competing causal models (“narratives”) that voters use to interpret the same data. The framework distinguishes disagreements that are resolvable by updating beliefs from those that are structurally underdetermined by any possible observational evidence, and derives conditions under which societies become trapped in low-epistemic-quality equilibria. Provides a novel empirical measurement of causal model divergence in political speech.

Critical Junctures and Institutional Uncertainty (with Michael Callen, Jonathan Weigel, and Noam Yuchtman)

Presents a new empirical approach to detecting critical junctures — periods during which multiple institutional futures become possible — as they unfold, using historical newspaper articles. Applies a large language model classifier to $\approx$ 13 million articles from The Times of London (1800–2019), producing a long-run text-based measure of institutional uncertainty that captures elevated uncertainty even in periods without eventual institutional change. Forthcoming in the Handbook of Political Economy (Acemoglu & Robinson, eds.). Appeared as an NBER Featured Working Paper (Aug 31, 2026).

Salience, Higher-Order Beliefs, and Coordination in Protest (with Lionel Page)

An experimental study of how salience and higher-order beliefs influence coordination in games of political uprising. In a stag-hunt framing where participants choose whether to protest against an authoritarian regime, we exogenously vary the salience of the protest option and players’ beliefs about this salience. Coordination rates rise from 39% under no information to 96% under common knowledge of salience, with the effect increasing monotonically through knowledge levels. Dictator-role bidding games confirm that higher knowledge levels are strategically valued. (Awarded UQ Synergies Consulting Honours Prize for Best Honours Thesis in Economics, 2022.)