Simulating policy scenarios
Like wealth and inheritance taxation, to explore interventions without the risks and costs of real experiments.
I am a machine learning researcher who wants to understand the mechanisms behind socio-economic inequality: how it arises, how it evolves, and what its consequences are. So far, I have used and developed methods that combine mechanistic models with data. As a next step, I want to tackle methodological bottlenecks in inequality research.
The principle behind my approach is simple:
A mechanistic model (differential equations, dynamics, agents, …) encodes what we know, together with our uncertainty about it.
The data then inform, in a very targeted way, the missing bits, such as parameters, latent forces or unobserved states.
What comes out are estimates that are interpretable and come with uncertainty.
Inference thus happens in a structured model with interpretable constraints.
I have worked mostly on the methods themselves, such as Bayesian state-space model inference, low-rank approximations, and probabilistic solutions to differential equations, always with the application in mind. More recently, I have used diffusion models as learned priors for principled Bayesian inference, e.g. for large-scale problems on weather and climate data. Diffusion models themselves fit the principle above, since only a part of the model, the score network, is learned from data, while being embedded in a structured Gauss–Markov framework.
Inequality is an interesting multi-scale mechanism, with many interacting parts and dynamics over lifetimes and generations. How it is generated and how it is reproduced are questions about mechanisms and dynamics, but the available data, like surveys, snapshots or aggregates, are often incomplete.
This is exactly where mechanistic models can help, and where honest bounds are often more useful than a single “X caused Y” answer. Beyond inequality itself, I am interested in its connection to other dynamics, like demography.
Methods are means to an end, and the problem in the domain takes center stage. I want to work on:
Like wealth and inheritance taxation, to explore interventions without the risks and costs of real experiments.
Connecting structural models to data, with uncertainty.
Of such simulations, for researchers and for the public, to inform debates.
If you work on inequality, mobility or demography and think this could be useful for your questions, I am happy to talk.