12-08, 18:00–18:30 (UTC), Machine Learning Track
Extreme events are ubiquitous, ranging from temperature records to stock market crashes or network outages. Using extreme weather events as an example we show how they can be modeled in a Bayesian way using PyMC. We start with simple models and ultimately move on to a more advanced model by implementing a Gaussian Process Latent Variable Model, which allows us to perform spatial modeling of extreme events.
The study of extreme events concerns asking questions about what a once every 100-year flood event would look like. The more common statistical methods are ill-suited to answer these types of questions since they focus on the usual rather than the unusual (extremes). Instead we need to turn to Extreme Value Theory, to which we give a brief introduction.
The rest of the presentation is very hands-on. We work out a few recent examples of extreme weather events in PyMC, which is a probabilistic programming library. This allows us to efficiently build Bayesian models. One advantage of building Bayesian models is that they can easily quantify the uncertainty of our predictions, which is crucial when making predictions about extreme events.
Finally we build a more advanced Bayesian model by implementing a Gaussian Process Latent Variable Model, which allows us to perform spatial modeling of extreme events. This not only allows us to make predictions for unobserved locations, it also helps reduce the uncertainty of our predictions.
In short you will learn something about
* Statistics of extreme (weather) events
* Bayesian in modeling in PyMC
* Gaussian Processes for spatial modeling
Prerequisites
* Familiar with Python
* Basics concepts of Statistics and Probability
Previous knowledge expected
Jorn Mossel works as a Data Scientist in Energy demand forecasting. Prior to that, he worked for a decade on Wall Street as a quant, building systematic trading strategies and risk models. Jorn holds a Ph.D. in Theoretical Physics.