Work in progress!!

As a follow-up to puffins which was aimed at modelling strictly periodic signals in univariate time series data, I began thinking that it would be good to generalise this framework to handle multiple streams of time-series data from the same source that have multiple, independent linear trends while having the same underlying periodic signal. Which brings us to my currently ongoing project: nostradamus - named for the task of our model, to predict the behaviour of time series.

This work aims to provide a general and flexible regression framework for modelling univariate and multivariate time series data where the number of parameters (i.e. model terms) can be arbitrarily large. To do this, we draw inspiration from the Facebook PROPHET model (Taylor & Letham, 2017) and our existing work with puffins, where we introduce feature weighting in our regression model following Hogg & Villar, 2021.