Mitchell O’Hara-Wild (Monash University)
Rob Hyndman (Monash University)
The fable ecosystem provides a tidy interface for time series modelling and forecasting, leveraging the data structure from the tsibble package to support a more natural analysis of modern time series. fable is designed to forecast collections of related (possibly multivariate) time series, and to provide tools for working with multiple models. It emphasises density forecasting, whilst continuing to provide a simple user-interface for point forecasting.
Existing implementations of time series models work well in isolation, however it has long-been known that ensembles of forecasts improve forecast accuracy. Hybrid forecasting (separately forecasting components of a time series) is another useful forecasting method. Both ensemble and hybrid forecasts can be expressed as forecast combinations. Recent enhancements to the fable framework now provide a flexible approach to easily combine and evaluate the forecasts from multiple models.
The fable package is designed for extensibility, allowing for easier creation of new forecasting models and tools. Without any further implementation, extension models can leverage essential functionality including plotting, accuracy evaluation, model combination and diagnostics. This talk will feature recent developments to the fable framework for combining forecasts, and the performance gain will be evaluated using a set of related time series.