Talk structure

[mitchelloharawild]

Topics discussed in abstract

• Tidy interface
• tsibble data structure
• modelling many series
• density forecasting
• combination forecasting
• forecast ensembling
• hybrid/decomposition forecasting
• forecast evaluation
• extensibility

Additional topics to consider

• Parallel support
• Forecast reconciliation
• fasster/fable.prophet

Conference specific changes

For ISF 2019 my talk will follow a related talk about feasts (https://github.com/robjhyndman/feasts-talk). In this talk, the concept of tsibbles and the tourism dataset is introduced. This is not the case for useR!2019, so these topics should be introduced briefly.

It is assumed that the audience at ISF will be more interested in how more sophisticated forecasting techniques are applied using fable. They may also be interested in how other forecasting models can be integrated with fablelite. Spend less time on code, but more time on combinations, reconciliation and extensibility.

The useR! audience is likely less experienced with forecasting, and so a more gradual introduction is appropriate. There will probably not be enough time to cover reconciliation in this talk. Instead focus on the interface, ensembling, extensibility, and parallel.

General talk structure

Talk dataset: tourism Key topics: combination forecasting and extensibility (less about fable, more about fablelite)

• [x] Briefly introduce tidyverts -> fable
• [x] Introduce tourism data
• [x] Produce a basic forecast of the aggregate tourism data
• [x] Produce an ensemble forecast of aggregate tourism data
• [x] Add in a model from a different package (say fasster)
• [x] Evaluate accuracy
• [x] Produce similar forecasts for all series in tourism
• [x] Reconcile
• [x] Evaluate accuracy

[mitchelloharawild]

Bonus slides

• [ ] Exogenous regressors
• [ ] Decomposition forecasting
• [ ] Multivariate modelling
• [ ] Prophet
• [ ] Broom
• [ ] Transformations

[mitchelloharawild]

Adjustment ideas for useR! 2019

• [x] Replace reconciliation section with decompositions
• [x] Spend more time exploring tsibble and forecasting verbs