High Frequency Forecast using Fable

Thanks a lot for creating this fantastic package !!

I am working on high frequency retail sales data for forecast, (weekly sales data), Which other models are available for these high frequency data?

TBATS ?
FASSTER ?
PROPHET ?

What model do you recommend for high frequency forecast? Do you have any good example except electricity demand?

Many Thanks
Gaurav

Have you had a look at this? It has been rewritten to use fable.

Hi Gaurav,

Weekly data often contains only one seasonal pattern (with annual seasonal period), and so most seasonal models should handle your data fairly well. While models that are capable of capturing multiple seasonal patterns (TBATS, FASSTER, and Prophet) may work, they are likely more complicated than necessary for your data.

There are two complications of weekly data.

  1. Non-integer seasonal period (there are approx 365.25/7 = 52.18 weeks in a year).
  2. Rounding to 52 seasonal periods (as bad as that is), gives many seasonal values to estimate.

One approach to handling these problems is to use fourier terms. This can be done with any model that accepts exogenous regressors (TSLM(), ARIMA(), etc.), and some models handle fourier terms directly (TBATS, FASSTER, and Prophet).

Try a few of these models and see what works best for your data.

As an example, you can add exogenous regressors to ARIMA with:
ARIMA(y ~ fourier("year", K = 8) + pdq(...).

The choice of K (number of harmonics) controls how smooth or flexible the seasonal pattern is. Again, you can experiment with this to find which suits your data best.

More details: https://robjhyndman.com/hyndsight/forecasting-weekly-data/

1 Like

Thanks a lot !! do we have tsibble, fable package with Anomalize?

There aren't any anomaly detection features for tidy time series yet.
Support for identifying outliers based on a given model is planned (https://github.com/tidyverts/fable/issues/160), but anomaly detection algorithms will come after this.

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.