tidyverts / feasts

Feature Extraction And Statistics for Time Series
https://feasts.tidyverts.org/
295 stars 23 forks source link

Should gg_tsdisplay have plot_type='histogram' as default? #68

Closed robjhyndman closed 5 years ago

robjhyndman commented 5 years ago

The most common use for this function is to look at residuals, and then the histogram is more useful than the PACF. Also, the use of ACF/PACF to pick an ARIMA model is largely about legacy rather than modern best practice, so perhaps it is time to retire it as a default.

mitchelloharawild commented 5 years ago

I'd think that the default should be an ensemble of plots that best display the features of the time series. Perhaps gg_tsdisplay should be more dynamic and show plots of seasonality when appropriate.

I think residual time series require a very different ensemble of graphics than the response time series, and so a separate function may be more appropriate here.

robjhyndman commented 5 years ago

OK. If we have a separate residuals function that wraps a particular form of gg_tsdisplay(), perhaps it should take the mable as the argument and generate residuals on the fly like forecast:::checkresiduals() does?

mitchelloharawild commented 5 years ago

Is there any extra information from the mable other than the residuals which would need to be used for the graphic?

robjhyndman commented 5 years ago

I don't think so

mitchelloharawild commented 5 years ago

Hmm - I'm torn for what is the best approach here. I think two functions is a good idea, but the ensemble of plots in gg_tsdisplay and gg_??? need further thinking.

What plots would you want to see when looking at a time series? How does this change based on the characteristics of the data? What if it is seasonal or multi-seasonal?

What plots should be looked at when checking the residuals?

robjhyndman commented 5 years ago

For a non-seasonal time series, probably a time plot, ACF, and either spectrum or histogram. For a seasonal time series with one period, probably time plot, ACF and season plot For a seasonal time series with multiple periods, I'm not sure. I don't think this has been thought about sufficiently in the literature or in practice. For residuals, a time plot, acf and histogram.

mitchelloharawild commented 5 years ago

Any suggestions for an appropriate name of the residual plot function? gg_tsresid, gg_residuals, gg_residualdisplay()...

robjhyndman commented 5 years ago

How about gg_tsresiduals() so we don't step on any other non-time series applications.