Interpolation

[flaviorbs]

Hey Mitch.

At some point you have shown interpolation using Fable + Fasster.

I think in one of the updates we lost functionality (I don't see interpolate method in Fasster code nor in previous commits). Any chances of getting this to work again or workarounds?

library(tidyverse)
library(lubridate)
library(tsibble)
library(fable)
library(fasster)

elec <- tsibbledata::vic_elec %>% 
  filter(Time < ymd("2014-03-01")) %>% 
  mutate(WorkDay = ifelse(wday(Time, week_start = 1) > 5, 0, 1))

interpolation <- elec %>% 
  mutate(Demand = ifelse(Time >= ymd("2014-02-01") & Time <  ymd("2014-02-14"), NA, Demand)) %>% 
  fasster(log(Demand) ~ WorkDay %S% (trig(48, 16) + poly(1))) %>% 
  interpolate()

[BirgerNi]

Same problem, although I am not sure whether the method interpolate.fasster is missing or that should be handled by interpolate.mdl_ts from fabletools package. @mitchelloharawild do you have an idea?

suppressWarnings(library(fabletools))
suppressWarnings(library(fable))
suppressWarnings(library(fasster))
suppressWarnings(library(tsibbledata))

fit_tslm <- olympic_running %>% 
  model(model = TSLM(Time ~ trend())) 
fit_fasster <- olympic_running %>%
  model(model = fasster(Time ~ trend()))

# work
fit_tslm %>% 
  interpolate(olympic_running)
#> # A tsibble: 312 x 4 [4Y]
#> # Key:       Length, Sex [14]
#>    Length Sex    Year  Time
#>     <int> <chr> <int> <dbl>
#>  1    100 men    1896  12  
#>  2    100 men    1900  11  
#>  3    100 men    1904  11  
#>  4    100 men    1908  10.8
#>  5    100 men    1912  10.8
#>  6    100 men    1916  10.8
#>  7    100 men    1920  10.8
#>  8    100 men    1924  10.6
#>  9    100 men    1928  10.8
#> 10    100 men    1932  10.3
#> # ... with 302 more rows

# does not work
fit_fasster %>%
  interpolate(olympic_running)
#> Error: Problem with `mutate()` input `interpolated`.
#> x nicht anwendbare Methode für 'interpolate' auf Objekt der Klasse "FASSTER" angewendet
#> i Input `interpolated` is `map2(model, new_data, interpolate, ...)`.

# there is no interpolate.fasster method
methods(interpolate)
#> [1] interpolate.ARIMA*      interpolate.mdl_df*     interpolate.mdl_ts*    
#> [4] interpolate.model_mean* interpolate.TSLM*      
#> see '?methods' for accessing help and source code

^{Created on 2021-03-22 by the reprex package (v1.0.0)}

[mitchelloharawild]

An interpolate(<fasster>) method is required for this to work. For the talk I had quickly included it as a proof of concept, but in the weeks after it was clear that this feature was better suited as a generic in {fabletools} (c94d74f). Seemingly I didn't add the necessary method into fasster for this.
The initial interpolate() method was simply used fitted values, however it would be better/nicer to use smoothed values instead.

[BirgerNi]

Thanks for your reply. I guess same problem for fable.prophet. Is that feature planned in the near future or do you know a work-around? I need to interpolate a multi-week gap in a district heating load (no forecast neccessary).

[mitchelloharawild]

I've re-added an interpolate() method that simply uses the fitted values like before. I hope to one day improve this by using smoothed values (a bit more computation, but much better results), but I expect this would take a while.