tidyverts / fasster

Forecasting with Additive Switching of Seasonality, Trend and Exogenous Regressors

tidyverts

Homepage

fasster's Issues

Confidence intervals for log() transformed time series

Hi!

I'm having some trouble figuring out how to formulate confidence intervals for a model where I am applying a log() transform to the outcome. The forecast plot looks perfectly reasonable but I need to get my hands on the numerical values of the confidence intervals.

This is what the model looks like:

fit <- data %>%
  fasster(
    log(y) ~ poly(1) + seas(24)
  )

I'm extracting data for the forecast as follows:

fit %>% 
  forecast(h=24) %>%
  pull(forecast) %>%
  .[[1]]

and the output looks like this:

   time                  mean    se distribution       
   <dttm>               <dbl> <dbl> <dist>             
 1 2018-09-25 08:00:00  887.   1.31 t(N(6.2, sd = 1.3))
 2 2018-09-25 09:00:00  339.   1.34 t(N(5.2, sd = 1.3))
 3 2018-09-25 10:00:00  399.   1.38 t(N(5.3, sd = 1.4))
 4 2018-09-25 11:00:00  367.   1.43 t(N(5.2, sd = 1.4))
 5 2018-09-25 12:00:00   40.4  1.47 t(N(3, sd = 1.5))  
 6 2018-09-25 13:00:00  103.   1.51 t(N(3.9, sd = 1.5))
 7 2018-09-25 14:00:00  580.   1.55 t(N(5.6, sd = 1.5))
 8 2018-09-25 15:00:00  858.   1.59 t(N(5.9, sd = 1.6))
 9 2018-09-25 16:00:00 2981.   1.62 t(N(7.2, sd = 1.6))
10 2018-09-25 17:00:00 3459.   1.66 t(N(7.3, sd = 1.7))

Am I right in assuming that the 95% confidence interval for the first forecast point would be

qlnorm(c(0.025, 0.975), 6.2, 1.3)

If there's a function for extracting the confidence intervals that would be great. However, I have no problem extracting them by hand. Just want to know that I'm doing the right thing!

Thanks,
Andrew.

Implement rolling lm heuristic

Do we need dplyr, tidyr, pillar as dependencies

Could you please have a look at ?tsibble::reexports? If there are lots of verbs you used that have been re-exported, you may like to remove them as explicit dependencies.

Getting errors while running fasster

First of, the package is really great, thanks a lot ! I have already used msts to define the time series (passenger_count) but I get the following warning of not using time series,

fasster:::fasster(numberOfPassengersTimeseries, passenger_count ~ temperature + Rain, heuristic = c("filterSmooth")) %>% ggfitted

Warning message:
In fasster:::fasster(numberOfPassengersTimeseries, passenger_count ~ :
Time series information not found, please provide a tsibble or time-series object.

I also add the coefficients of AR and MA as vectors but I get error. Can you please give an example of a arma added to fasster?

Thanks in advance

Add support for conditional modelling

Potential UI:

FASSTER(y ~ poly(1) + G %S% (trig(12)) + (G == "A") %only% (x + z))

Error when running "forecast"

I just followed the code example on the README but when I tried to run the forecasting function, I got an error message "Error: x is not a tsibble". Any ideas? Thanks in advance!

library(fable)
fit <- as_tsibble(USAccDeaths) %>% 
  model(fasster = FASSTER(value ~ poly(1) + trig(12)))
fit %>% 
  forecast(h=24) %>%
  autoplot(as_tsibble(USAccDeaths))

Dynamically specifying the formula says 'No supported inverse for this function'

Hi, Great work on the package by the way.

I want to generate my formula dynamically from a list of variable names. So I tried to paste them into a string and convert that into a formula. However, this is throwing an error saying "No supported inverse for this function"

Example:

# Preparing the data
airquality %>% mutate(
    date = as.Date(paste('2018', Month, Day, sep = '-')), 
    Ozone = round(imputeTS::na.interpolation(Ozone, option = 'stine')),
    Solar.R = round(imputeTS::na.interpolation(Solar.R, option = 'stine'))
  ) %>% 
  as_tsibble()

# Creating formula
f <- as.formula('y ~ Wind + Temp + Solar.R + trig(12) + poly(1) + ARMA(1, 1)')

# Building fasster without formula parsing
fsModel <- FASSTER(data = train, 
    formula = y ~ Wind + Temp + Solar.R + trig(12) + poly(1) + ARMA(1, 1))

# Building fasster - throws an error 
fsModel <- FASSTER(data = train, formula = f)

I am not sure if I am doing anything wrong here. The same formula when passed directly worked perfectly fine.

Add support for multiple keys

Error in .subset2(public_bind_env, "initialize")

fit <- USAccDeaths %>% 
  as_tsibble %>% 
  model(fasster = FASSTER(value ~ poly(1) + trig(12)))

gives an error
Error in .subset2(public_bind_env, "initialize")(...) : argument ".env" is missing, with no default

Conflict: tsCV() requires forecast package, but to use the argument 'forecastfunction' in tsCV() forecast function has to be detached

Hi Mitchell,

Would you please check the forecast function?

library(fasster)
library(tsibble)
library(tidyverse)
ped_sthx <- pedestrian %>% 
  filter(
    Sensor == "Southern Cross Station",
    yearmonth(Date_Time) < yearmonth("2015 Apr")
  )
library(lubridate)
fit <- ped_sthx %>% 
  model(
    mdl = fasster(Count ~ poly(1) + (trig("week", 8)))
  )
fit %>% forecast(h = 10)

Error in FUN(left, right) : non-numeric argument to binary operator
In addition: Warning message:
In mean.default(x, na.rm = TRUE) :
  argument is not numeric or logical: returning NA

Another example directly from the forecast.TSLM {fable} R documentation

library(fable)
as_tsibble(USAccDeaths) %>% 
  model(lm = TSLM(log(value) ~ trend() + season())) %>% 
  forecast()

It also doesn't work.

Error in as.data.frame.default(x[[i]], optional = TRUE, stringsAsFactors = stringsAsFactors) : 
  cannot coerce class ‘"mdl_ts"’ to a data.frame
In addition: Warning message:
In mean.default(x, na.rm = TRUE) :
  argument is not numeric or logical: returning NA

Could you please help me with this?
Thank you for checking this post!

Implement heuristic specific options

Such as allowing the include option to apply only to the heuristic estimation step.

Implement streaming data

Include example datasets

Short forecast horizons can give incompatible matrices

When computing specials, if some elements never switch in then the resulting specials may be of incorrect dimension.

Rewrite dlm component

Ideally, improving speed, functionality, and object structures.

Write vignettes

Allow manual parameter specification

Drop usage of `poly()`

To avoid conflicts with potentially useful xreg function stats::poly()

How can I get multicore performance using fable and fasster?

I have a tsibble from this competion https://www.kaggle.com/c/demand-forecasting-kernels-only/overview, fasster is truly fast and cool, but I think I can get 4x performance if I can fit the model on 4 cores.

Add include option for fasster

Evaluate accuracy

Edit suggestion for slide 42 (http://www.mitchelloharawild.com/melburn17)

I was trying to reproduce your example and notice one small edit that may help future users :)

library(lubridate)
library(forecast)

tbl_taylor <- as_tsibble(taylor) %>%
  mutate(index = seq(ymd_h("2000-6-5 00"), by="30 mins", length.out=length(taylor)),
         DayType = ifelse(wday(index) %in% 2:6, "Weekday", "Weekend")) %>% 
  filter(month(index) == 8)

tbl_taylor %>%
  filter(index < ymd("2000-8-21")) %>%

  # original line
  # fasster(taylor~ DayType %S% (poly(1) + trig(48, 16))) %>%

  # revised line
  fasster(value ~ DayType %S% (poly(1) + trig(48, 16))) %>%
 
  forecast(newdata=tbl_taylor %>%
             filter(index >= ymd("2000-8-21"))) %>%
  autoplot() + 
  ylab("Electricity Demand (Megawatts)") + 
  xlab("Time (5 June 2000 - 27 August 2000)")

stream()/forecast() fails for short new_data

Hi @mitchelloharawild,

When using a switching model, both stream() and forecast() parse the new_data in a way that is inconsistent with the already trained model.

Consider the following model:

elec_tr <- tsibbledata::vic_elec %>%
  filter(
    yearmonth(Time) == yearmonth("2012 Mar")
  ) %>% 
  mutate(WorkDay = wday(Time) %in% 2:6 & !Holiday)

elec_fit <- elec_tr %>%
  model(
    fasster = fasster(Demand ~ 
      WorkDay %S% (trig(48, 16) + poly(1)) + Temperature + I(Temperature^2)
    )
  )

elec_fit

It is trained on the boolean switch WorkDay, however if the new_data input is too short, like in this case:

elec_update <- tsibbledata::vic_elec %>%
  filter(
    yearmonth(Time) == yearmonth("2012 Apr")
  ) %>% 
  head(2) %>%
  mutate(WorkDay = wday(Time) %in% 2:6 & !Holiday)

elec_fit_upd <- elec_fit %>% stream(elec_update)
elec_fit_upd

then internal DLM states and X will have wrong dimensions as they don't account for unseen states (say, holidays/weekends when the data contains only weekdays):

Error in rbind(object$dlm$X, X) : 
  number of columns of matrices must match (see arg 2)
Calls: <Anonymous> ... .f -> stream.mdl_ts -> stream -> stream.FASSTER -> rbind

Just to try out the streaming functionality, I fixed it by patching stream() to always use longer new_data, but then tail()'ing X/est/residuals when needed. Obviously, this is extremely ugly.

Not sure how to fix it at the moment. Am I missing something? If you have any ideas to share, I'd love to try fixing it and submitting a PR.

Thanks!

Improve heuristic optimisation

Potential directions include loess and sampling techniques.
Automatic model specification should also be considered with this.

can't reproduce USAccDeaths example: Error in FUN(left, right)

fit <- USAccDeaths %>% 
  as_tsibble() %>% 
  model(fasster = FASSTER(value ~ poly(1) + trig(12)))
fit %>% 
  forecast(h=24) 

gives Error in FUN(left, right) : non-numeric argument to binary operator

Heuristic (fasster version)

• Divide data into blocks of length 4*N_par
• For short series, overlap blocks, long series have no overlapping
• Randomly sample data blocks (up to nblock parameter)
• Fit model
• Pool and take variances

Levels in state-switching

Can state-switching only handle 2 levels? {0,1}

I had tried with three, but the forecast resulted in a flat line (y_hat=0)

Smarter parsing of lhs of %S% special

No need to parse this into a dlm, treat it as data

Implement accuracy() for fasster forecasts

Create model object structure

With associated methods for print, residuals, and other typical methods

Implement support for multiple group structures

Improve plotting functions

Especially the fortify method, should aggregate the like terms appropriately.

parse_model no longer returns args

Hello, I appreciate this is a very new bug which has come with a recent update in the fablelite package but I thought I would flag. I'm really enjoying exploring the capabilities of this package so hopefully this can be resolved soon.

model.R has a bug where

dlmModel <- model_inputs$args %>%
    unlist(recursive = FALSE) %>%
    reduce(`+`)

should be

dlmModel <- model_inputs$specials %>%
    unlist(recursive = FALSE) %>%
    reduce(`+`)

this is also the case for other places where parse_model is used. I can't seem to push my fixed code which is why this is an issue and not a pull request.

Error running vignette example with poly()

As per subject; taking out poly() term avoids the error.

library(fasster)
library(tsibble)
library(lubridate)
  
daytype <- function(date){
  factor(ifelse(wday(date) %in% 2:6, "Weekday", "Weekend"), levels = c("Weekday", "Weekend"))
}

SthCross_Ex <- pedestrian %>% 
  filter(Sensor == "Southern Cross Station") %>% 
  mutate(DayType = daytype(Date)) %>% 
  as_tsibble()

SthCross_fasster_fit <- SthCross_Ex %>%
  fasster(Count ~ DayType %S% (poly(1) + trig(24)))
#> Error in dlmSmooth.dlmFiltered(filtered): error code 1 from Lapack routine dgesdd

SthCross_fasster_fit <- SthCross_Ex %>%
  fasster(Count ~ DayType %S% (poly(1)))

Created on 2018-07-29 by the reprex package (v0.2.0).

Does Fasster handle transformations during fit?

Does Fasster perform similarly to the forecast package in that it can handle transformations in model and then back transform at forecasting, such as taking the difference or a seasonal decomposition?

Create tbl_forecast

How to specify new_data

Copied comments of 63f53d7 for more visibility.

Hi Mitchell, I have been your R package 'fable' and 'FASSTER' fan for a while. After FASSTER changes the way it incorporates newdata to the forecast function, my original code couldn't work.

Can you provide an example of how to include new_data into the forecast function (like how you do it to forecast the Australian electricity demand)?

The newdata I need to provide is "daytype" before the switching term %S% ( exactly like what you did for the Australian electricity forecast). I read your code and relevant information online regarding this change but still couldn't figure out how to do it by myself.

Appreciated,
Yijun (@yijunwang0805)

Consider adding pillar to print methods

Fix heuristic for models with near-singular Z matrix

Drop dlm unnecessary structures

Add automatic modelling suport

If no formula is specified, a model can be automatically identified.
There are some complexities with this:

• Identifying time periods to switch on (this likely requires selection from a set of supported switching functions)
• multiple seasonal patterns of potentially arbitrary period would need to be identified

Check first switched forecast observational variance

Add progress indicators

object 'object' not found

Hi,

I viewed your presentation from useR2018 and was deeply impressed by what's possible with fasster. This looks really amazing!

I've been working my way through the examples of the README.

> fit <- tsibbledata::UKLungDeaths %>%
+   FASSTER(fdeaths ~ mdeaths)
> 
> class(fit)
[1] "mable"      "lst_ts"     "tbl_df"     "tbl"        "data.frame"
> 
> fit %>% summary
FASSTER Model:
 fdeaths ~ mdeaths 

Estimated variances:
 State noise variances (W):
  mdeaths
   1.7119e-34

 Observation noise variance (V):
  1.6631e+03

Everything up to this point works perfectly. But when I try

fit %>% components

or

fit %>% forecast(h=24) %>% autoplot

I get an error:

Error in .f(.x[[i]], ...) : object 'object' not found

Please let me know if you need more information.

Best regards,
Andrew.

Here's my session information:

> devtools::session_info()
Session info ------------------------------------------------------------------------------------------------------------------------------------------------
 setting  value                       
 version  R version 3.5.1 (2018-07-02)
 system   x86_64, linux-gnu           
 ui       RStudio (1.1.456)           
 language en_ZA:en                    
 collate  en_ZA.UTF-8                 
 tz       Africa/Johannesburg         
 date     2018-09-14                  

Packages ----------------------------------------------------------------------------------------------------------------------------------------------------
 package     * version    date       source                                
 assertthat    0.2.0      2017-04-11 CRAN (R 3.5.1)                        
 base        * 3.5.1      2018-07-03 local                                 
 bindr         0.1.1      2018-03-13 CRAN (R 3.5.1)                        
 bindrcpp    * 0.2.2      2018-03-29 CRAN (R 3.5.1)                        
 colorspace    1.3-2      2016-12-14 CRAN (R 3.5.1)                        
 compiler      3.5.1      2018-07-03 local                                 
 curl          3.2        2018-03-28 CRAN (R 3.5.1)                        
 datasets    * 3.5.1      2018-07-03 local                                 
 devtools      1.13.6     2018-06-27 CRAN (R 3.5.1)                        
 digest        0.6.17     2018-09-12 cran (@0.6.17)                        
 dlm           1.1-5      2018-06-13 cran (@1.1-5)                         
 dplyr         0.7.6      2018-06-29 CRAN (R 3.5.1)                        
 fable       * 0.0.0.9000 2018-09-14 Github (tidyverts/fable@6e81dda)      
 fablelite   * 0.0.0.9000 2018-09-14 Github (tidyverts/fablelite@98271e9)  
 fasster     * 0.1.0.9000 2018-09-14 Github (tidyverts/fasster@7633a99)    
 forecast      8.4        2018-06-21 CRAN (R 3.5.1)                        
 fracdiff      1.4-2      2012-12-02 CRAN (R 3.5.1)                        
 ggplot2       3.0.0.9000 2018-09-14 Github (tidyverse/ggplot2@1e01e68)    
 glue          1.2.0      2017-10-29 CRAN (R 3.5.1)                        
 graphics    * 3.5.1      2018-07-03 local                                 
 grDevices   * 3.5.1      2018-07-03 local                                 
 grid          3.5.1      2018-07-03 local                                 
 gtable        0.2.0      2016-02-26 CRAN (R 3.5.1)                        
 lattice       0.20-35    2017-03-25 CRAN (R 3.5.0)                        
 lazyeval      0.2.1      2017-10-29 CRAN (R 3.5.1)                        
 lmtest        0.9-36     2018-04-04 CRAN (R 3.5.1)                        
 lubridate     1.7.4      2018-04-11 CRAN (R 3.5.1)                        
 magrittr      1.5        2014-11-22 CRAN (R 3.5.1)                        
 memoise       1.1.0      2017-04-21 CRAN (R 3.5.1)                        
 methods     * 3.5.1      2018-07-03 local                                 
 munsell       0.5.0      2018-06-12 CRAN (R 3.5.1)                        
 nlme          3.1-137    2018-04-07 CRAN (R 3.5.0)                        
 nnet          7.3-12     2016-02-02 CRAN (R 3.5.0)                        
 numDeriv      2016.8-1   2016-08-27 CRAN (R 3.5.1)                        
 parallel      3.5.1      2018-07-03 local                                 
 pillar        1.2.3      2018-05-25 CRAN (R 3.5.1)                        
 pkgconfig     2.0.1      2017-03-21 CRAN (R 3.5.1)                        
 plyr          1.8.4      2016-06-08 CRAN (R 3.5.1)                        
 purrr         0.2.5      2018-05-29 CRAN (R 3.5.1)                        
 quadprog      1.5-5      2013-04-17 CRAN (R 3.5.1)                        
 quantmod      0.4-13     2018-04-13 CRAN (R 3.5.1)                        
 R6            2.2.2      2017-06-17 CRAN (R 3.5.1)                        
 Rcpp          0.12.18    2018-07-23 cran (@0.12.18)                       
 rlang         0.2.2      2018-08-16 cran (@0.2.2)                         
 scales        1.0.0      2018-08-09 CRAN (R 3.5.1)                        
 stats       * 3.5.1      2018-07-03 local                                 
 stringi       1.2.4      2018-07-20 CRAN (R 3.5.1)                        
 stringr       1.3.1      2018-05-10 CRAN (R 3.5.1)                        
 tibble        1.4.2      2018-01-22 CRAN (R 3.5.1)                        
 tidyr         0.8.1      2018-05-18 CRAN (R 3.5.1)                        
 tidyselect    0.2.4      2018-02-26 CRAN (R 3.5.1)                        
 timeDate      3043.102   2018-02-21 CRAN (R 3.5.1)                        
 tools         3.5.1      2018-07-03 local                                 
 tseries       0.10-45    2018-06-04 CRAN (R 3.5.1)                        
 tsibble       0.5.2.9000 2018-09-14 Github (tidyverts/tsibble@0bc8ceb)    
 tsibbledata   0.0.0.9000 2018-09-14 Github (tidyverts/tsibbledata@3c2baef)
 TTR           0.23-3     2018-01-24 CRAN (R 3.5.1)                        
 urca          1.3-0      2016-09-06 CRAN (R 3.5.1)                        
 uroot         2.0-9      2017-01-29 CRAN (R 3.5.1)                        
 utils       * 3.5.1      2018-07-03 local                                 
 withr         2.1.2      2018-03-15 CRAN (R 3.5.1)                        
 xts           0.11-1     2018-09-12 CRAN (R 3.5.1)                        
 yaml          2.2.0      2018-07-25 CRAN (R 3.5.1)                        
 zoo           1.8-3      2018-07-16 CRAN (R 3.5.1)

Add broom functionality

• augment
• tidy
• glance

Update seasonal periods to use fablelite::get_frequencies

Support transformations in the formula specification

MRE:

fasster(USAccDeaths * 1000 ~ poly(1) + trig(12))

Implement automatic ARMA coefficient specification

Heuristic ideas

• Filter smooth and get state variances
• lm on first terms with exact model specification, rescale variance based on saturated model specification

Reproducible issues of the examples from the README

I'm interested to try this package. But I can't reproduce the examples from the README. It seems it's due to the changed dataset from tsibbledata and maybe also some api changes. Could you update the tutorial? And I tried the following example. I got the error like "Error in attr(data, "tsp") <- c(start, end, frequency) : object is not a matrix". Any thoughts?

library(fasster)
library(fable)
library(tidyverse)
library(forecast)
tbl_taylor <- as_tsibble(taylor) %>%
  mutate(index = seq(ymd_h("2000-6-5 00"), by="30 mins", length.out=length(taylor)),
         DayType = ifelse(wday(index) %in% 2:6, "Weekday", "Weekend")) %>% 
  filter(month(index) == 8)
tbl_taylor %>%
  filter(index < ymd("2000-8-21")) %>%
  fasster(taylor ~ DayType %S% (poly(1) + trig(48, 16))) %>%
  fablelite::forecast(h=40) %>%
  autoplot() + 
  ylab("Electricity Demand (Megawatts)") + 
  xlab("Time (5 June 2000 - 27 August 2000)")

Add super fast option which doesn't filter through the first few observations