Let's talk about Forecasting!

• Time Series data is everywhere! (especially in Business Settings)
• Forecasting ≈ Zeitreihenanalyse + Invariance Hypothesis
• Forecasting for S&OP-Planning
• Classical Stuff:ETS, Arima, Decompositions

►► Robust, automatic forecasting is essential in practice!

Forecasting-Tools in R?

• CRAN Task View: cran.r-project.org/web/views/TimeSeries.html
• -> Dates, classes, univariate timeseries, regression, ...
• Different ts (xts, timeseries)-objects not standardized!

A great introduction

Application: Predicting Gas Prices

• from https://creativecommons.tankerkoenig.de/
• wouldn't it be nice to know what the sprit price is going to be next week?

Data-Exploration 1

• Trend
• Seasonality

Gas Prices over time

Data-Exploration 2

https://anomaly.io/seasonal-trend-decomposition-in-r/

Data-Exploration 3

Data-Exploration 4

Nesting a dataframe*

ets.models <- ets.params %>%
    # Add in the training set and the modelling function.
    mutate(fn = replicate(forecast::ets, n = n()),
           train = replicate(list(train), n = n())) %>%
    # Create a "param" column to pass to `fn`.
    mutate(params = purrr::transpose(list(
        "y" = train, "model" = model, "damped" = damped
    ))) %>%
    select(kind, desc, train, fn, params)

#> # A tibble: 2 x 5
#>    kind    desc     train     fn     params
#>   <chr>   <chr>    <list> <list>     <list>
#> 1   ETS (M,N,M) <S3: xts>  <fun> <list [3]>
#> 2   ETS (M,N,N) <S3: xts>  <fun> <list [3]>

Fitting many models

forecast_all <- models_all %>%
  mutate(fit = purrr::invoke_map(fn, params),
         forecast = map(fit, forecast::forecast, h = h))

#> # A tibble: 6 x 8
#>     kind           desc   model     fn     params     train
#>    <chr>          <chr>   <chr> <list>     <list>    <list>
#> 1  naive          Naive default  <fun> <list [2]>    <NULL>
#> 2 snaive Seasonal Naive default  <fun> <list [2]>    <NULL>
#> 3    ETS        (M,N,M)    <NA>  <fun> <list [3]> <S3: xts>
#> 4    ETS        (M,N,N)    <NA>  <fun> <list [3]> <S3: xts>
#> 5  ARIMA     auto.arima    <NA>  <fun> <list [1]> <S3: xts>
#> 6 nnetar          Neuro default  <fun> <list [1]>    <NULL>
#> # ... with 2 more variables: fit <list>, forecast <list>

Model Evaluation 1👌

forecast_all %>%
    mutate(test = replicate(list(test), n = n()),
           glance = map2(forecast, test, broom::glance))

#> # A tibble: 6 x 8
#>     kind           desc  is_rmse oos_rmse    is_mae  oos_mae   is_mape
#>    <chr>          <chr>    <dbl>    <dbl>     <dbl>    <dbl>     <dbl>
#> 1 snaive Seasonal Naive 16.94010 19.35697 10.655327 15.10381 0.9483623
#> 2    ETS        (M,N,M) 10.60398 21.10344  6.787350 17.43475 0.5934255
#> 3 nnetar          Neuro  8.93484 35.58820  5.133682 29.83273 0.4562150
#> 4  ARIMA     auto.arima 13.69798 58.15478  7.037829 45.87266 0.6229313
#> 5    ETS        (M,N,N) 20.71179 70.56093  8.197112 57.80277 0.7243085
#> 6  naive          Naive 20.71209 70.56093  8.178921 57.80277 0.7227123
#> # ... with 1 more variables: oos_mape <dbl>

Model Evaluation 2

Summary/ Discussion?

• What potential does this appoach have?

• What else could be done this way?

• How good is R as a forecasting tool? What is good about it? Where are challenges?

Forecasting challenges in practice

• many timeseries
• intermittend demand
• forecast value add-analysis

• error measures ("Zerrspiegel")

• "Once you have selected the right model, forecasting becomes easy!" (Prof. Stoffer)

What else is out there?

References

• github-repo for this talk: github.com/user/foRcasting
• Managing many models with R - Hadley Wickham (Youtube)

Data Science @ Iqast