One of the most powerful benefits of sweep is that it helps forecasting at scale within the “tidyverse”. There are two common situations:

In this vignette we’ll review how sweep can help the second situation: Applying multiple models to a time series.

Prerequisites

Before we get started, load the following packages.

library(forecast)
library(tidyquant)
library(timetk)
library(sweep)

Forecasting Gasoline Prices

To start, let’s get some data from the FRED data base using tidyquant. We’ll use tq_get() to retrieve the Gasoline Prices from 1990 through today (2017-07-25).

gas_prices_monthly_raw <- tq_get(
    x    = "GASREGCOVM", 
    get  = "economic.data", 
    from = "1990-01-01", 
    to   = "2016-12-31") 
gas_prices_monthly_raw

## # A tibble: 316 x 2
##          date price
##        <date> <dbl>
##  1 1990-09-01 1.258
##  2 1990-10-01 1.335
##  3 1990-11-01 1.324
##  4 1990-12-01    NA
##  5 1991-01-01    NA
##  6 1991-02-01 1.094
##  7 1991-03-01 1.040
##  8 1991-04-01 1.076
##  9 1991-05-01 1.126
## 10 1991-06-01 1.128
## # ... with 306 more rows

Upon a brief inspection, the data contains 2 NA values that will need to be dealt with.

summary(gas_prices_monthly_raw$price)

##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
##   0.900   1.138   1.615   1.974   2.697   4.002       2

We can use the fill() from the tidyr package to help deal with these data. We first fill down and then fill up to use the previous and then post days prices to fill in the missing data.

gas_prices_monthly <- gas_prices_monthly_raw %>%
    fill(price, .direction = "down") %>%
    fill(price, .direction = "up")

We can now visualize the data.

gas_prices_monthly %>%
    ggplot(aes(x = date, y = price)) +
    geom_line(color = palette_light()[[1]]) +
    geom_point(color = palette_light()[[1]]) +
    labs(title = "Gasoline Prices, Monthly", x = "", y = "USD") +
    scale_y_continuous(labels = scales::dollar) +
    theme_tq()

Monthly periodicity might be a bit granular for model fitting. We can easily switch periodicity to quarterly using tq_transmute() from the tidyquant package along with the periodicity aggregation function to.period from the xts package. We’ll convert the date to yearqtr class which is regularized.

gas_prices_quarterly <- gas_prices_monthly %>%
    tq_transmute(mutate_fun = to.period, period = "quarters") 
gas_prices_quarterly

## # A tibble: 106 x 2
##          date price
##        <date> <dbl>
##  1 1990-09-01 1.258
##  2 1990-12-01 1.324
##  3 1991-03-01 1.040
##  4 1991-06-01 1.128
##  5 1991-09-01 1.109
##  6 1991-12-01 1.076
##  7 1992-03-01 1.013
##  8 1992-06-01 1.145
##  9 1992-09-01 1.122
## 10 1992-12-01 1.078
## # ... with 96 more rows

Another quick visualization to show the reduction in granularity.

gas_prices_quarterly %>%
    ggplot(aes(x = date, y = price)) +
    geom_line(color = palette_light()[[1]], size = 1) +
    geom_point(color = palette_light()[[1]]) +
    labs(title = "Gasoline Prices, Quarterly", x = "", y = "USD") +
    scale_y_continuous(labels = scales::dollar) +
    scale_x_date(date_breaks = "5 years", date_labels = "%Y") +
    theme_tq()

Performing Forecasts Using Multiple Models

In this section we will use three models to forecast gasoline prices:

ARIMA
ETS
BATS

Multiple Models Concept

Before we jump into modeling, let’s take a look at the multiple model process from R for Data Science, Chapter 25 Many Models. We first create a data frame from a named list. The example below has two columns: “f” the functions as text, and “params” a nested list of parameters we will pass to the respective function in column “f”.

df <- tibble(
  f = c("runif", "rpois", "rnorm"),
  params = list(
    list(n = 10),
    list(n = 5, lambda = 10),
    list(n = 10, mean = -3, sd = 10)
  )
)
df

## # A tibble: 3 x 2
##       f     params
##   <chr>     <list>
## 1 runif <list [1]>
## 2 rpois <list [2]>
## 3 rnorm <list [3]>

We can also view the contents of the df$params column to understand the underlying structure. Notice that there are three primary levels and then secondary levels containing the name-value pairs of parameters. This format is important.

df$params

## [[1]]
## [[1]]$n
## [1] 10
## 
## 
## [[2]]
## [[2]]$n
## [1] 5
## 
## [[2]]$lambda
## [1] 10
## 
## 
## [[3]]
## [[3]]$n
## [1] 10
## 
## [[3]]$mean
## [1] -3
## 
## [[3]]$sd
## [1] 10

Next we apply the functions to the parameters using a special function, invoke_map(). The parameter lists in the “params” column are passed to the function in the “f” column. The output is in a nested list-column named “out”.

df_out <- df %>% 
    mutate(out = invoke_map(f, params))
df_out

## # A tibble: 3 x 3
##       f     params        out
##   <chr>     <list>     <list>
## 1 runif <list [1]> <dbl [10]>
## 2 rpois <list [2]>  <int [5]>
## 3 rnorm <list [3]> <dbl [10]>

And, here’s the contents of “out”, which is the result of mapping a list of functions to a list of parameters. Pretty powerful!

df_out$out

## [[1]]
##  [1] 0.59102984 0.20873637 0.87339386 0.24989660 0.24076947 0.20023536
##  [7] 0.81866094 0.88582756 0.91040161 0.07579531
## 
## [[2]]
## [1]  6 14 13 17  9
## 
## [[3]]
##  [1]  11.645398   2.280693  12.513776   5.753345 -10.439134   2.414632
##  [7] -27.233099  11.065930   7.532832  -9.286955

Take a minute to understand the conceptual process of the invoke_map function and specifically the parameter setup. Once you are comfortable, we can move on to model implementation.

Multiple Model Implementation

We’ll need to take the following steps to in an actual forecast model implementation:

Coerce the data to time series
Build a model list using nested lists
Create the the model data frame
Invoke a function map

This is easier than it sounds. Let’s start by coercing the univariate time series with tk_ts().

gas_prices_quarterly_ts <- gas_prices_quarterly %>% 
    tk_ts(select = -date, start = c(1990, 3), freq = 4)
gas_prices_quarterly_ts

##       Qtr1  Qtr2  Qtr3  Qtr4
## 1990             1.258 1.324
## 1991 1.040 1.128 1.109 1.076
## 1992 1.013 1.145 1.122 1.078
## 1993 1.052 1.097 1.050 1.014
## 1994 1.008 1.078 1.144 1.060
## 1995 1.059 1.186 1.108 1.066
## 1996 1.129 1.243 1.200 1.233
## 1997 1.197 1.189 1.216 1.119
## 1998 1.014 1.048 0.994 0.923
## 1999 0.961 1.095 1.239 1.261
## 2000 1.498 1.612 1.525 1.418
## 2001 1.384 1.548 1.506 1.072
## 2002 1.221 1.341 1.363 1.348
## 2003 1.636 1.452 1.616 1.448
## 2004 1.689 1.910 1.841 1.800
## 2005 2.063 2.123 2.862 2.174
## 2006 2.413 2.808 2.501 2.284
## 2007 2.503 3.024 2.817 2.984
## 2008 3.215 3.989 3.709 1.669
## 2009 1.937 2.597 2.480 2.568
## 2010 2.742 2.684 2.678 2.951
## 2011 3.509 3.628 3.573 3.220
## 2012 3.774 3.465 3.801 3.256
## 2013 3.648 3.576 3.474 3.209
## 2014 3.474 3.626 3.354 2.488
## 2015 2.352 2.700 2.275 1.946
## 2016 1.895 2.303 2.161 2.192

Next, create a nested list using the function names as the first-level keys (this is important as you’ll see in the next step). Pass the model parameters as name-value pairs in the second level.

models_list <- list(
    auto.arima = list(
        y = gas_prices_quarterly_ts
        ),
    ets = list(
        y = gas_prices_quarterly_ts,
        damped = TRUE
    ),
    bats = list(
        y = gas_prices_quarterly_ts
    )
)

Now, convert to a data frame using the function, enframe() that turns lists into tibbles. Set the arguments name = "f" and value = "params". In doing so we get a bonus: the model names are the now convieniently located in column “f”.

models_tbl <- enframe(models_list, name = "f", value = "params")
models_tbl

## # A tibble: 3 x 2
##            f     params
##        <chr>     <list>
## 1 auto.arima <list [1]>
## 2        ets <list [2]>
## 3       bats <list [1]>

We are ready to invoke the map. Combine mutate() with invoke_map() as follows. Bada bing, bada boom, we now have models fitted using the parameters we defined previously.

models_tbl_fit <- models_tbl %>%
    mutate(fit = invoke_map(f, params))
models_tbl_fit

## # A tibble: 3 x 3
##            f     params         fit
##        <chr>     <list>      <list>
## 1 auto.arima <list [1]> <S3: ARIMA>
## 2        ets <list [2]>   <S3: ets>
## 3       bats <list [1]>  <S3: bats>

Inspecting the Model Fit

It’s a good point to review and understand the model output. We can review the model parameters, accuracy measurements, and the residuals using sw_tidy(), sw_glance(), and sw_augment().

sw_tidy

The tidying function returns the model parameters and estimates. We use the combination of mutate and map to iteratively apply the sw_tidy() function as a new column named “tidy”. Then we unnest and spread to review the terms by model function.

models_tbl_fit %>%
    mutate(tidy = map(fit, sw_tidy)) %>%
    unnest(tidy) %>%
    spread(key = f, value = estimate)

## # A tibble: 16 x 4
##                 term auto.arima          bats          ets
##  *             <chr>      <dbl>         <dbl>        <dbl>
##  1             alpha         NA  8.159321e-01 0.6770332202
##  2   ar.coefficients         NA            NA           NA
##  3                 b         NA            NA 0.0592275021
##  4              beta         NA            NA 0.0129535504
##  5 damping.parameter         NA            NA           NA
##  6             gamma         NA            NA 0.0001718402
##  7      gamma.values         NA -2.019471e-02           NA
##  8                 l         NA            NA 1.0436968053
##  9            lambda         NA  1.549104e-08           NA
## 10   ma.coefficients         NA            NA           NA
## 11               phi         NA            NA 0.8000459550
## 12                s0         NA            NA 1.0463243371
## 13                s1         NA            NA 0.9764888261
## 14                s2         NA            NA 0.9439714264
## 15              sar1  0.9551837            NA           NA
## 16              sma1 -0.8299847            NA           NA

sw_glance

Glance is one of the most powerful tools because it yields the model accuracies enabling direct comparisons between the fit of each model. We use the same process for used for tidy, except theres no need to spread to perform the comparison. We can see that the ARIMA model has the lowest AIC by far.

models_tbl_fit %>%
    mutate(glance = map(fit, sw_glance)) %>%
    unnest(glance, .drop = TRUE)

## # A tibble: 3 x 13
##            f             model.desc     sigma    logLik       AIC
##        <chr>                  <chr>     <dbl>     <dbl>     <dbl>
## 1 auto.arima ARIMA(0,1,0)(1,0,1)[4] 0.3024768 -23.33351  52.66701
## 2        ets            ETS(M,Ad,M) 0.1117591 -75.41773 170.83545
## 3       bats BATS(0, {0,0}, -, {4}) 0.1163602 160.49138 176.49138
## # ... with 8 more variables: BIC <dbl>, ME <dbl>, RMSE <dbl>, MAE <dbl>,
## #   MPE <dbl>, MAPE <dbl>, MASE <dbl>, ACF1 <dbl>

sw_augment

We can augment the models to get the residuals following the same procedure. We can pipe (%>%) the results right into ggplot() for plotting. Notice the ARIMA model has the largest residuals especially as the model index increases whereas the bats model has relatively low residuals.

models_tbl_fit %>%
    mutate(augment = map(fit, sw_augment, rename_index = "date")) %>%
    unnest(augment) %>%
    ggplot(aes(x = date, y = .resid, group = f)) +
    geom_line(color = palette_light()[[2]]) +
    geom_point(color = palette_light()[[1]]) +
    geom_smooth(method = "loess") +
    facet_wrap(~ f, nrow = 3) +
    labs(title = "Residuals Plot") +
    theme_tq()

Forecasting the model

Creating the forecast for the models is accomplished by mapping the forecast function. The next six quarters are forecasted withe the argument h = 6.

models_tbl_fcast <- models_tbl_fit %>%
    mutate(fcast = map(fit, forecast, h = 6))
models_tbl_fcast

## # A tibble: 3 x 4
##            f     params         fit          fcast
##        <chr>     <list>      <list>         <list>
## 1 auto.arima <list [1]> <S3: ARIMA> <S3: forecast>
## 2        ets <list [2]>   <S3: ets> <S3: forecast>
## 3       bats <list [1]>  <S3: bats> <S3: forecast>

Tidying the forecast

Next, we map sw_sweep, which coerces the forecast into the “tidy” tibble format. We set fitted = FALSE to remove the model fitted values from the output. We set timetk_idx = TRUE to use dates instead of numeric values for the index.

models_tbl_fcast_tidy <- models_tbl_fcast %>%
    mutate(sweep = map(fcast, sw_sweep, fitted = FALSE, timetk_idx = TRUE, rename_index = "date"))
models_tbl_fcast_tidy

## # A tibble: 3 x 5
##            f     params         fit          fcast              sweep
##        <chr>     <list>      <list>         <list>             <list>
## 1 auto.arima <list [1]> <S3: ARIMA> <S3: forecast> <tibble [112 x 7]>
## 2        ets <list [2]>   <S3: ets> <S3: forecast> <tibble [112 x 7]>
## 3       bats <list [1]>  <S3: bats> <S3: forecast> <tibble [112 x 7]>

We can unnest the “sweep” column to get the results of all three models.

models_tbl_fcast_tidy %>%
    unnest(sweep)

## # A tibble: 336 x 8
##             f       date    key price lo.80 lo.95 hi.80 hi.95
##         <chr>     <date>  <chr> <dbl> <dbl> <dbl> <dbl> <dbl>
##  1 auto.arima 1990-09-01 actual 1.258    NA    NA    NA    NA
##  2 auto.arima 1990-12-01 actual 1.324    NA    NA    NA    NA
##  3 auto.arima 1991-03-01 actual 1.040    NA    NA    NA    NA
##  4 auto.arima 1991-06-01 actual 1.128    NA    NA    NA    NA
##  5 auto.arima 1991-09-01 actual 1.109    NA    NA    NA    NA
##  6 auto.arima 1991-12-01 actual 1.076    NA    NA    NA    NA
##  7 auto.arima 1992-03-01 actual 1.013    NA    NA    NA    NA
##  8 auto.arima 1992-06-01 actual 1.145    NA    NA    NA    NA
##  9 auto.arima 1992-09-01 actual 1.122    NA    NA    NA    NA
## 10 auto.arima 1992-12-01 actual 1.078    NA    NA    NA    NA
## # ... with 326 more rows

Finally, we can plot the forecasts by unnesting the “sweep” column and piping to ggplot().

models_tbl_fcast_tidy %>%
    unnest(sweep) %>%
    ggplot(aes(x = date, y = price, color = key, group = f)) +
    geom_ribbon(aes(ymin = lo.95, ymax = hi.95), 
                fill = "#D5DBFF", color = NA, size = 0) +
    geom_ribbon(aes(ymin = lo.80, ymax = hi.80, fill = key), 
                fill = "#596DD5", color = NA, size = 0, alpha = 0.8) +
    geom_line(size = 1) +
    facet_wrap(~f, nrow = 3) +
    labs(title = "Gasoline Price Forecasts",
         subtitle = "Forecasting multiple models with sweep: ARIMA, BATS, ETS",
         x = "", y = "Price") +
    scale_y_continuous(labels = scales::dollar) +
    scale_x_date(date_breaks = "5 years", date_labels = "%Y") +
    theme_tq() +
    scale_color_tq()

Forecasting Using Multiple Models

Matt Dancho

2017-07-25