How to Run Multiple Synthetic Forecasts with Furrr • SynthCast

This is a walk through of how the main package function SynthCast::run_synthetic_forecast() can be used with the furrr package. The goal of furrr is to combine purrr’s family of mapping functions with future’s parallel processing capabilities. Therefore, the objective of this article is to show how you could run multiple forecasts in a parallel manner.

Please fell free to contribute if you think of a batter way to do so! If you want see an usage example of the function SynthCast::run_synthetic_forecast() see the article How to Run a Synthetic Forecast.

Lets load the packages we will need:

library(dplyr)
library(purrr)
library(future)
library(furrr)

The Dataset

The first thing that a forecast needs a data to be forecasted. The SynthCast provides a example of how it expected a dataset to look like, the code bellow loads the package and the example dataset:

library(knitr)
library(SynthCast)
data('df_example')
kable(head(df_example))

unit	time_period	x1	x2	x3	x4	x5	x6	x7	x8	x9	x10	x11	x12	x13	x14	x15	x16	x17	x18	x19	x20	x21	x22	x23	x24	x25	x26	x27	x28
1	1	0.4279268	0.2329316	0.4531898	0.5010649	0.0140657	0.5	0.0103704	0.0126492	0.0061209	0.0016722	0.0020701	0.0229175	0.1717596	0.0028440	0.2961483	0.2777202	0.0179579	0.5	0.0186335	0.0196256	0.0140659	0.5	0.0191083	0.0193874	0.0280014	0.5	0.0062926	0.0193874
1	2	0.3923215	0.0661752	0.4300946	0.4639223	0.1523873	0.5	0.0167901	0.1340623	0.0940312	0.0016722	0.0063536	0.0896040	0.1362349	0.0028440	0.2961483	0.2352990	0.1657939	0.5	0.1428571	0.1479287	0.1589145	0.5	0.1974522	0.1750037	0.1949374	0.5	0.0181592	0.1750037
1	3	0.4420440	0.1649872	0.4336537	0.5034269	0.2919640	0.5	0.0395062	0.2602215	0.1796289	0.0016722	0.0137895	0.1695727	0.1045988	0.0028440	0.2961483	0.2088865	0.3180237	0.5	0.3167702	0.2890312	0.3442300	0.5	0.3949045	0.3201550	0.2198580	0.5	0.0167533	0.3201550
1	4	0.4545717	0.1076923	0.4433019	0.5427364	0.4315704	0.5	0.0501235	0.3791298	0.2685505	0.0016722	0.0172917	0.2420208	0.0822586	0.0028440	0.2961483	0.1556901	0.4694968	0.5	0.4223602	0.4250857	0.5346481	0.5	0.5859873	0.4600435	0.2291281	0.5	0.0072638	0.4600435
1	5	0.4223203	0.1391912	0.4767905	0.5474351	0.5673960	0.5	0.0501235	0.4999604	0.3522328	0.1638796	0.0279551	0.3139178	0.0689121	0.2787148	0.0835851	0.1119981	0.6177005	0.5	0.6149068	0.5627327	0.7247700	0.5	0.7834395	0.5979929	0.2351954	0.5	0.0072638	0.5979929
1	6	0.3827364	0.1078405	0.5021293	0.5456524	0.6992290	0.5	0.0688889	0.6161397	0.4334900	0.3311037	0.0335161	0.3829171	0.0602702	0.2787148	0.0835851	0.0985164	0.7600335	0.5	0.7826087	0.6957559	0.9102858	0.5	0.9745223	0.7413431	0.2458748	0.5	0.0072638	0.7413431

The dataset is expected to have 3 types of columns:

1. A unit column: containing a numeric identification of the unit. In the credit card example this could the the customer, a group of customer, etc.,;
1. A time columns: containing the time in integer. In the credit card example this would be the age in months of the respective unit (say 1 for first month, 2 for the second month, etc.,);
1. Feature Columns: Numeric features, with both the serie(s) that will be forecasted as well as features to use to forecast. In the credit card this could be the profitability and transactional features.

Setting Up the Requirements

The solution that I present here consists of creating another dataset that will contain the parameters as columns and each row the combination of parameters.

So lets work with a practical example: Lets supose we want to forecasts 12 and 40 time periods for the units 30 and 40, for the columns x1 and x3.

Dataframe with Parameters

The first thing that we need is a dataframe with the parameters as columns, with the exeception of the dataset itself:



## parametros variaveis
units_of_interest <- c(30,40)
series_of_interest <- c('x1', 'x3')
periods_to_forecast = c(12, 20) 

col_time = c('time_period')
col_unit_name = c('unit')
periods_to_forecast = tidyr::tibble(
  unit_of_interest = units_of_interest,
  periods_to_forecast = periods_to_forecast
)

crossArg <- cross_df(
  list(
    unit_of_interest = units_of_interest,
    serie_of_interest = series_of_interest,
    col_time=col_time,
    col_unit_name=col_unit_name
  )
) %>%
  dplyr::left_join(periods_to_forecast, by='unit_of_interest')

kable(crossArg)

unit_of_interest	serie_of_interest	col_time	col_unit_name	periods_to_forecast
30	x1	time_period	unit	12
40	x1	time_period	unit	20
30	x3	time_period	unit	12
40	x3	time_period	unit	20

Wrapper Function

The second thing that we need is to define a function that will wrap the SynthCast::run_synthetic_forecast(). This is done because in order to map over multiple parameters with the furrr package with need the parameters in a dataframe. But we cannot put a dataframe in a column, without using column lists.

Note that the object df_example that is called inside f() is a global parameter defined above. This is a workaround, but gets the job done.

f = function(col_unit_name, unit_of_interest, col_time, periods_to_forecast, serie_of_interest){
  synth_forecast = SynthCast::run_synthetic_forecast(
    df=as.data.frame(df_example), # Global environment
    col_unit_name=col_unit_name,
    unit_of_interest=unit_of_interest,
    col_time=col_time,
    periods_to_forecast=periods_to_forecast,
    serie_of_interest=serie_of_interest
  )
  return(synth_forecast)
}

Running the Forecasts Parallelized

Setup the parallel backend, as indicated in the furrr documentation:

future::plan(multisession) # Setup to use multiple core

Now we can use the furrr:future_pmap() to run the forecasts:

synthetic_forecasts <- furrr::future_pmap(
  crossArg,
  f
)
#> [1] "Forecasting Unit:  30 . Serie:  x1"
#> 
#> X1, X0, Z1, Z0 all come directly from dataprep object.
#> 
#> 
#> **************** 
#>  searching for synthetic control unit  
#>  
#> 
#> **************** 
#> **************** 
#> **************** 
#> 
#> MSPE (LOSS V): 0.005105562 
#> 
#> solution.v:
#>  0.03795838 0.02953412 0.03356642 0.01533716 0.1226315 0.1285906 0.05816525 0.02318678 0.01465216 0.01080646 0.06187415 0.0289542 0.01702719 0.08006876 0.009607601 0.01627082 0.1278952 0.02615566 0.01342692 0.04431671 0.04468165 0.01097563 0.04431671 
#> 
#> solution.w:
#>  1.1452e-05 0.0002666085 0.0001873182 0.0002686277 0.0001636778 0.0003347625 0.0004905744 0.0005939929 0.0005203545 0.5502766 4.8739e-06 0.0007708196 0.0003844661 0.001002508 0.000803214 0.000999687 0.1285913 0.3143298 
#> 
#> [1] "Forecasting Unit:  40 . Serie:  x1"
#> 
#> X1, X0, Z1, Z0 all come directly from dataprep object.
#> 
#> 
#> **************** 
#>  searching for synthetic control unit  
#>  
#> 
#> **************** 
#> **************** 
#> **************** 
#> 
#> MSPE (LOSS V): 0.01126689 
#> 
#> solution.v:
#>  0.00451887 0.1198275 0.001215545 0.01511271 0.0005242874 0.0207889 0.1270972 0.03638089 0.1136289 0.07111816 0.06962764 0.03914189 0.00266398 0.003293383 0.003680319 0.002879397 0.0001513126 0.1898451 0.003094035 0.0532468 0.06650722 0.002409148 0.0532468 
#> 
#> solution.w:
#>  2.892e-07 7.7e-09 8.85e-08 2.97e-08 2.778e-07 1.147e-07 1.083e-07 3.687e-07 1.0996e-06 6.876e-07 2.4978e-06 0.0002996523 0.00148343 0.4549331 0.0003138607 0.001220969 0.526065 0.0005761354 5.26296e-05 0.01504971 
#> 
#> [1] "Forecasting Unit:  30 . Serie:  x3"
#> 
#> X1, X0, Z1, Z0 all come directly from dataprep object.
#> 
#> 
#> **************** 
#>  searching for synthetic control unit  
#>  
#> 
#> **************** 
#> **************** 
#> **************** 
#> 
#> MSPE (LOSS V): 0.004444729 
#> 
#> solution.v:
#>  0.0008966688 0.05006106 0.04589722 0.01215437 0.1364756 0.1380885 0.07131376 0.02247537 0.06353448 0.001899661 0.05256282 0.02746854 0.02928654 0.06565559 0.001513901 0.01546533 0.1404815 0.008559098 0.0178953 0.02120212 0.04279917 0.01311125 0.02120212 
#> 
#> solution.w:
#>  1.11103e-05 0.0004018295 0.000251307 0.0004156746 0.0002930494 0.0004620298 0.0007838085 0.001376503 0.0006554179 0.5237144 0.001810757 0.001349141 0.0005160587 0.002039712 0.001832605 0.002641334 0.03367259 0.4277733 
#> 
#> [1] "Forecasting Unit:  40 . Serie:  x3"
#> 
#> X1, X0, Z1, Z0 all come directly from dataprep object.
#> 
#> 
#> **************** 
#>  searching for synthetic control unit  
#>  
#> 
#> **************** 
#> **************** 
#> **************** 
#> 
#> MSPE (LOSS V): 0.0003451711 
#> 
#> solution.v:
#>  0.0838238 0.01402602 0.02662944 0.0001756046 0.08158252 0.04514179 0.03369849 0.0432683 0.04256028 0.01812175 5.0628e-05 0.04334514 0.1161261 0.0002120456 6.248e-05 0.01880793 0.05472154 0.03005398 1.5966e-06 0.142656 0.0129223 0.04935617 0.142656 
#> 
#> solution.w:
#>  4.158e-07 1.9116e-06 4.2001e-06 3.7943e-06 0.05303326 0.02782933 0.0003817437 1.66687e-05 1.06607e-05 4.89541e-05 8.04749e-05 3.8502e-06 8.695e-07 0.6799389 2.1e-08 3.14e-06 0.001466123 0.151224 0.08586669 8.49833e-05

The synthetic_forecasts object is a list of output from SynthCast::run_synthetic_forecast().

Bind Result Tables

We can now bind the tables from the different forecasts together:

mape_backtest = bind_rows(lapply(synthetic_forecasts, function(x) x$mape_backtest))

kable(mape_backtest)

execution_date	projected_unit	projected_serie	max_time_unit_of_interest	periods_to_forecast	elegible_control_units	number_control_units	mape
2022-03-08	30	x1	21	12	17	9	13.009279
2022-03-08	40	x1	11	20	19	6	21.482920
2022-03-08	30	x3	21	12	17	12	7.898514
2022-03-08	40	x3	11	20	19	6	3.090236