In aflatoune/etalonnage: Simplify Macroeconomic Forecasting

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

This document shows how to use the etalonnage package to facilitate GDP forecasting/nowcasting using real French data.

library(etalonnage)

Data

First, load a dataframe containing regressors (fr_x) and another one containing the target (fr_y). To emulate a real forecast situation fr_x covers a wider period (one more quarter) than fr_y.

data("fr_x", "fr_y")

fr_x is composed of:

1. monthly survey indicators in three sectors (services, industry, construction, retail trade)
2. "hard" series (household consumption, industrial production index - IPI).

head(fr_x, n = 3)

fr_y is composed of real GDP values at a quarterly frequency.

head(fr_y, n =3)

Processing

During a forecasting exercise, the preprocessing to be done is often the same (set the predictors to the same frequency as the target, add dummies, ...). The package etalonnage provides functions to facilitate the realization of these treatments.

Convert the target to growth rate using build_target:

fr_y <- build_target(fr_y,
                     growth_rate = TRUE,
                     date_freq = "quarter")
head(fr_y, n = 3)

Add all regressors first-diff to fr_x using add_diff and pivot regressors to have one column for each month in a quarter (and thus match the frequency of the target) using month_to_quarter:

fr_x <- fr_x %>%
    add_diff(exclude = "date") %>%
    month_to_quarter()

Now suppose that one wants to forecast the French GDP growth rate in 2000Q2. Depending on the horizon, the forecast is computed conditional on the information released until April 2000, May 2000 or June 2000. Consider June 2000. At this horizon, 3 months (April 2000, May 2000 and June 2000) of Insee survey variables are available since these variables are released with no delay (i.e. during the month to which they relate) but only 2 months (April 2000 and May 2000) of Banque de France survey variables are available since these variables are released with a delay of 1 month. As a result the final dataset contains some values that shouldn't be observed and must be dropped. Regarding household consumption and IPI, only one month is available but it would be too costly to remove these variables given their importance to the forecast. To deal with these, etalonnage package contains a function acquis that transform series with NA by computing a "granted" growth (acquis de croissance).

fr_x <- fr_x %>%
  acquis(cols = c("ipi", "conso"), month = 1) %>%
  dplyr::select(-dplyr::contains(c("Bdf_fd1_3", "Bdf_3")))

To deal with NA or structural breaks, it is common to add dummies to the regressors. This can be done using add_dummy: pass a list of names for the dummies and add the corresponding conditions one-by-one. Here, the columns retailInsee and batBdf have NA respectively until 1999 and 2009Q4:

fr_x <- fr_x %>%
    add_dummy(
        names = list("dummy_retailInsee", "dummy_batBdf"),
        (date < "1999-01-01"),
        (date < "2009-04-01")
    )

It is also possible to convert columns values to growth rate using to_growth_rate.

Make sure that fr_x and fr_y starts at the same date and replace NA with some value (here 0):

fr_y <- fr_y[-1,]
fr_x <- fr_x[-c(1,2),]
fr_x[is.na(fr_x)] <- 0

Forecasting

During a forecast exercise, it is common that the forecast is performed using information provided by series released at a higher frequency than the target. These series are released with various delays so that the forecast is conditioned on the sample of series that are known at the time the estimation is performed. In order to take into account the non-synchronicity of data publications (and thus to properly assess the performances of a amodel), the forecast accuracy is assessed on the basis of a pseudo real-time experiment. This kind of evaluation aims at replicating the timeliness of the releases of the series by taking into account their publications lags. In this framework, the series are truncated in order to consider only those values of the series that would have been available on the date on which the forecasts were calculated.

Validation scheme

When data are not i.i.d., the validation scheme has to take into consideration the time dependent structure of the data to avoid the creation of non-independent training and test sets. To assess a model performances, etalonnage package implements "rolling-origin-update evaluation" (ROUE), meaning that the forecast origin rolls ahead in time. At each step, ROUE increments the traning set by one observation of the test set. Here, ROUE is implemented in such a way that the size of the training set increases at each iteration (expanding window) rather than remaining constant (fixed window). In doing so, all the available information is used but equal importance is given to all observations of the training set, regardless of their "distance" to the forecast origin. At the end of the validation, a set of forecasts is available, making it possible to compare models.

Models

All that remains is to choose a forecast origin and fit the models:

rf <- etalonnage(
    name = "Random Forest",
    X = fr_x,
    y = fr_y$target,
    regressor = "randomForest",
    forecast_origin = "2014-10-01",
    scale = "none",
    mtry = 15,
    ntree = 500,
    nodesize = 3,
    importance = TRUE
)

xgb <- etalonnage(
    name = "XGBoost",
    X = fr_x,
    y = fr_y$target,
    regressor = "xgboost",
    forecast_origin = "2014-10-01",
    scale = "none",
    nrounds = 1500,
    eta = 0.05,
    max_depth = 6,
    verbose = FALSE
)

For each of the two models, this is what is done by etalonnage:

1. Drop column date from the regressors,
2. Keep all data until forecast_origin,
3. If scale != "none", process the data,
4. Fit a model of type regressor on the data,
5. Predict a value for the growth rate at forecast_origin + 1 quarter,
6. Repeat steps from step 2 to step 5 until the last raw in fr_x is reached.

Other arguments like mtry or eta come directly from the packages used to fit the models, i.e.:

1. randomForest for regressor = "randomForest",
2. xgboost for regressor = "xgboost",
3. glmnet for regressor = "glmnet".

Plot a model predictions using graph method:

graph(rf, annotation_y = -0.01, annotation_x = 200)

Directly access to the predicted values using rf$predicted_values or valuate the models using their attributes:

rf$test_rmse
rf$test_mae
rf$test_mda

xgb$test_rmse
xgb$test_mae
xgb$test_mda

Graph the two models using graph_models:

graph_models(rf, xgb, start_graph = "2000-01-01")

aflatoune/etalonnage documentation built on May 30, 2022, 1:57 p.m.