CEWP 21-05

Forecasting Canadian GDP Growth with Machine

Learning

Shafiullah Qureshi Ba Chu Fanny S. Demers

Carleton University Carleton University Carleton University

May 17, 2021

CARLETON ECONOMICS WORKING PAPERS

Department of Economics
1125 Colonel By Drive
Ottawa, Ontario, Canada
K1S 5B6
 Forecasting Canadian GDP Growth with Machine Learning
Shafiullah Qureshi* Ba Chu† Fanny S. Demers‡

May 17, 2021

Abstract
This paper applies state-of-the-art machine learning (ML) algorithms to forecast monthly real GDP growth in
Canada by using both Google Trends (GT) data and official macroeconomic data (which are available ahead of
the release of GDP data by Statistics Canada). We show that we can forecast real GDP growth accurately ahead of
the release of GDP figures by using GT and official data (such as employment) as predictors. We first pre-select
features by applying up-to-date techniques, namely, XGBoost’s variable importance score, and a recent variable-
screening procedure for time series data, namely, PDC-SIS+. These pre-selected features are then used to build
advanced ML models for forecasting real GDP growth, by employing tree-based ensemble algorithms, such as
XGBoost, LightGBM, Random Forest, and GBM. We provide empirical evidence that the variables pre-selected
by either PDC-SIS+ or the XGBoost’s variable importance score can have a superior forecasting ability. We find
that the pre-selected GT data features perform as well as the pre-selected official data features with respect to
short-term forecasting ability, while the pre-selected official data features are superior with respect to long-term
forecasting ability.
We also find that (1) the ML algorithms we employ often perform better with a large sample than with a small
sample, even when the small sample has a larger set of predictors; and (2) the Random Forest (that often produces
nonlinear models to capture nonlinear patterns in the data) tends to under-perform a standard autoregressive model
in several cases while there is no clear evidence that the XGBoost and the LightGBM can always outperform each
other.

* Department of Economics, Carleton University, 1125 Colonel By Dr., Ottawa, Ontario, Canada; and Department of Economics, Sector

H-9, Islamabad, Pakistan. Email: [email protected]. Tel: +1.613.520.2600 (ext. 3778).

†
Department of Economics, Carleton University, 1125 Colonel By Dr., Ottawa, Ontario, Canada. Email: [email protected]. Tel: +1
613-520-2600 (ext. 1546).
‡
Department of Economics, Carleton University, 1125 Colonel By Dr., Ottawa, Ontario, Canada. Email: [email protected]. Tel:
+1 613-520-2600 (ext. 3775).

1
 1 Introduction
Gross domestic product (GDP) is the primary measure for assessing the performance of an economy. It enables
policymakers to judge whether the economy is expanding or contracting, and permits them to make appropriate
monetary or fiscal policy decisions accordingly. In this respect, the accurate and timely forecast of GDP growth
ahead of the release of official GDP figures is vital. Both quarterly and monthly GDP data are published with a two-
month delay by Statistics Canada due to multiple revisions. Yet, some other official macroeconomic information,
such as employment, housing starts, and retail trades, often produced by national statistical bureaus, central banks,
or other government institutions are available much sooner in some cases, with only a few days’ delay. In addition,
the use of Google Trend (GT) data for economic forecasting has become popular since the seminal work of Choi
and Varian (2009) and Choi and Varian (2012). GT data has been exploited to predict retail sales, automotive
sales, home sales, and travel volume, among many others. GT allows users to download the time series data for a
keyword search of particular interest in index form by selecting a specific geographic region and a specific time.
Notably, Tkacz (2013) used GT data to predict the 2008-09 recession in Canada with a Probit model, and observed
an enormous surge in the search frequency of the word ‘recession’ about one month before the actual date of the
2008-09 recession. In the context of GDP forecasting, Götz and Knetsch (2019) used GT and other survey data to
forecast German GDP with the Bridge regression.1 They obtained good forecasts when GT data was used instead
of survey data (suggesting that GT data can be a good substitute for survey data when the latter are not available),
but the forecast improvement tended to be smaller when both GT and survey data were used together. Ferrara and
Simoni (2019) also used a Bridge regression to nowcast Euro-area GDP with GT data. They employed Fan and Lv’s
(2008) Sure Independence Screening (SIS) to pre-select GT features as the first step. A Ridge regression was then
used in the second step to construct forecasts given the pre-selected features. They then evaluated their forecasts
using various combinations of data sources, and found GT data to be a useful predictor only before official data
becomes available.
In this paper, we apply state-of-the-art machine learning (ML) algorithms to ‘nowcast’ monthly and quarterly
Canadian GDP growth by using both Google Trends (GT) and official macroeconomic data that are available ahead
of Statistics Canada’s release of GDP data. We hereafter refer to the latter as Official data. (See Table 3 for a list
of the variables we use and their release dates). We show that we can forecast real GDP growth accurately ahead of
the release of GDP figures by using GT data and Official data variables as predictors. In particular, we investigate
whether GT data are still useful for forecasting GDP in the presence of Official variables. As mentioned above,
most previous works on forecasting with GT data found this dataset to be useful when no Official data are available,
and that its importance diminishes as soon as Official data become available (as in e.g., Ferrara and Simoni (2019)
and Götz and Knetsch (2019)). We find that the forecasting accuracy of GT data alone is quite comparable to that
of using both GT and Official data for short term forecasting (such as one- or three-steps ahead forecasts), but that
Official variables yield somewhat greater forecasting accuracy for longer term forecasting (such as six-steps ahead
forecasts). Hence, our results indicate that GT data can be very useful for forecasting GDP growth in the short term.
Work on forecasting with ML has recently been growing. It would be impossible to provide a complete list
here. We shall unfortunately be able to discuss only a few, certainly nonrepresentative, contributions that are closely
related to the topic of the present paper. Biau and D’Elia (2012) used the Random Forest (RF) algorithm to predict
Euro-area GDP, and found that a combination of the RF and a linear model can outperform the benchmark autore-
gressive model. Other authors, such as Tiffin (2016) and Jung, Patnam, and Ter-Martirosyan (2018), have applied
the Elastic Net, the SuperLearner, and the Deep learners algorithms to produce forecasts that can outperform those
made by traditional statistical methods. A comprehensive survey of applications of ML methods for macroeconomic
forecasting is also provided by Fuleky (2019). Recently, Medeiros, Vasconcelos, Veiga, and Zilberman (2021) ex-
tensively applied standard ML methods to forecast the U.S. inflation rate, and find that the Random Forest (RF)
algorithm can deliver the best forecasts among all the ML algorithms that they considered. With respect to fore-
1
The Bridge regression has been used by policymakers to make short-term forecasts since at least the 1980s. Bridge models are linear dynamic
models that use short-term indicators (available at a higher frequency) to forecast low-frequency variables such as aggregate GDP or its
components.

2
 casts of Canadian GDP, Tkacz (2001) adopted a neural network approach, while Chernis and Sekkel (2017) used
a dynamic-factors model. As the success of a supervised ML method depends on its capability to capture nonlin-
earities in the relationship between the response and predictive variables as well as in the (nonlinear) interactions
between two predictive variables, in this study we attempt to forecast Canadian GDP growth rate with (nonlinear)
tree-based ensemble algorithms that have not been fully explored for economic forecast, namely, Gradient Boosting
Machine (GBM), eXtreme Gradient Boosting (XGBoost), and Microsoft’s Light Gradient Boosting Machine (Light-
GBM).2 Such ensemble methods have proved their empirical success in many other applications because averaging
several nonlinear models with possibly different sets of predictive variables is perhaps the best way to cope with
model uncertainty which is very common in practice [see, e.g., Athey, Bayati, Imbens, and Qu (2019)].
To implement the aforementioned ML algorithms, we used the automated machine learning (AutoML) – an
application interface which automates the process of simultaneously training a large number of models using the
GBM, the XGBoost, the Generalized Linear Model (GLM), the Distributed Random Forest (DRF), Deep Learning
algorithms, and Stacked Ensembles.3 Our proposed forecasting procedure is a three-step procedure detailed in
Section 3. In the first step, we pre-select the most relevant GT features and Official variables through the variable
importance measure of XGBoost, which has recently become a very popular and powerful algorithm in applied
ML and Kaggle competitions.4 It should be noted that variable pre-selection in our procedure plays a crucial role
in improving forecast performance. Therefore, we rigorously train and validate several thousand models in order
to achieve optimal forecast accuracy. In the second step, the selected variables are then introduced as inputs to
train and validate the RF regression and boosted tree models. In the last step, the trained models are used to make
out-of-sample forecasts.
With respect to forecast evaluation, a common tradition is to calculate the root mean squared error (RMSE), the
mean absolute error (MAE), or R2 to evaluate and rank forecasts. While these metrics are useful in differentiating
among the forecasting accuracy of various models, using only RMSE, MAE or R2 does not always provide an
accurate assessment of the forecasting ability of an individual model against the actual target variable. Thus, as
recently emphasized by Brown and Hardy (2021), who refer to this phenomenon as “the mean squared prediction
error paradox,” a model may yield a poor prediction of the target variable even if its RMSE is low. They propose
the correlation among variables as a superior assessment method, and show analytically, empirically and graphically
that a forecast with the lowest RMSE may have the worst correlation with the target variable, and thus be a poor
predictor. Only under certain regularity conditions, these two metrics coincide. In this paper, we provide a graphical
representation of the actual versus predicted trajectories to give a clear picture of how well our predicted GDP growth
rate (with both GT and Official variables) tracks actual GDP growth. We have also calculated Hansen, Lunde, and
Nason’s (2011) model confidence set (MCS) p-values of all models considered. This MCS procedure exploits all
information available in the data to decide if a model belongs to the set of the best models by implementing a
sequence of statistical tests to verify the null hypothesis of equal predictive ability.
The rest of this paper is outlined as follows: Section 2 provides a brief description of the ML algorithms em-
ployed in this paper. Although this description may appear non-technical, interested readers can find a rigorous
treatment of these algorithms in many good ML textbooks, for example, Hastie, Tibshirani, and Friedman (2009)
and Murphy (2012). Section 3 explains our three-step forecasting procedure. Section 4 presents GT and Official
data that we used to predict GDP growth rate. Section 5 provides the main empirical findings of this paper. Section
6 concludes this paper. Finally, further details about the software packages used to implement the ML methods and
other tables and figures are collected in two appendices at the end of the paper.
2
For example, the XGBoost by default utilizes a nonlinear booster, called gbtree. However, one could use the linear booster gblinear to
generate linear models.
3
AutoML is available in H2O (an open-source platform for machine learning).
4
https://www.kaggle.com/competitions.

3
 2 Description of ML algorithms
We start by explaining the concept of decision tree which is used as the building block for the RF regression and
other (tree-based) boosting algorithms.
Decision Trees: This model provides a tree-based method of splitting the data recursively using one feature at a
time. At the end of the process, we end up with leaf nodes – that is, nodes where there are no further splits and
where decisions are made. To take a simple example as an illustration, suppose the dependent variable Y is a binary
(true-false) variable representing the statement “I will go out to play football today.” The outcome of Y depends
on two independent variables (or features): the chance of raining (X) and the temperature (Z). Given a training
sample, say, X = {0.1, 0.8, 0.2, 0.1} , Z = {30o C, 20o C, 7o C, 15o C} , and Y = {T rue, F alse, F alse, T rue} ,
the algorithm can then automatically learn the following set of rules:

X
X ≤ 0.6 X > 0.6

Z False
Z ≤ 10o C Z > 10o C

False True

Here, X represents the root, Z represents a branch node, and True and False represent the leaf node. Starting
with X, if X is sufficiently high (say, above 0.6), then the leaf node Y takes the value ‘False’. Otherwise, we
continue to split Z. If Z is high enough (say, above 10o C), then the leaf node Y takes the value ‘True’; otherwise,
it takes the value ‘False’. With this set of decision rules in hand, one can easily predict any future value of Y
given values of X and Z. A critical issue is to find an appropriate threshold at which to split a node (for exam-
ple, the numbers ‘0.6’ and ‘10o C’ used to construct this decision tree). An ideal split would partition the sample
into two non-overlapping sets. In our case, we were able S to partition theTtraining sample into two distinct sets:
{True: (X ≤ 0.6) (Z > 10 C)} and {False: (X > 0.6) ((X ≤ 0.6) (Z ≤ 10o C))} . In reality, there may
o
T
be many features with multiple classes, and two classes may be overlapping. We cannot always get a pure split as
we create more branches further down a tree, but we can make the leaf nodes as pure as possible by setting the node
thresholds so as to minimize the Gini impurity or, equivalently, the cross entropy (the entropy is minimal if all mem-
bers in a node belong to the same class and the entropy is maximal if all these members are uniformly distributed
across classes).5
Although the preceding example illustrates the concept of a decision tree based on categorical variables (specif-
ically, a binary variable in this case), the same concept applies if our dependent variable (Y ) and features (X) are
real numbers. Imagine that the set of all possible values [that Y can take] has a finite dense set, say w1 , . . . , wM , and
each element wm , m = 1, . . . , M, (which can be viewed as a class label) is associated with a region Rm constructed
by recursively partitioning the sample spaceP of X with a given set of thresholds. The tree regression function can
then be approximated as: E[Y |X = x] ≈ M m=1 wm I (x ∈ Rm ) , where I(A) is an indicator function that takes
the value of one if the event A is true and zero otherwise. A tree grown using real-valued data on Y and X is called
a regression tree.
Thus, a decision (or regression) tree is a sequence of greedy searches for the best splitting points for each feature,
based on a training sample. However, this procedure tends to cause overfitting as the optimal splitting points may
only work well for the training sample, but will not generalize well enough to new data. To prevent overfitting, it is
necessary to prune the tree by setting limits for the maximal depth of the tree, the minimum number of observations
needed in a level to allow further splitting, and the minimum number of observations needed in a level to allow
it to become a leaf node. These parameters may be obtained via cross-validation (CV), which employs the model
5
The mathematical expressions of these criteria are given in Murphy (2012, Chapter 16).

4
estimated from a training sample to predict the dependent variable on a validating sample (that does not overlap
with the training sample) such that the accuracy level on the validating sample is maximal. However, there are other
more effective algorithms that can prevent overfitting, and at the same time, reduce the high variance that a tree
regression may suffer from. We shall next describe two ensemble methods, namely the Random Forest and boosting
algorithms.
Random Forests: The method of random forests was first introduced by Ho (1995) and later formalized by Breiman
(2001). A random forest can be considered as an ensemble of trees. The main idea is to average many different
trees that individually suffer from high variance to build a more robust model that can produce better out-of-sample
forecasts whilst being less susceptible to overfitting. The RF algorithm can be implemented in the following four
steps:

1. Draw a bootstrap sample of size less than or equal to the original sample size.

2. Grow a decision tree from each bootstrap sample. At each node,

a. Randomly select a subset of features without replacement.

b. Split the node using the feature that provides the best split using the threshold obtained by mini-
mizing the Gini impurity or the cross entropy.

3. Repeat the above two steps k times to create k independent trees.

4. Aggregate the predictions produced by the k trees.

For time series data [which have a natural ordering], the step 1 of the above RF algorithm is usually implemented
via a “block bootstrap” method according to which a sufficiently long block of time-series observations is resampled
in order to capture serial dependence in the block. The regular block bootstrap does this with a fixed block length
while the stationary bootstrap uses random block lengths, where the length may be drawn from an auxiliary distri-
bution (see, e.g., Lahiri (2003) for a detailed discussion of dependent bootstrap methods). Decreasing the size of the
bootstrap sample (like the case with bootstrap sample 3 below) improves the diversity of trees as the proba-
bility that a particular observation is included in all the bootstrap samples is lower. This diversity may yield a more
random forest, which can help to reduce the effect of overfitting. On the contrary, increasing the size of the bootstrap
sample can exacerbate the overfitting problem as trees are more likely to become similar to each other, and therefore
learn to fit the original training data more closely. However, an issue with block bootstrap is that points around each
joint of two blocks may not well mimic the serial dependence structure in the original sample. Therefore, if the time
series is long, one could implement sub-sampling (i.e., sampling just one block at a time) instead. Another potential
sampling scheme that could be used in the first step of the RF algorithm is Dahl and Sørensen’s (2021) resampling
using generative adversarial networks. This method can effectively generate many time series that mimic the serial
dependence structure of the original series.

original sample block 1 block 2 block 3

bootstrap sample 1 block 3 block 1 block 2

bootstrap sample 2 block 2 block 3 block 1

bootstrap sample 3 block 3 block 1

Boosting: The idea of ‘boosting’ was first introduced by Freund and Schapire (1996) and further developed by
Breiman (1998) and Friedman (2001). Unlike a random forest, which is made of trees grown independently using
different bootstrap samples of the original dataset, a boosted tree consists of individual trees trained in a sequence
such that the next tree corrects the error made by the previous one, where the error is defined as the negativity of the

5
 gradient of a loss function. The gradient boosted trees algorithm builds one tree at a time and combines the results
recursively as in the following diagram:

Training sample Error 1 Error 2

(Y, X) (1 = Y − f1 (X))(2 = 1 − f2 (X))

Tree 1 (f1 (X)) Tree 2 (f2 (X)) Tree 3 (f3 (X))

Boosted tree (f1 (X) + f2 (X) + f3 (X))

As mentioned above, in prediction problems, growing a very deep decision tree at each boosting step may lead
to overfitting because doing so may reduce the expected loss on the training data beyond some point, but it can
cause the expected loss on the testing data to stop decreasing and often to start increasing. The XGBoost developed
by Chen and Guestrin (2016) overcomes the overfitting problem by using regularization at every boosting step to
regularize the degree to which expected loss on the training data can be minimized.6 As a consequence, regression
trees built by the XGBoost are usually simple, but they can achieve a high level of accuracy when being applied to
predict new data. One novel aspect of this algorithm is the fast node splitting method proposed for handling sparse
data. Therefore, the XGBoost runs more than ten times faster than the standard boosting algorithms and it can scale
to big datasets in shared-memory or distributed-computing environments.
From a theoretical perspective, tree-based boosting algorithms (especially the XGBoost) can outperform the
RF algorithm because the former algorithms can be much more immune to the curse of dimensionality [see, e.g.,
Nielsen (2016)].7 The XGBoost learns the similarity between data points by adapting the neighbourhoods according
to the amount of data in each region through various types of regularization while the RF employs only one type of
regularization (i.e., row and column sampling as described above).

3 Forecasting Method
We implement the rolling-window strategy where the sample in each window is split into three sub-samples used to
train, validate, and test a model respectively.
6
To give an example of regularization, suppose that one tries to fit a quadratic polynomial, say y = a + bx + cx2 , into points that appear to
follow the linear model: y = a + bx. The regression algorithm may produce curves that score well but look more complicated than what the
original data would be like. We can influence this algorithm so that it can produce straight lines instead of curves by adding a penalty to the
loss function, say loss(x, y) + λ |c| , where λ is a positive constant that controls how much we want to penalize c. If λ is set to zero, then
there is no regularization. As λ is set to larger values, the larger c will then be heavily penalized. In general, regularization sacrifices some of
the flexibility of a model in order to achieve better generalization to unseen data.
7
The curse of dimensionality is defined as a phenomenon where the RMSE of a nonparametric estimator increases as the number of explanatory
variables (or features) becomes larger, ceteris paribus. It is very common across most nonparametric estimation problems, such as splines,
kernels, or trees.

6
 train validate test
rolling window 1:
train validate test
rolling window 2:
train validate test
rolling window 3:
train validate test
rolling window 4:

We take the following steps to produce forecasts in each rolling window:

Step 1: In each training and validating sample, pre-select the most relevant predictors from the Official and GT
variables by using the XGBoost or a variable screening procedure [such as Yousuf and Feng’s (2021) PDC-
SIS+].
Step 2: Use the above pre-selected predictors as inputs to train and validate random forest regression and boosted
tree models.
Step 3: Use the models obtained in Step 2 to produce out-of-sample forecasts for three horizons: h = 1, 3, and 6.
We use data from January 2004 to 2019 (192 months) as the earliest date for which GT data is available is January
2004. The training periods range from January 2004 to December 2015 (143 months), the validating periods range
from January 2016 to December 2017 (24 months), and the test periods range from January 2018 to December 2019
(24 months). We take the first differences of the logarithmic transformations of the data, then multiply them by 100
for both Official and GT data. We have also implemented the Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test for
stationarity on the data, and found that all the transformed variables are indeed stationary.
We evaluate our forecasts by using the following three metrics: the root mean squared error (RMSE), the mean
absolute error (MAE) – which is more robust to outliers than the RMSE, and the out-of-sample (OoS) R2 :
v
u T −h 
u 1 X 2
RM SE = t Yt+h − Ybt+h , (1)
T − h − t1
t=t1 +1
T −h
1 X  
M AE = Yt+h − Ybt+h  , (2)
 
T − h − t1
t=t1 +1
PT −h  2
t=t1 +1 Yt+h − Ybt+h
R2 = 1 − P  , (3)
(AR) 2

T −h
t=t1 +1 Yt+h − Ybt+h

where t1 is the first observation of the out-of-sample period, T is the total number of observations, Ybt+h , t =
(AR)
t1 + 1, . . . , T − h, are the h-period ahead forecasts of Yt+h , and Ybt+h , t = t1 + 1, . . . , T − h, are the baseline
h-period ahead forecasts of Yt+h calculated from an autoregressive (AR) model. The OoS R2 takes value in between
−∞ and one. Therefore, a negative value of the OoS R2 means that the baseline forecast is better (in terms of the
RMSE) while a positive value merely means that our forecast is better than the baseline forecast, and as the value of
R2 is closer to one, the better our forecast becomes.
When there are many methods or models available to forecast the same economic variable, an important question
is to assess which models provide superior predictive ability. As pointed out by Hansen et al. (2011), since different
competing models are usually built to answer a specific econometric question, a single model may not dominate
all other models either because they are statistically equivalent or because there is not enough information coming
from the data to discriminate the models. Therefore, we apply Hansen et al.’s (2011) model confidence set (MCS)
procedure to determine if a method is one of the best forecasting methods. A method with a small MCS p-value is
unlikely to belong to the set of best methods.

7
 4 Data
Official data: We use both Canadian and U.S. variables to forecast the Canadian GDP growth rate. It has long
been noted that some U.S. variables have predictive power for Canadian GDP because of the extensive Canada-U.S.
trading relationship. Some variables that we use here have also been used by Chernis and Sekkel (2017) and the
Bank of Nova Scotia’s nowcasting model for the Canadian economy. There are nine Canadian economic variables
and five U.S. variables. All the variables (together with their actual release and reference periods) used to predict
the monthly real GDP growth rates are provided in Table 3.

GT data: This data set (available from www.google.com/trends) provides all search volumes and shows the
percentage of web searches performed on Google. GT data set reports data in terms of an index of queries rather
than in terms of the raw number of queries. This index is then normalized so that the maximum search volume at a
given point in time and location is set to 100%, and every other observation is calculated relative to that maximum
value. The resulting numbers on the scale from zero to 100 are thus based on the total number of searches for a
specific category relative to all other search queries. Choi and Varian (2009) have used the term “Google Trends”
while Götz and Knetsch (2019) and Ferrara and Simoni (2019) have used the term “Google search data”. GT data is
an unbiased sample of the Google search data. According to the Canadian internet usage survey in 2018, 94% of the
population had internet access at home, and the share of all the Canadians aged 15 or above who use the internet was
91%. Interestingly, approximately 84% of the internet users bought about $57.4 billions worth of goods or services
online in 2018 compared to $18.9 billions in 2012. We obtained all the data used in this study from Bloomberg,
Statistics Canada, Google, and the Fed. GT data are only available for normal users from January 2004 to date.
We proceeded to divide our GT dataset into two main categories. The search keywords in the first category, which
we call general search terms, are extracted from the R package ‘gtrendsR’. We have downloaded 26 categories and
269 subcategories of GT data for a total of 1776 search terms. Only 464 search terms remain after removing columns
with missing values and outliers. The second category, which we call the relevant search terms, has the same variable
names as those in the monthly and quarterly Canadian GDP components. There are 80 search terms in this category.
If a column in the second category has no more than two missing values, we replaced these missing values with the
column mean. We also used the program X-13ARIMA-SEATS developed by the U.S. Census Bureau and the Bank
of Spain to seasonally adjust GT data since these data may not be seasonally adjusted.

5 Empirical Results
5.1 Selection of Predictors
We implemented Step 1 of the rolling window strategy described in Section 3.8
Using the Official data, we have selected the best eight variables out of 14 official macroeconomic variables by us-
ing the variable importance measure of XGBoost.9 This variable selection method is different from other traditional
methods, such as PCA, PLS, SIS, and LASSO, in the sense that the former selects variables with an aim to maximiz-
ing their out-of-sample predictive ability for a target variable. We can notice how the model performance changes as
we add or drop any variable based on the variable importance measure. Specifically, we run 1000 XGBoost models
(constructed from decision trees) at a time, and compute the variable importance measures. We obtained the best
predictive performance [using the validating sample] when using 1000 models (instead of 800 or 1600 models).
Training a larger number of models for the current AutoML project may be accomplished by using different random
seeds. These 1000 models have different parameter values, such as “depth of tree ”, “sample rate ”, or type of “grow
8
We made use of the XGBoost function of the AutoML package.
9
One of the advantages of using gradient boosting is that it provides importance scores for each feature after the boosted tree is constructed.
Variables with high importance are, by definition, the major drivers of the target variable. We use the H2O to determine variable importance.
In H2O, the variable importance score measures the relative influence of each feature according to the extent to which this feature can reduce
the overall RMSE.

8
 policy or booster”. We chose the model with minimum RMSE (called leader model), and use the scaled importance
measure to choose variables. A plot of the selected variables versus their importance scores is presented on the left
panel of Figure 3. Here, we sequentially dropped the variables with less than 4-5 % of importance score, repeating
the same process until we obtained sufficient improvement in the RMSE.

Using GT data, we select GT features by using either the XGBoost’s variable importance scores or Yousuf and
Feng’s (2021) variable screening procedure (explained below). As illustrated in Figure 4, we first apply the XGBoost
variable importance measure on the general search terms [listed in Table 5] and run 1000 tree-based models.10 The
whole process was repeated until we obtained a sufficient improvement in the RMSE. We dropped variables with
importance scores less than 5-10 %. Here, we used slightly stricter variable selection criteria relative to the one used
for Official data because GT data have a much larger number of features. We have ended up with 26 features from
this category of search terms. Next, the same procedure was applied on the relevant search terms [listed in Table 4].
We eventually obtained 23 features for this category. Finally, we combined the top features from both the categories
of GT data: there is a total of 22 features to be used for forecasting the GDP growth.
Yousuf and Feng’s (2021) variable screening procedure based on Székely and Rizzo’s (2014) partial distance
correlation, namely PDC-SIS+, can be used to select time-series variables. This procedure defines the set of possible
predictors as: Sk,l = (Yt−1 , . . . , Yt−L , Xt−1,k , . . . , Xt−l+1,k ) for 1 ≤ k ≤ K and 1 ≤ l ≤ L, where K is the
number of features and L is a maximum number of lags to be set. Using the partial distance correlation to screen
out irrelevant variables has an important advantage over the conventional partial correlation as a null partial distance
correlation between two variables implies conditional independence between these two variables while this is not
necessarily the case with other measures of conditional correlation. Therefore, if an element in the set of predictors
has zero partial distance correlation with the target variable, then this element has no predictive ability power. All
the ML algorithms used in this paper actually train nonlinear models in which features and the target variable may
have complex relationships. PDC-SIS+ can effectively select predictors that can be either linearly or nonlinearly
associated with the target variable. To apply PDC-SIS+, we first combine general search terms and relevant search
terms to make a set of total 544 predictors, then take square roots of these predictors. There are 24 variables being
selected by this procedure, as listed in Table 7.

5.2 Comparison of Forecasts

Given the set of predictors selected in Section 5.1 above, we proceeded to Step 2 of the rolling window strategy
described in Section 3. The following ML algorithms were implemented: Random Forest, Gradient Boosting Ma-
chine (GBM), Light Gradient Boosting Machine (LightGBM) described in Appendix A, and the XGBoost, together
with a simple benchmark autoregressive (AR) model of order one. After training and validating forecast models by
employing these ML algorithms, we then used the trained models to compute out-of-sample forecasts (as described
in Step 3 in Section 3).
All the forecasts [at three horizons, h = 1, 3, and 6 steps ahead] of the monthly GDP growth rate using both
Official and GT data for a shorter period of time (the entire sample period is from February 2004 to December 2019
and the out-of-sample period is from January 2018 to December 2019) are presented in Table 1. The first column
lists ML algorithms being implemented, the second column presents types of data being used as the predictors of the
GDP growth rate together with the procedures used to pre-select variables [explained in Section 5.1]. The rest of the
columns provide the RMSE, the MAE, the OoS R2 , and the MCS p-value of each forecast method. As mentioned
earlier, the Official data features (for both Canada and U.S.) presented here were pre-selected with the XGBoost.
The key takeaways from this table are:

• All of the ML models have a superior forecasting accuracy than the AR model. (The only exception is the RF
which performs less well than the AR in the case of the one-step-ahead forecast.) As expected, one-step-ahead
10
Note that we can only provide a few GT data feature names (out of 1776 search terms in this category) in this table. The full list of features
is available upon request.

9
 forecasts are more accurate than two- or three- steps-ahead forecasts.

• The RF does not seem to perform as well as the XGBoost or the GBM across all the three forecast horizons.
When it is applied to variables pre-selected (with PDC-SIS+) from both Official and GT variables, the RF’s
performance can be on par with that of other boosting algorithms using only GT or Official data.

• The LightGBM using Official data can perform better than other models in terms of the RMSE and the MCS
p-value at the one-step-ahead forecast.

• The XGBoost and the GBM provide more accurate long-horizon forecasts. Specifically, the GBM with GT
features selected using the XGBoost achieves the minimum RMSE among other models and various data
categories at the three-steps-ahead forecast. Meanwhile, the GBM with GT features selected using PDC-SIS+
provides the minimum MAE for three-steps-ahead forecasts.

• The GT features pre-selected by PDC-SIS+ can be better predictors than those pre-selected by the XGBoost.
Both the GT features and the Official variables perform equally well in a shorter horizon for the models trained
by the XGBoost and the LightGBM. In a longer horizon, forecasts with the Official variables can outperform
those with GT data. However, we do not observe an improvement in the forecasts using both GT and Official
data. This is consistent with Ferrara and Simoni (2019) and Götz and Knetsch (2019) who conclude that
forecast accuracy does not improve by using both GT and Official data.

A more detailed analysis of Table 1 also permits a comparison of the predictive success of models depending on the
different forecast horizons and different data sets. Given the overall lower performance of the RF model in our case,
we will only compare the GBM, XGBoost and LightGBM.

One-step ahead forecasts: In general, it seems that the GBM model is the best overall model across (almost) all data
and variable selection categories and in terms of all the four metrics. The only exception is the case when Official
data is used alone. In this case, the LightGBM model performs the best of all models, and has the lowest RMSE
(0.00137), the lowest MAE (0.00086), and the best MCS p-value (1.000).
As for the XGBoost model, we can see that the latter is a relatively close second to the GBM in terms of its
forecasting accuracy. The greatest difference in accuracy between the two models is observed when only GT data
are used, but their accuracy in terms of RMSE is almost identical when both Official and GT data are used together.
However, even in this case, the GBM is superior to XGBoost according to its MAE and its MCS p-value.
It is also interesting to note that for the one-step-ahead forecast, the GBM model yields smaller RMSE’s in the
case when only GT features are selected (0.00322 and 0.00077 respectively for the two selection criteria) than in
the case when only Official data features are selected (0.0123), or in the case when both GT and Official data are
selected (0.02627). This indicates that the GBM is best suited to produce very good short-term forecasts when only
GT data are available.

Three-step ahead forecasts: While forecasting accuracy clearly diminishes across all models as the forecasting
horizon lengthens, here too, we reach the same general conclusion as for the one-step-ahead forecasting horizon:
GBM is still the best overall model with the XGBoost being a close second for all data and variable selection
categories. The XGBoost’s performance is very close to that of the GBM especially when GT variables are selected
by PDC-SIS+ (which also emerges, overall, as the more successful variable selection method). Comparing the
GBM model’s performance when using both the Official variables and GT features selected with PDC-SIS+, with
its performance when using only GT variables (selected with PDC-SIS+), we observe that Official data contribute to
greater forecasting accuracy in terms of a higher MCS p-value (1.000 compared to 0.759), while the OoS R2 is high
(0.96) in both cases. The RMSE’s and MAE’s of the two cases are also comparable, with the RMSE being slightly
higher (0.04831 compared to 0.04526) but the MAE being slightly lower (0.01272 compared to 0.01433).
Alternatively, when we compare the GBM’s performance where only Official data is used with the case when
both sets of data are used, we observe an improvement in forecasting accuracy according to all four metrics. (RMSE:

10
0.04831 versus 0.06164; MAE: 0.01272 versus 0.01875; OoS R2 0.96 versus 0.94; MCS p-value 1.000 versus
0.759). Clearly, GT data contributes to forecasting accuracy here (in contrast to Ferrara and Simoni’s (2019) finding).
Hence, we can conclude that the most successful model for the three-steps ahead forecast horizon is the GBM
model using both pre-Official variables and GT variables selected with PDC-SIS+. The LightGBM model does not
perform as well as the other two models in the three-steps ahead forecasting horizon even when Official data are
used alone.

Six-step ahead forecasts: At this longer forecasting horizon, forecasting accuracy diminishes yet again across mod-
els. However, interestingly, we observe that now the XGBoost model tends to dominate the GBM model in terms of
RMSE across all data and variable selection categories, but with the GBM model having somewhat superior MCS
p-values. We can generally say that the XGBoost model is the most successful model except when Official data are
used alone. In this latter case, once more, (as in the one-step-ahead-forecast) the LightGBM model performs the
best of all models (according to its RMSE, its MAE, its OoS R2 and its MCS p-value.)

Table 2 presents forecast results using only Official data for a longer period of time (the entire sample period
is from January 1981 to December 2019 and the out-of-sample period is from January 2018 to December 2019).
We implemented the same set of ML algorithms mentioned in the first column of Table 1. In general, forecast
performance in terms of the RMSE and the MAE significantly improves with a longer sample of data. The values of
the OoS R2 reported in these tables are consistent with the values reported in other papers on GDP nowcasting. For
example, de Valk, de Mattos, and Ferreira (2019) nowcasted Brazilian GDP growth with a dynamic factor model
and found an OoS R2 value of roughly 0.90. The tree-based boosting algorithms continue to outperform the RF
algorithm in this case as well. Forecasts using the Canadian variables as predictors may be slightly better in a
shorter horizon, and can deteriorate over longer horizons. This implies that the effect of the U.S. economy on the
Canadian economy becomes visible over an extended period of time. Moreover, forecasts using only the Canadian
employment data with the lags and a rolling average of GDP growth rates perform better than those using other
Canadian variables and/or U.S. variables.
Figures 1 and 2 plot the out-of-sample forecast performance (for the period from January 2018 to December
2019) with GT plus Official data and Official data only, respectively. These plots show that one-step-ahead forecasts
are quite close to the observed GDP growth rates, thus they can track very well the GDP growth path. All the models
can produce good forecasts with employment as the only predictor. Indeed, Foroni, Marcellino, and Stevanovic
(2020) have mentioned that employment and industrial production are good predictors of GDP. However, we could
use only employment as it is available with a ten-day lag while industrial production is only available with a 50-day
lag. Therefore, we can forecast the Canadian GDP growth well prior to the release of official GDP figures.

6 Conclusion
This paper compares the performance of the tree-based ML algorithms (i.e., the LightGBM, the XGBoost, the GBM,
and the RF) in a GDP growth forecasting exercise using Official and GT data as predictors. In particular, we used
the automated machine learning (AutoML) with the XGBoost algorithm to pre-select the most important variables
from a broad set of potential candidate predictors. We run 1000 XGBoost models and implement AutoML 5 to 15
times until the RMSE can be sufficiently improved. We have dropped all variables with less than 5-10 % of variable
importance.
We also employed another novel variable screening procedure, namely PDC-SIS+, to pre-select the best GT
features. We have found that the features pre-selected by this variable screening procedure can produce a better
forecast accuracy than the features pre-selected with the XGBoost. All the algorithms that we use can forecast the
Canadian GDP growth very well, except for the RF. This result indicates that the GBM, the LightGBM, and the
XGBoost can be very useful ML algorithms for macroeconomic prediction. We have also found that GT data can be
a good substitute for Official data as the predictor of the monthly Canadian real GDP growth rate in shorter forecast

11
horizons. However, the forecast ability of GT data can decrease over a longer horizon.
An important limitation of tree-based algorithms is that, although they are quite good at capturing the seasonal
and nonlinear patterns in the data, these algorithms may not track time trends or structural breaks very well. We will
study this issue further in a future research. We conjecture that the accuracy of forecasts in different scenarios can
be enhanced by combining various boosting methods (for instance, the XGBoost with the LightGBM).

References
Athey, S., M. Bayati, G. Imbens, and Z. Qu (2019). Ensemble methods for causal effects in panel data settings. AEA
Papers and Proceedings 109, 65–70.

Biau, O. and A. D’Elia (2012). Euro area GDP forecasting using large survey datasets: A random forest approach.
mimeo, available from https://ec.europa.eu/eurostat/web/products-eurostat-news/-/
EWP-2011-002.

Breiman, L. (1998). Arcing classifier. Annals of Statistics 26(3), 801–849.

Breiman, L. (2001). Random forests. Machine Learning 45(1), 5–32.

Brown, P. P. and N. Hardy (2021). The mean squared prediction error paradox. mimeo.

Chen, T. and C. Guestrin (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd
international conference on knowledge discovery and data mining, pp. 785–794. ACM.

Chernis, T. and R. Sekkel (2017). A dynamic factor model for nowcasting Canadian GDP growth. Empirical
Economics 53(1), 217–234.

Choi, H. and H. Varian (2009). Predicting initial claims for unemployment benefits. mimeo, available from
https://static.googleusercontent.com/media/research.google.com/en//archive/
papers/initialclaimsUS.pdf.

Choi, H. and H. Varian (2012). Predicting the present with Google Trends. Economic Record 88, 2–9.

Dahl, C. M. and E. N. Sørensen (2021, February). Time series (re)sampling using generative adversarial networks.
mimeo.

de Valk, S., D. de Mattos, and P. Ferreira (2019). Nowcasting: An R package for predicting economic variables
using dynamic factor models. The R Journal 11(1), 1–15.

Fan, J. and J. Lv (2008). Sure independence screening for ultrahigh dimensional feature space. Journal of the Royal
Statistical Society: Series B (Statistical Methodology) 70(5), 849–911.

Ferrara, L. and A. Simoni (2019). When are Google data useful to nowcast GDP? an approach via pre-selection and
shrinkage. mimeo, available from https://arxiv.org/abs/2007.00273.

Foroni, C., M. Marcellino, and D. Stevanovic (2020). Forecasting the Covid-19 recession and recovery: Lessons
from the financial crisis. International Journal of Forecasting (in press).

Freund, Y. and R. E. Schapire (1996). Experiments with a new boosting algorithm. In L. Saitta (Ed.), Proceedings
of the Thirteenth International Conference on Machine Learning (ICML 1996), San Francisco, California, pp.
148–156. Morgan Kaufmann.

Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. Annal of Statistics 29(5),
1189–1232.

12
Fuleky, P. (2019). Macroeconomic Forecasting in the Era of Big Data: Theory and Practice, Volume 52. Springer.

Götz, T. B. and T. A. Knetsch (2019). Google data in Bridge equation models for German GDP. International
Journal of Forecasting 35(1), 45–66.

Hansen, P. R., A. Lunde, and J. M. Nason (2011). The model confidence set. Econometrica 79(2), 453–497.

Hastie, T., R. Tibshirani, and J. Friedman (2009). The Elements of Statistical Learning: Data Mining, Inference,
and Prediction (second ed.). Springer.

Ho, T. K. (1995). Random decision forests. Proceedings of the 3rd International Conference on Document Analysis
and Recognition 1(1), 278–282. Montreal, QC, Canada.

Jung, J.-K., M. Patnam, and A. Ter-Martirosyan (2018). An algorithmic crystal ball: Forecasts-based on machine
learning. IMF working paper.

Lahiri, S. N. (2003). Resampling Methods for Dependent Data. Springer.

Medeiros, M. C., G. F. R. Vasconcelos, A. Veiga, and E. Zilberman (2021). Forecasting inflation in a data-rich
environment: The benefits of machine learning methods. Journal of Business & Economic Statistics 39(1), 1–22.

Murphy, K. P. (2012). Machine Learning: A Probabilistic Perspective. The MIT Press.

Nielsen, D. (2016, ). Tree boosting with XGBoost: Why does XGBoost win “every” machine learning competition?
Master’s thesis, Norwegian University of Science and Technology.

Székely, G. J. and M. L. Rizzo (2014). Partial distance correlation with methods for dissimilarities. Annals of
Statistics 42(6), 2382–2412.

Tiffin, A. (2016). Seeing in the dark: A machine-learning approach to nowcasting in Lebanon. IMF working paper.

Tkacz, G. (2001). Neural network forecasting of Canadian GDP growth. International Journal of Forecasting 17(1),
57–69.

Tkacz, G. (2013). Predicting recessions in real-time: Mining Google trends and electronic payments data for clues.
CD Howe Institute Commentary No. 387.

Yousuf, K. and Y. Feng (2021). Targeting predictors via partial distance correlation with applications to financial
forecasting. Journal of Business & Economic Statistics (forthcoming).

13
Figure 1: Predicted vs. actual real GDP growth (January 2018 - December 2019) with both GT and Official data
(Full sample period: January 2004 - December 2019)

14
Figure 2: Predicted vs. actual real GDP growth (January 2018 - December 2019) with only Official data (Full
sample period: January 1981 - December 2019)

15
 Table 1: Summary of the out-of-sample forecasts using both the GT and Official data features for the period from January 2018 to December 2019 (Full
sample period: January 2004 - December 2019)

Data with Variable h=1 h=3 h=6

Algorithm
Selection Method
RM SE M AE OoS R2 M CS p-value RM SE M AE OoS R2 M CS p-value RM SE M AE OoS R2 M CS p-value

Autoregression (AR) GDP 0.09831 0.09831 - 0.028 0.11301 0.09397 - 0.008 0.17979 0.14967 - 0.015

Random Forest GT with XGBoost 0.13843 0.10571 0.72 0.023 0.14085 0.10702 0.71 0.031 0.18387 0.14823 0.49 0.005

GBM GT with XGBoost 0.00322 0.00065 0.99 0.580 0.01957 0.07968 0.91 0.759 0.11207 0.04441 0.81 0.390

XGBoost GT with XGBoost 0.03905 0.01772 0.97 0.153 0.07161 0.03155 0.92 0.428 0.11054 0.06311 0.82 0.245

LightGBM GT with XGBoost 0.09029 0.06887 0.87 0.000 0.15989 0.12814 0.62 0.020 0.16378 0.09864 0.60 0.258

Random Forest GT with PDC-SIS+ 0.17098 0.13193 0.56 0.024 0.11830 0.09431 0.79 0.041 0.18709 0.15048 0.48 0.027

GBM GT with PDC-SIS+ 0.00077 0.00380 0.99 0.378 0.04526 0.01433 0.96 0.759 0.14085 0.06312 0.71 0.390

XGBoost GT with PDC-SIS+ 0.02501 0.00907 0.99 0.138 0.04921 0.01862 0.96 0.759 0.09347 0.04911 0.87 0.384

16
LightGBM GT with PDC-SIS+ 0.00313 0.00129 0.99 0.378 0.04561 0.02460 0.96 0.759 0.12045 0.09707 0.78 0.190

Random Forest Official∗ 0.12515 0.09576 0.76 0.039 0.14551 0.11450 0.68 0.049 0.15430 0.12334 0.64 0.049

GBM Official 0.01230 0.00251 0.99 0.378 0.06164 0.01875 0.94 0.759 0.10321 0.04651 0.84 0.390

XGBoost Official 0.01411 0.00721 0.99 0.028 0.06443 0.01992 0.93 0.759 0.09057 0.04839 0.87 0.390

LightGBM Official 0.00137 0.00086 0.99 1.000 0.07868 0.02538 0.90 0.759 0.07501 0.03820 0.91 1.000

Official plus
Random Forest 0.16612 0.12916 0.59 0.029 0.12575 0.09940 0.76 0.759 0.17953 0.14680 0.52 0.039
GT with PDC-SIS+

Official plus
GBM 0.02627 0.00536 0.98 0.378 0.04831 0.01272 0.96 1 0.09170 0.03752 0.87 0.390
GT with PDC-SIS+

Official plus
XGBoost 0.02632 0.01291 0.98 0.186 0.07479 0.03939 0.91 0.759 0.07916 0.04024 0.90 0.282
GT with PDC-SIS+

Official plus
LightGBM 0.13310 0.10642 0.73 0.029 0.14541 0.11757 0.68 0.759 0.16264 0.12763 0.60 0.017
GT with PDC-SIS+

∗ The variable selection procedure used for Official data is always the XGBoost.
 Table 2: Summary of the out-of-sample forecasts using only the Official data features for the period from January 2018 to December 2019 (Full sample
period: January 1981 - December 2019)

h=1 h=3 h=6

Algorithm Data
RM SE M AE OoS R2 M CS p-value RM SE M AE OoS R2 M CS p-value RM SE M AE OoS R2 M CS p-value

Autoregression (AR) GDP 0.01617 0.01617 - 0.017 0.12956 0.08113 - 0.014 0.19603 0.15953 - 0.010

Random Forest Canada & U.S. 0.08605 0.06485 0.89 0.029 0.07972 0.05983 0.90 0.013 0.08321 0.06706 0.89 0.004

GBM Canada & U.S. 0.02635 0.00537 0.98 0.575 0.05858 0.02045 0.94 0.417 0.07318 0.03198 0.92 0.231

XGBoost Canada & U.S. 0.01766 0.00804 0.99 0.090 0.03133 0.02237 0.98 0.025 0.05453 0.03717 0.95 0.218

17
LightGBM Canada & U.S. 0.01865 0.00458 0.99 0.264 0.03932 0.01742 0.97 0.409 0.05959 0.04407 0.94 0.152

Random Forest Canada 0.08519 0.06269 0.89 0.029 0.08054 0.05914 0.90 0.014 0.08211 0.06340 0.90 0.005

GBM Canada 0.01432 0.00292 0.99 0.469 0.04953 0.01525 0.96 0.417 0.04356 0.01813 0.97 0.231

XGBoost Canada 0.01298 0.00680 0.99 0.036 0.04489 0.01815 0.97 0.417 0.06595 0.04340 0.93 0.166

LightGBM Canada 0.01214 0.00348 0.99 0.264 0.03312 0.01846 0.98 0.212 0.07438 0.05833 0.91 0.051

Random Forest Employment only 0.01908 0.01541 0.99 0.001 0.01925 0.01537 0.99 0.005 0.02285 0.01874 0.99 0.013

GBM Employment only 0.01228 0.00250 0.99 0.469 0.00847 0.00235 0.99 0.816 0.01907 0.00879 0.99 0.225

XGBoost Employment only 0.00283 0.00210 0.99 0.575 0.00760 0.00529 0.99 1.000 0.01176 0.00647 0.99 1.000

LightGBM Employment only 0.00248 0.00135 0.99 1.000 0.00829 0.00369 0.99 0.816 0.01183 0.00985 0.99 0.981
A Implementation Softwares
A.1 LightGBM
The LightGBM algorithm introduced by Microsoft is faster than the XGBoost. The LightGBM is different from
the XGBoost in that the former algorithm grows trees vertically (and leaf-wise) whereas the latter algorithm grows
trees level-wise. The LightGBM selects and splits the leaf that contributes the most to loss reduction; it uses a
technique, called ‘Gradient-based One-Side Sampling’ (GOSS), which filters out data samples to select an optimal
split point. The XGBoost utilizes a pre-sorted algorithm and a histogram-based algorithm to find the best split point.
The XGBoost typically uses the nonlinear booster gbtree or the linear booster gblinear while the LightGBM uses the
same booster type as the Gradient Boosting Decision, Random Forest, Dropouts meet Multiple Additive Regression
Trees (DART), or GOSS.

A.2 Hyper-tuning Parameters

We have used both R and Python packages of H2O ver. 3.26.0.10 for the XGBoost. We also used the Scikit-learn,
XGBoost, and LightGBM libraries of Python to calculate rolling-window forecasts, and the PDC-SIS+ R rou-
tine to select predictors. All the codes to replicate the results reported in the paper can be downloaded from
https://github.com/Shafi2016/Canadian-GDP-Forecast.
Hyper-tuning parameters is one of the most complicated and vital tasks in ML in order to improve forecast
performance. It is relatively easy to build a XGBoost model, but to fine-tune this model is a daunting task as the
XGBoost has a large number of parameters. We used the package AutoML of H2O to initialize the parameters.
We then utilized the routine RandomizedSearchCV with rolling time-series based cross-validation to hyper-tune
the parameters. RandomizedSearchCV searches randomly through a large set of all possible combinations of the
parameters, then calculates the optimal parameters. We initialized this routine with many different random seeds in
order to obtain the parameter values with minimum RMSE while avoiding the effects of overfitting.
The XGBoost has many tuning parameters that may help us to cope with the overfitting issue. The first param-
eter controls the complexity of a decision tree. It splits the node only when the resulting split improves the loss
function. Other regularization parameters are penalty levels to balance loss and the amount of penalization defined
by an Elastic Net function. The parameter max depth specifies the maximum depth of a tree. By increasing this
parameter, we essentially increase the model complexity, and thus this can lead to overfitting. The parameter learn-
ing rate shrinks the step size of the stochastic gradient descent algorithm used to minimize the loss function. The
parameter min child weight defines the minimum summation of the weights of all observations required in a child.
The parameter sample rate specifies the fraction of observations randomly sampled for each tree. The parameter
booster sets the learner type, either linear or nonlinear.

B Other Tables and Figures

Table 6: Final GT data features selected by AutoML XGBoost

Property.Management Army Ford

Small.Business Financial.Markets Volvo
Car.Electronics Property.Development Transportation
Finance rental consumption
Construction Farm Mining
Energy.sector Services US.employment
US.retail Export

18
 Table 3: Official data variables for Canada and the U.S. used to forecast the Canadian GDP growth rate

Variable Reference period Release date Publication lag (days) Frequency

US : All Employees, Total Nonfarm (PAYEMS) Dec 2019 Jan 8, 2020 8 M
CAN : Employment rate - Jan 10, 2020 10 M
US: Industrial Production: Total Index (INDPRO) - Jan 15, 2020 15 M
Motor Vehicle Assemblies: Autos and Light Truck Assemblies (MVAAUTLTTS) - Jan 15, 2020 15 M
Housing Starts: Total: New Privately Owned Housing Units Started (HOUST) - Jan 21, 2020 21 M
CAN: Retail trade - Jan 22, 2021 22 M

19
US : Light Weight Vehicle Sales: Autos and Light Trucks (ALTSALES) - Jan 29, 2020 29 M
CAN : Housing starts - Jan 30, 2020 30 M
CAN: Export - Feb 5, 2020 37 M
CAN : Import - Feb 5, 2020 37 M
CAN : Manufacturing new orders (total) - Feb 18, 2020 50 M
CAN: Industrial production - Feb 18, 2020 50 M
Manufacturing inventories (total) - Feb 18, 2020 50 M
CAN : wholesale trade - Feb 24, 2020 56 M
GDP by Industry - Feb 28, 2019 60 M
 Table 4: Google Trends (GT) customized search keywords (relevant search terms)

Energy sector Content Construction Rental

Business sector Media Manufacturing Leasing
Business industries Cannabis Wholesale trade Professional
Industrial production Agriculture Retail trade Scientific
Manufacturing industries Forestry Transportation Technical services
Durable good fishing and hunting Warehousing Management companies
Information Mining cultural industries Management enterprises
communication technology Quarrying Cultural Administrative support
Technology sector Oil extraction Finance Waste management
Public Sector Utilities Real estate Health care
Social assistance Arts Entertainment Recreation
Accommodation Food services Public administration Perishable goods
Export Import Trade balance US industrial
US Shipment Building Permits Consumer confidence index Inflation
Employment US employment GDP growth Price Index
Housing starts US Retail Shipment US Shipment
US pmi global pmi Inventories Retail trade
US retail US housing Consumption Expenditure
Services Investment Residential Machinery
Equipment Intellectual.property Farm

Table 5: GT general search terms

Porsche Aquaculture Office.Supplies Import.Export
Rolls.Royce Food.Production Office.Furniture Maritime.Transport
Saab Forestry Printers Packaging
Saturn Horticulture Scanners Parking
Subaru Livestock Outsourcing Public.Storage
Suzuki Business.Education Signage Rail.Transport
Toyota Business.Finance Civil.Engineering Computer.Hardware
Scion Investment.Banking Electricity Computer.Components
Volkswagen Risk.Management Nuclear.Energy Computer.Memory
Volvo Venture.Capital Oil...Gas Hard.Drives
Auto.Interior Financial.Markets Waste.Management Network.Storage
Car.Electronics Business.Operations Recycling Copiers
Car.Audio Human.Resources Data.Management Desktop.Computers
Car.Video Management Hospitality.Industry Computer.Security
Body.Art Business.Process Event.Planning Network.Security
Cosmetic.Surgery Project.Management Food.Service Audio.Equipment
Fitness Project.Management.Software Restaurant.Supply Headphones
Bodybuilding Strategic.Planning Generators Speakers
Hair.Care Supply.Chain.Management Heavy.Machinery Camera.Lenses
Hair.Loss Business.Services Manufacturing Cameras
Massage.Therapy Consulting Small.Business Nintendo
Weight.Loss Corporate.Events Home.Office Sony.PlayStation
Marketing.Services Knowledge.Management Aviation Infectious.Diseases

20
 Table 7: Final GT data features selected by PDC-SIS+

Finance lag2 Food Production Consumption Farm

Construction lag2 Agriculture Ford lag1 Energy sector
Finance Construction lag1 Electricity lag1 Finance lag1
Food.Service lag1 Audi Administrative.support lag2 Consumption lag2
Audi lag2 Electricity Expenditure lag2 Automotive.Industry lag2
Business. sector lag1 Ford Food Production lag2 Business.sector

Figure 3: Variable importance plot

Official data GT data

Figure 4: Variable selection procedure for GT data

21