See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/331837874

Big Data and Machine Learning for Economic Cycle Prediction: Application of
Thailand’s Economy

Chapter · March 2019

DOI: 10.1007/978-3-030-14815-7_29

CITATIONS READS

0 200

2 authors:

Chukiat Chaiboonsri Satawat Wannapan

Chiang Mai University Chiang Mai University
89 PUBLICATIONS 307 CITATIONS 27 PUBLICATIONS 10 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Econometrics project development View project

All content following this page was uploaded by Satawat Wannapan on 29 May 2020.

The user has requested enhancement of the downloaded file.

 Big Data and Machine Learning
for Economic Cycle Prediction: Application
of Thailand’s Economy

Chukiat Chaiboonsri and Satawat Wannapan(&)

Puey Ungphakorn Centre of Excellence in Econometrics, Faculty of Economics,

Chiang Mai University, Chiang Mai, Thailand
[email protected], [email protected]

Abstract. Since traditional econometrics cannot guarantee that the parametric

estimation based on some of time-series variables provides the best solution for
economic predictions. Interestingly, combining with mathematics, statistics, and
computer science, the big data analysis and machine learning algorithms are
becoming more and more computationally highlighted. In this paper, 29 yearly
collective factors, which are qualitative information, quantitative trends, and
social movement activities, are employed to process in three machine learning
algorithms such as k-Nearest Neighbors (kNN), Tree models and random forests
(RF), and Support vector machines (SVM). Technically, collective variables
using in this paper were observed from the source agents who successfully
accumulated data details from trends of the world for easily accessing, for
instance, Google Trends or World Bank Database. With advanced artificial
calculations, the empirical result is very precise to real situations. The predicting
result also clearly shows Thailand economy would be very active (peak) in the
upcoming quarters. Consequently, this advanced artificial learning successfully
done in this paper would be the new approach to helpfully provide policy
recommendations to authorities, especially central banks.

Keywords: Macroeconomics
Machine learning
Big data

Econometric forecast

1 Introduction

It is inevitable to state that the world is closed to be completely united by an enormous

amount of data. Every day we talk, criticize, analyze, and even mobilize. These
activities create a huge category of information, which is interestingly implied to
humans’ evolutional thinking. However, the problem is information, frequently refer-
red as “big data”. It is extremely vast and difficult to observe and predict simultane-
ously. Additionally, the study on big data calculations essentially needs more than one
science and one innovation. Consequently, it seems to be very rare that the big data
analysis is considered into the area of social science, especially in fiscal and monetary
economics.
Since more than one hundred years economists have been relied on only the tra-
ditional estimating assumption, which is to fix non-considerable variables to be

© Springer Nature Switzerland AG 2019

H. Seki et al. (Eds.): IUKM 2019, LNAI 11471, pp. 347–359, 2019.
https://doi.org/10.1007/978-3-030-14815-7_29
348 C. Chaiboonsri and S. Wannapan

constant. Graphically, this assumption is just one of sciences used for data analyses,
which is displayed in Fig. 1. It is not enough and makes a cause for many researchers
have to put more assumptions to guarantee that their estimating models can compu-
tationally fit for data. Furthermore, this unmitigated mistake is being still employed to
many econometric predictions and used to provide policy recommendations to
authorities, especially central banks. Accordingly, to escape from it, Machine Learning
(ML) is invented to efficiently combine with big data analyses, and this would be the
solution for clarifying and forecasting the economy’s trends, hidden signs, and crises in
the modern era of academic researches.

Fig. 1. The conceptual framework of research

Considering into central banks, it is obvious their outcomes are also microeconomic
decisions and interactions. These responsibilities came with the collection and access to
a wealth of new data sources, which moves central banks into the realm of big data
(Chakraborty and Joseph 2017). Historically, applying data science investigations and
machine learning coined in 1959 by Samuel (1959) is rare in econometrics for eco-
nomic researches. For central banks, however, these data computing analyses are
becoming more highlighted. For example, the paper of Bholat (2015) was proposed to
situate topics within the context of strategic plans and initiatives and summarized the
article linking central banks’ emerging interest in Big Data approaches. The letter of
Bank of Italy proposed by Signorini (2018) was mentioned the actual and potential
value of big data is more and more crucial for economic researches. Moreover, the
academic issue stated by Daniel Hinge (2017) was expressed that machine learning
may not yet be at the stage where central bankers are being replaced with robots, but
the field is recently bringing powerful tools to bear on big economic questions.
Accordingly, there is no reason why big data and machine learning cannot be the
possibly suitable solution for monetary-econometric researches in Thailand.
 Big Data and Machine Learning for Economic Cycle Prediction 349

2 The Objective and Scope of Research

The fundamental aim of this paper is to computationally predict Thailand’s economic

structural trends by applying big data and machine learning. Interestingly, mixed
observations such as qualitative survey details and time-trend data series during 2004 to
2017 are being employed to do an econometric estimation by artificial intelligent
approaches. All of variables are described and presented in Table 1. Technically,
collective variables using in this paper were observed from the reliable source agents
who successfully accumulated information from trends of the world for convenient
accessing, for instance, Google Trends, World Bank Database, or YouTube Searching
Engine.

Table 1. The details of collective information used to data science analyses reference from the
Thailand’s Key Macroeconomic (Identified by BOT) and Big data from Google Trends database
Variable Detail Symbol Time-series Source
range
GDP (growth rate) RBC_GDP 2004–2017 The World Bank
Database
Population (growth rate) POP 2004–2017 The World Bank
Database
Unemployment (% change) UN_EM 2004–2017 The World Bank
Database
Industrial value added (% per GDP) IND 2004–2017 The World Bank
Database
Agricultural value added (% per GDP) AGR 2004–2017 The World Bank
Database
Gross domestic investment (% per GDI 2004–2017 The World Bank
GDP) Database
Service imports (% per GDP) SER_IM 2004–2017 The World Bank
Database
Service exports (% per GDP) SER_EX 2004–2017 The World Bank
Database
Military expenditures (% per GDP) MIL 2004–2017 The World Bank
Database
Public debt (% per GDP) PU_DEP 2004–2017 The World Bank
Database
Consumer price index (% per annual) CPI 2004–2017 The World Bank
Database
Foreign direct investment (US$) FDI 2004–2017 The World Bank
Database
Available lands (km2) LAND 2004–2017 The World Bank
Database
Preserved forests (km2) FOR 2004–2017 The World Bank
Database
(continued)
350 C. Chaiboonsri and S. Wannapan

Table 1. (continued)
Variable Detail Symbol Time-series Source
range
International reserves INT 2004–2017 The World Bank
Database
Fixed interest rate FIX_I 2004–2017 The World Bank
Database
Exchange rate (Baht per US$) EX 2004–2017 The World Bank
Database
Official development assets (US$) ODA 2004–2017 The World Bank
Database
Values of stock market trading MKT 2004–2017 The World Bank
Database
Overall Thailand economic situations Big_data 1 2004–2017 Google shopping
database
Investment situations Big_data 2 2004–2017 Google shopping
database
Stock market situations Big_data 3 2004–2017 Google shopping
database
Thailand business movements Big_data 4 2004–2017 Google shopping
database
Employments Big_data 5 2004–2018 Google shopping
database
Thailand international trades and Big_data 6 2004–2019 News search
investments
Thailand agricultural situations Big_data 7 2004–2020 New search
Thailand industrial situations Big_data 8 2004–2017 Image search
Thailand banking situations Big_data 9 2004–2018 Google shopping
database
Thailand political atmospheres Big_data 2004–2019 YouTube search
10
Thailand social atmospheres Big_data 2004–2020 YouTube search
11

3 Theory and Methodology

3.1 The Fundamental Concept of Machine Learning in Data Science
With complex components to perform the machine learning computation, there are
essential approaches described as follows: firstly, decision tree learning is used to make
a decision tree as a predictive model, which maps observations about an item in order
to conclude about the item’s target value. Second, an artificial neural network
(ANN) learning algorithm is employed to model complex relationships between inputs
and outputs or to capture the statistical structure in an unknown joint probability
distribution between collective variables. Third, inductive logic programming (ILP) is
 Big Data and Machine Learning for Economic Cycle Prediction 351

an approach used to rule learning using logic programming as a uniform representation

for input samples, background knowledge, and hypotheses. Forth, support vector
machines (SVMs) are an array of related supervised learning methods used for clas-
sification and regression. Fifth, clustering analysis is a method of unsupervised
learning, and a common technique for statistical data analysis. Sixth, Bayesian net-
works are a belief network that presents a set of random variables and their conditional
independencies via a directed acyclic graph (DAG). This modern statistics can effi-
ciently performs algorithm exist for inference and learning. Seventh, representation
learning (RL) is employed to disentangle the underlying factors of variation that
explain the observed data (Bengio 2009). Eighth, a genetic algorithm (GA) is a search
heuristic that mimics the process of natural selection, which is used to improve the
performance of genetic and evolutionary algorithms (Zhang et al. 2011). Ninth, rule-
based machine learning is a contrast model to other machine learners that commonly
verifies a singular model for predicting (Bassel et al. 2011). Rule-based machine
learning approaches include learning classifier systems, association rule learning, and
artificial immune systems. Tenth, feature selection approach is employed to select an
optimal subset of relevant features for use in model construction.

3.2 Algorithms for Machine Learning Analyses

(a) k-Nearest Neighbors (kNN)
k-Nearest Neighbors (kNN) is a non-parametric method for both classification and
regression problems (Chakraborty and Joseph 2017). Data is performed as its k-nearest
to others in the feature space. For solving the problem, data observations are assigned
to the majority class of its nearest observations, for example, the mean value of its
nearest neighbors in the regression model. The performance of k-NN can be analyzed
by focusing on the error rate of miss-classified examples for classification or squared
errors for regression problems, respectively. The output value for a single sample xi is
achieved as following the steps, which are neighbor selection and value assignment.
The former is the calculation of the distance xi to all other points in the feature space.

k
The stage determines its k closet neighbors xj i . Euclidean distance is commonly
used as a distance measure. The latter is conducted to assign the output yi the class
membership (i.e.yi 2 fC1 ; C2 ; . . .; Cc g 8i ¼ 1; 2; . . .; n, where c is a number of class
layers) by the majority vote of its k nearest neighbors. The k-NN regression, yi , is
delegated to be the average value of its single nearest neighbor, yi ¼ 1=k
Pk
k xj . One can be also considered as a distance-weighted average.
j¼1;xi 2fxj gi

(b) Support vector machines (SVM)

Support vector machines (SVM) are a powerful technique for both classification and
regression type problems (Chakraborty and Joseph 2017). Generally, two-class clas-
sification problems are modeled by logistic regressions (Logit model), the position to
hyperplane in the feature space is projected to a (0,1) interval, which can be defined as
probabilities of class membership. Computationally, An SVM will try to find a decision
boundary to separate these two classes by the maximal margin, represented by the gray
area (as shown in Fig. 2). The separating boundary (black line), which maximizes the
352 C. Chaiboonsri and S. Wannapan

margin, is, however, not straightforward in the general case. Thus, the concept behind
SVMs is now two-fold. First, to solve a presentation of the feature spaces in which the
data is linearly separable and, second, to identify the points in the input space which
define, or support, the maximal margin, the support vectors. As seen in Fig. 2, the
separating boundary, decision rule and error function for two-class classification can be
expressed as
 T 
x
b  1 support vectors
SV  
hðxi ; bÞ ¼ sign xTi :b P hypothesis
m
ERRðX; Y; bÞ ¼  2m 1
i¼1 ðjyi  hðxi ; bÞjÞ error function:

Fig. 2. Schematic presentation of a two-class (green and red dots) support-vector classifier in a
two-dimensional feature space. (Color figure online)

The coefficients b define the hyperplane satisfying the decision boundary equation.
The support vectors lying on the boundary of the gray area and the target values, yi ,
take the values yi 2 1; 1. The coefficients b can be provided by solving the (dual)
optimization problem:

X
m
1
LðaÞ ¼ ai  aT :H:a;
i¼1
2

X
m
with b ¼ ai yi xi ; Hij  yi yj xTi :xj ; ð1Þ
i¼1

under the constraints a:Y ¼ 0 and ai  0. Considering into the Lagrangian LðaÞ, it
focuses though there is a large number of free parameters ai , namely one for each
observation xi . The input data X enters the Lagrangian only via an inner product,
resulting in a scalar. This allows to define transformations T(.) and inner products via a
so-called kernel,
 Big Data and Machine Learning for Economic Cycle Prediction 353

K ð^x; x0 Þ ¼ hT ð xÞ; T ðx0 Þi: ð2Þ

Commonly, the radial basis function (Gaussian kernel) relied on the polynomial
expansion of the exponential function is employed to be the choice for a kernel
function,
h i
K ð^x; x0 Þ ¼ exp cjjx  x0 jj2 : ð3Þ

(c) Tree models and random forests

Tree models are one of popular non-parametric machine learning techniques for
regression and classification problems. Conceptually, the training dataset based on the
input features until assignment criterion with respect to the target variable is consec-
utively divided into a data basket (leaf). The purpose is to minimize the entropy
H ðYjX Þ (objective function) within areas of the baskets conditioned on the features
X. At the beginning, the full set X of m observations is identified for conditioning on the
features x, which leads to the highest information gain (I), and this can be expressed as
follows:
P
I ðYjxÞ ¼ H ðYjX Þ  v2fxg jjXmvjj H ðYgXv Þ information gain ðclassificationÞ
P
H ðYjX Þ ¼  Cc¼1 pðY ¼ cjX Þ logðpðY ¼ cjX ÞÞ entropy ðclassificationÞ
P Pmj  2
H ðYjX Þ ¼ m1 kj¼1 i¼1 yi  lj jxi 2Xj MSE ðregressionÞ;

where pðY ¼ cjX Þ is the connected frequency of class c observations in X. jXv j is the set
of observations which take on each value. In a regression setting, the entropy can be
replaced by the mean squared error (MSE) and splits are performed along the
dimensions which most reduce the error (Galton 1907). A schematic representation of a
tree model with two features, x1 and x2, is displayed in Fig. 3.

Fig. 3. Schematic representation of a tree model. Left: the target data space is systematically
segregated by the tree model based on the input features. Right: a tree representation of the final
model. The tree is grown from the top to the bottom. (Modified from Chakraborty and Joseph
2017)
354 C. Chaiboonsri and S. Wannapan

3.3 Model Validation

Similar to econometric models, it is difficult to give the best result from the beginning
when doing the computation of machine learning systems. Adjustments and robustness
checking are essential in most cases. In this paper, Cohen’s kappa coefficient is
employed to validate the best prediction from machine learning algorithms. Cohen’s
kappa coefficient ðjÞ is a statistic tool measuring inter-rater agreement for qualitative
(categorical) items. It is mostly implied to be a more robust measurement than simple
percent agreement calculation. However, the complexity of Cohen’s kappa interpre-
tation is the obstruction to use this coefficient. Some researchers have suggested that it
is conceptually simpler to evaluate disagreement between items (Pontius and Millones
2011). The calculation of Cohen’s kappa measurements is to verify the agreement
between two raters who each classify N items into C mutually exclusive categories. The
coefficient ðjÞ is defined as:

p0  pe 1  p0
j ¼1 ; ð4Þ
1  pe 1  pe

where p0 is the relative observed agreement among raters, which identifies to accuracy,
and pe is the hypothetical probability of chance agreement. The observed data is used to
calculate the probabilities of each randomly observable seeing in each category. If ðjÞ
equals one, this implies that the raters are in complete agreement. On the other hand, if
ðjÞ equals zero then here is no agreement among the raters other than what would be
expected by chance. Cohen’s kappa coefficient is possible to be negative. This implies
that there is no effective agreement between the two raters or the agreement is worse
than random. For categories k, a number of items N and nki is the number of time raters
i predicted category k,

1 X
pe ¼ nk1 nk2 : ð5Þ
N2 k

Interestingly, the magnitude of kappa reflects adequate agreement is depended on

two important factors, which are prevalence (doing equiprobable codes) and bias (the
marginal probabilities for the two observers similar or different). For codes, Kappas are
obviously higher when codes are distributed asymmetrically by the two observers. This
is contrast to probability variations, which the effect of bias is greater when Kappa is
small than when it is large (Sim and Wright 2005). The Kappa calculation shown in
Fig. 4 is from the same simulated binary data. Each point on the graph is computa-
tionally generated from a pairs of judges randomly rating 10 subjects for having a
diagnosis of X or not. Experimentally, a Kappa equals zero is approximately equivalent
to an accuracy = 0.5.
 Big Data and Machine Learning for Economic Cycle Prediction 355

Fig. 4. Kappa accuracy and coefficient presentation

4 Empirical Results
4.1 Descriptive Information
First of all, taking consideration into Thailand economic trends, yearly collective GDP
during 2004 to 2017 is displayed in Fig. 5. The classification of the trend obviously
shows that Thailand economy is dramatically fluctuated. This fluctuation causes the
prediction to become more difficult, doing traditionally econometric tools alone cannot
precisely provide the best model. In particularly, general means are not reliable any-
more. Consequently, the optimal algorithm calculation called Newton Method is
employed to extend the ability of data explanations.

Fig. 5. The presentation of economic trends in Thailand during 2004 to 2017

To mathematically solve the issue, the basic idea of Newton’s method is typically
used as a basis on linearization. Graphically, the convergence processing of Newton’s
method is approximated by its tangent line, which is already gone closer to the root at
x* (as shown in Fig. 6). Empirically, the result represents that the real middle value of
data observations is quite different from the mean (Seen details in Table 2).
From the result of the optimal value, the economic trend can be clarified as a period
of real business cycles. The peak period is defined at the level, which is higher than the
optimal value (3.55%). Expansion is in the interval between 2.1% and 3.5%. Recession
belongs to the interval between 0% and 2%, and fall is defined as the level below 0%.
The details are demonstrated in Table 3.
356 C. Chaiboonsri and S. Wannapan

Fig. 6. Presentation the convergence processing in Newton-optimal approach (Modified from

Satawat et al. 2018)

Table 2. The comparison between the optimal value and normal mean of yearly Thailand GDP
Variables (Growth rate) General mean Optimal value (Newton’s method)
GDP 3.67% 3.55%

Table 3. The demonstration of real business cycles in Thailand economic trends

Growth rate Description
>3.55% Peak
2.1% to 3.5% Expansion
0% to 2% Recession
<0% Fall

4.2 The Results of Machine Learning Algorithm

This stage is the model comparison. For the comparison of different models, an
approach of training, validation and testing suitable to the given dataset are adopted.
Three different such approaches would raise an alert if either one, two or three or more
outliers are detected within the four model features such as k-NN, random forest, and
SVM, respectively. The results are shown in Table 4. In this computation, the random
forest (rf) is the best model and contains the highest parametric values when selecting
by both the values of accuracy and Kappa coefficient, which are 0.7 and 0.4167,
respectively. Moreover, this comparison result is graphically displayed in Fig. 7.
Obviously, as the random forest model is chosen to be the algorithm predictor for
Thailand macroeconomic variables, the forecasting result strongly confirms that Big
Data using in this computational paper can provide a very sensibility in a similar way to
the real situation, more precisely than traditional econometric estimations. The details
are represented in Table 5. Empirically, the predicting result clearly shows Thailand
economy would be very active (peak) in the upcoming quarters. In other words, the rate
of economic expansion would be 4.5%, higher than the optimal value (3.55%) during
the third quarter to fourth quarter.
 Big Data and Machine Learning for Economic Cycle Prediction 357

Table 4. Comparative performance statistics of various machine learning models for predicting
Thailand macroeconomic variables
Method Cross-validation The final values Accuracy Kappa’s coefficient
Random forest (rf) Cohen’s kappa k = 9 0.7 0.4167
k-NN (kn) Cohen’s kappa c = 0.25 0.6 0
sigma = 1.286084e-18
SVM (svm) Cohen’s kappa mtry = 15 0.5417 0

Fig. 7. The model selection in machine learning systems by Cohen’s kappa

Table 5. The overall prediction from three machine learning algorithms

Years RBC_GDP RBC_GDP RF*** KNN SVM
(Current values) (Description) (Prediction) (Prediction) (Prediction)
2018(Q1) (4.8%) Peak Peak Peak Peak
2018(Q2) (4.2%–4.7%) Peak Peak Peak Peak
2018(Q1–Q4)f (4.5%) Peak Peak Peak Peak

5 Conclusion

The huge challenge to apply the advanced computational method, as called “machine
learning algorithms” for predicting the big data in macroeconomic variables is suc-
cessfully done in this paper. Since it is totally different from traditionally parametric
estimations and is more powerful. The machine learning systems can capture an
enormous amount of informative details in databases, including qualitative data,
quantitative factors, and even time-series trends. In this paper, 29 time-series variables
regarding to Thailand economic structures during 2004 to 2017 were included to
predict the sensible upcoming trend.
With the minimization of modeling assumptions, machine learning systems can
efficiently calculate both stationary and non-stationary data. Methodologically, the data
was processed into two sections, algorithm calculations and model selections.
358 C. Chaiboonsri and S. Wannapan

Empirically, the random forest algorithm is the best model selected by Cohen’s kappa
coefficient. Consequently, the predicting result clearly shows Thailand economy would
be very active (peak) in the upcoming quarters (2018q3 to 2018q4).
To interpret this computational result and apply to policy recommendations, it is
obvious that the big data using in this paper dose not includes only economic factors,
but it contains social variables such as political activities, environmental issues, and
even IT social behaviors. With machine learning algorithms, the result which is con-
cluded that Thailand economy would be the sunshine time is extremely reliable much
more than traditional econometric models for time-series forecasting. Since the com-
plexity in multi-processing estimations by AI, machine learning methods can explain
the outliners in the mixed observation rather than traditional econometric methods,
which inevitably need assumptions. The solution can confirm that political stability,
social network updating, fluctuation controlling in financial market systems, online
news, and even efficiently environmental managements are inevitably connected, and
these points have to be empirical implemented. Hence, there is no reason why big data
and machine learning cannot be the suitable answer for monetary-econometric
researches in Thailand, especially data mining for researching into the responsibilities
of BOT (Bank of Thailand).

References
Bassel, G.W., Glaab, E., Marquez, J., Holdsworth, M.J., Bacardit, J.: Functional network
construction in Arabidopsis using rule-based machine learning on large-scale data sets. Plant
Cell 23(9), 3101–3116 (2011)
Bholat, D.: Big data and central banks. Q. Bull. Q1, pp. 86–93 (2015). https://www.researchgate.
net/publication/276101527_Big_Data_and_central_banks
Bishop, C.M.: Pattern Recognition and Machine Learning. Springer, New York (2006). ISBN
0-387-31073-8
Chakraborty, C., Joseph, A.: Machine learning at central banks. Staff Working Paper No. 647.
Bank of England (2017)
Galton, F.: Vox populi. Nature 75, 450–451 (1907)
Hinge, D.: Big Data in Central Banks. Published by Infopro Digital Services Ltd, Central
Banking Publications, London (2017)
Landis, J.R., Koch, G.G.: The measurement of observer agreement for categorical data.
Biometrics 33(1), 159–174 (1977). https://doi.org/10.2307/2529310
Pontius, R., Millones, M.: Death to Kappa: birth of quantity disagreement and allocation
disagreement for accuracy assessment. Int. J. Remote Sens. 32, 4407–4429 (2011)
Samuel, A.L.: Some studies in machine learning using the game of checkers. In: Levy, D.N.L.
(ed.) Computer Games, pp. 335–365. Springer, New York (1988). https://doi.org/10.1007/
978-1-4613-8716-9_14
Sim, J., Wright, C.C.: The Kappa statistic in reliability studies: use, interpretation, and sample
size requirements. Phys. Ther. 85, 257–268 (2005)
Signorini, L.F.: Harnessing big data & machine learning technologies for central banks. The
Printing and Publishing Division, Bank of Italy, Rome (2018)
 Big Data and Machine Learning for Economic Cycle Prediction 359

Wannapan, S., Chaiboonsri, C., Sriboonchitta, S.: Macro-econometric forecasting for during
periods of economic cycle using bayesian extreme value optimization algorithm. In:
Kreinovich, V., Sriboonchitta, S., Chakpitak, N. (eds.) TES 2018. SCI, vol. 753, pp. 706–723.
Springer, Cham (2018). https://doi.org/10.1007/978-3-319-70942-0_51
Zhang, J., et al.: Evolutionary computation meets machine learning: a survey. IEEE 6(4), 68–75
(2011)

View publication stats