Mining and Linguistically Interpreting Data from Questionnaires
Influence of Financial Literacy to Behaviour
Miroslav Hudec
1
and Zuzana Brokešová
2
1
Faculty of Economic Informatics, University of Economics in Bratislava, Dolnozemská cesta 1, Bratislava, Slovakia
2
Faculty of National Economy, University of Economics in Bratislava, Dolnozemská cesta 1, Bratislava, Slovakia
Keywords: Financial Literacy, Questionnaire, Data Mining, Quantified Sentence of Natural Language, Flexible Data
Summarization, Fuzzy Logic.
Abstract: This paper is focused on mining and interpreting information about effect of financial literacy on
individuals’ behavior from the collected data by soft computing approach. Fuzzy sets and fuzzy logic allows
us to formalize linguistic terms such as most of, high literacy and the like and interpret mined knowledge by
short quantified sentences of natural language. This way is capable to cover semantic uncertainty in data and
concepts. The preliminary results in this position paper have shown that for majority of people of low
financial literacy angst and other treats represent serious issues, whereas about half of people with high
literacy do not consider these treats as significant. Finally, influence of literacy to anchoring questions is
mined and interpreted. Eventually, the paper emphasises needs for further data analysis and comparison.
1 INTRODUCTION
Adequate level of financial knowledge was
identified as important for sound financial decisions
(Lusardi, 2008). In addition, previous researches
indicate that individuals’ financial decisions are
affected by different psychological heuristics and
biases (Tversky and Kahneman, 1975). They include
emotional aspects in decision-making, loss aversion,
anchoring, framing and many others (see Gilovich et
al., 2002). According to our knowledge, the effect of
financial literacy to these heuristics and biases has
not been widely examined in the literature. Hence, in
the paper, we test the effect of high financial literacy
to (i) feeling like angst, nervousness, loss of control
and fear regarding possible catastrophic scenarios;
(ii) how these people are influenced by anchor
questions; (iii) how they decide about risky
investments.
The second task is mining and interpreting this
effect from the data collected by surveys in an easily
understandable and interpretable way. First,
summaries from the data are better understandable if
they are not as terse as numbers (Yager et al., 1990).
Secondly, there are often uncertainties in answers,
which should not be neglected (Hudec, 2015; Viertl,
2011). Thirdly, answers to respective questions
might be numbers, categorical data and short texts.
Hence, mining and interpreting survey data by
computational intelligence is beneficial. Fuzzy sets
and fuzzy logic allows us to mathemiatically
formalize linguistic terms such as most of, low
literacy, high angst and the like. They make possible
partial membership degrees of elements near the
borderline cases to these concepts.
In order to meet both aforementioned challenges,
we have conducted survey in Slovakia and adjusted
Linguistic Summaries (LSs) in order to mine and
interpret relational knowledge between financial
literacy and respective attributes.
2 LINGUISTIC SUMMARIES
AND THEIR QUALITY
LSs have been initially introduced by Yager (1982)
in order to express summarized information from the
data by linguistic terms instead of numbers.
Overview of recent development can be found in
(Boran et al., 2016). LSs of the structure Q R entities
in a data set are (have) S developed by Rasmussen
and Yager (1997) are relevant for our research. One
example of such summary is most of low financially
literate people feel high level of angst. The validity
of such a LS is computed as
Hudec, M. and Brokešová, Z.
Mining and Linguistically Interpreting Data from Questionnaires - Influence of Financial Literacy to Behaviour.
DOI: 10.5220/0006462402490254
In Proceedings of the 6th International Conference on Data Science, Technology and Applications (DATA 2017), pages 249-254
ISBN: 978-989-758-255-4
Copyright © 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
249

)
)(
))(),((
())((
1
1





n
i
iR
iR
n
i
iS
Q
xμ
xμxμt
μPxQxv
(1)
where n is the cardinality of elements in a data set
(in our case number of respondents),




n
i
iR
iR
n
i
iS
xμ
xμxμt
1
1
)(
))(),((
is the proportion of the elements
that belong to restriction R and satisfy summarizer S
(fully or partially), t is a t-norm, µ
S,
µ
R
and µ
Q
are
membership functions explaining summarizer S,
restriction R and relative quantifier Q, respectively.
The truth value (validity) v gets value from the
unit interval. If v is closer to the value of 1, then the
relation between R and S explained by Q is more
significant. The goal of LS is to reveal such
relations. LSs are graphically illustrated in Figure 1,
where grey areas between sets low, medium and high
emphasizes the uncertain area, i.e. area where
unambiguous belonging to a particular set cannot be
arranged.
Flexible summarizers, restrictions and
quantifiers are mathematically formalized by fuzzy
sets. For instance, fuzzy set around anchor value of
m can be expressed as triangular fuzzy set (Figure
2).
attribute in
summarizer (S)
high validity
r
1
r
n
...
low
medium
high
r
n
...
low
medium
high
r
1
attribute in
restriction (R)
medium validity
Figure 1: Graphical illustration of LSs with restriction.
μ
A
(x)
0
1
a
X
b
m
Figure 2: Triangular fuzzy set around m.
otherwise0
)(
)(
1
)(















m,bx
mb
xb
a,mx
m-a
x-a
mx
x 
(2)
where belonging to the set decreases when distance
to the anchor m increase.
The benefit against interval is in the intensity of
belonging to a set. The closer is element to
boundaries, the lower matching degree it has. Thus,
defining boundaries is not as sensitive as for
classical intervals. Similarly, fuzzy sets expressing
concepts small and high are plotted in Figure 3. For
categorical data matching degree is directly assigned
to each element, i.e.
))}(,()),...,(,{(
nn11
xxxxFL  
(3)
where FL is a fuzzy set expressing concept in
summarizer or restriction.
If LS with restriction has high validity v (1), it
does not straightforwardly mean that such LS is
relevant. In order to solve this problem several
quality measures were suggested in (Hirota and
Pedrycz, 1990). Due to their complexity and partial
overlapping, simplified quality measure merging
validity and coverage is suggested in (Hudec, 2017)
otherwise0
5.0) ,(





CCvt
Q
c
(4)
where C is data coverage and t is a non-idempotent
t-norm, e.g. product one. Coverage is calculated
form the index of coverage.
))(μ),(μ(
1
n
xxt
i
iR
n
i
iS
C



(5)
where parameters have the same meaning as in (1)
and n is cardinality of data set by the transformation
(Wu et al, 2010)
,1
2
,)(21
2
,)(2
,0
)(
2
2
21
2
12
2
21
1
2
12
1
1





















ri
ri
rr
rr
ir
rr
ir
rr
ri
ri
ifC
c
c
c
c
c
c
C
(6)
where r
1
= 0.02 and r
2
= 0.15, because in LSs with
restriction only small subset of data is included in
both sets R and S (Figure 1).
μ
H
(x)
0
1
a
m
X
μ
L
(x)
0
1
m
b
X
highlow
Figure 3: Fuzzy set expressing concepts small and high.
DATA 2017 - 6th International Conference on Data Science, Technology and Applications
250

3 MINING RELATIONAL
KNOWLEDGE AMONG
ATTRIBUTES
This section explains procedure, results, discussion
and challenges for further work. The data set
consists of 644 answers with no missing data.
Examples of questions, possible answers and fuzzy
sets are provided in successive sections.
3.1 Experiments
The first step consists of construction of membership
functions for attributes and their respective linguistic
labels appearing in summaries.
Question regarding financial literacy is used in
all experiments. The level of financial literacy was
not included in the questionnaire. It was aggregated
from answers to several questions. Financial literacy
is expressed on the [0, 5] scale, where 0 is the lowest
and 5 the highest level. Other questions were
directly focused on feelings to potential kinds of
threats, willingness to participate in financial games
and influences by anchor questions.
The level of financial literacy attribute is
fuzzified into three fuzzy sets: low literacy (LL),
medium literacy (ML) and high literacy (HL) using
notation (3) in the following way
HL = {(5, 1), (4, 0.75), (3, 0.45)}
ML = {(1, 0.35), (2, 1), (3, 1), (4, 0.35)}
LL = {(0, 1), (2, 0.75), (3, 0.45)}
It means that level 5 is without doubt high, level
4 significantly high but not fully and level 3 is
partially high. Other two levels do not belong to this
set. The recognized drawback in fuzzification is in
set ML. It contains two elements with matching
degree equal to 1. Hence, higher numbers of
elements belong to this set. The ideal option is using
scale of odd number of elements. Anyway, sets HL
and ML do not cope with this issue.
3.1.1 Relation between Financial Literacy
and Attributes of Fear, Angst,
Nervousness and Loss of Control
Emotional heuristics and biases including fear,
angst, nervousness or loss of control could
significantly shape financial decision (e.g. Lee and
Andrade, 2015). For example, in the case of
insurance purchase, Zaleskiewicz et al. (2002)
identified increase in the flood insurance demand
after the recent flood. The role of financial literacy
in this relation is questionable.
Our respondents were asked about emotions
(angst, fear, nervousness and loss of control) evoked
by imagination of possible negative events that
could affect their life. These imagination was
visualized in the pictures of unexpected natural
disaster, death of relatives, war and deadly epidemic.
Possible answers are: not at all, weakly, moderately,
intensely and very intensely. These categorical
attributes are fuzzified into sets low and high by (3):
L = {(not at all, 1), (weakly, 0.75), (moderately, 0.45)}
H = {(very intensely, 1), (intensely, 0.75), (moderately,
0.45)}
In order to get summaries, covered by sufficient
number of respondents, we have strengthened Eq.
(4) by condition C ≥ 0.75, instead of 0.5.
Mined results are shown in tables 1 and 2, where
terms low and high corresponds with terms
illustrated in Figure 1. Financial literacy is
restriction (R) and summarizers (S) are considered
attributes. First row in Table 1 means that LS about
half respondents with high financial literacy have
low angst has significantly higher validity than
sentence explained by quantifier most of.
Mined relational knowledge (Table 2) has shown
that people with low financial literacy have high
level of fear, angst, loss of control and nervousness.
The most problematic attribute is angst, where
significant proportion of respondents with low
Table 1: Result of summaries Q respondents of high
financial literacy have low values of respective attributes.
high financial literacy
low
validity for quantifier
Eq. (1)
coverage Eq.
(6)
angst
about half - 0.7129
most of - 0.1411
0.7888
loss of control
about half – 1
1.0000
fear
about half – 1
0.9988
nervousness
about half – 1
0.9346
Table 2: Result of summaries Q respondents of low
financial literacy have high values of respective attributes.
low financial literacy
high
validity for quantifier
(1)
coverage (6)
angst
most of - 0.9996
about half – 0.0004
1
loss of control
most of - 0.5863
about half – 0.4137
1
fear
most of - 0.5818
about half – 0.4182
1
nervousness
most of - 0.7674
about half – 0.2326
1
Mining and Linguistically Interpreting Data from Questionnaires - Influence of Financial Literacy to Behaviour
251

financial literacy consider angst as a treat. These
people might think that potential dangerous
situation, if appear, will significantly devalue their
lives and properties. On the other side, majority of
people with high financial literacy does not have
straightforwardly low intensities of these treats
(Table 1). But at least about half of them have low
values of respective threats.
3.1.2 Relation between Financial Literacy
and Answer to Anchor Question
Anchoring represents a subconscious tendency of
individuals to rely on onward information in the
decision-making. This information does not have to
be necessarily important for the decision. The role of
financial literacy in the process of anchoring has not
been previously studied.
In this experiment, we worked with numerical
attribute: number of inhabitants in Iowa. Anchor
value was set to 1 500 000 inhabitants. In order to
reveal behaviour of people who have answered
around this anchor value, we had relaxed crisp value
1 500 000 to the fuzzy number around 1 500 000,
which is a convex fuzzy set with limited support and
for value of 1 500 000 the matching degree is equal
to 1. Fuzzy set around 1 500 000 is represented by
(2) with the following values of parameters:
a = 1200000, m = 1500000, b = 1800000 (Figure 2).
Benefits against classical interval are explained in
Section 2.
In this experiment, significant relations between
literacy and answers to anchor questions have not
been recorded by LSs. From the opposite view
(restriction is anchor question and summarizer is
financial literacy), we recorded high validity for
summary: most of respondents with answer around
anchor have medium literacy. These relations are
summarized in Table 3.
Table 3: Result of summaries Q respondents with answer
around anchor value have (low, medium, high) values of
financial literacy.
number of inhabitants around the anchor value
validity for quantifier
(1)
coverage (6)
low financial
literacy
few – 0.5686
about half – 0.4314
0,1249
medium finan.
literacy
most of – 1
0,9456
high financial
literacy
few - 0.3018
about half – 0.6981
0,1918
We could conclude that individuals with high level
of financial literacy do not reflect the anchor in their
decision-making. However, also respondents with
low level of financial literacy do not react to
anchoring. Anyway, this experiment requires further
analysis. We consider trying different terms
regarding answers, for instance, far from 1 500 000,
significantly lower than 1 500 000, significantly
higher than 1 500 000 and the like. These terms can
be formalized by fuzzy sets plotted in Figure 3.
Another option is approach based on fuzzy
functional dependencies, which might be helpful due
to comparing conformance of each two respondents
answer on both attributes.
3.1.3 Relation between Financial Literacy
and Risk Taking
Risk aversion is an important parameter in financial
decisions as majority of these decisions include risk.
However, the effect of level of financial literacy is
not clear. In economic literature, risk taking is often
measured by the decisions in the lottery, where
individuals could choose between sure and risky
alternative (see e.g. Holt and Laury, 2002).
In our experiments, participation in lottery is
numerical attribute. Respondents were asked how
much they are willing to pay for the participation in
the lottery, where one out of 10 players win 1000 €.
Based on the expected value theory, those who are
willing pay less than 100 € could be considered as
risk averse, those who would like to pay 100 € are
assumed as risk neutral and those who will pay more
than 100 € are recognized as risk loving. The lowest
recorded value was 0 and the highest 333, whereas
mean value was 21.30 €. Data distribution is not
uniform, which means that we cannot create three
granules: low, medium and high by uniformly
covering domain. The options are statistical mean
based method (Tudorie, 2008) or logarithmic
transformation (Hudec and Sudzina, 2012).
Applying the former, fuzzy set low participation is
expressed by parameters m = 25 and b = 37.5,
whereas set high participation is expressed by
parameters a = 62.5 and m = 75 (Figure 3). Results
from this experiment are summarized in Table 4.
We have recorded that respondents belonging to
all three categories of financial literacy prefer low
participation in lottery. That is an expected result as
the majority of the population is risk averse and
prefer sure alternative over the risky one. We have
not identified effect of financial literacy on risk
taking. Other relations have not been recorded, due
to low data coverage.
DATA 2017 - 6th International Conference on Data Science, Technology and Applications
252

3.2 Discussion
The preliminary results are promising, but further
research is highly advisable. Regarding questions
angst, loss of control, fear and nervousness, we
reveal that improving financial literacy might be
beneficial for the population (tables 1 and 2).
Although, majority of people with high financial
literacy do not have straightforwardly low angst and
other threats, at least about half of them do not
consider the treats significant. For respondents with
low financial literacy the treats represent problems.
On the other hand, reaction to anchoring is low
in the group with high financial literacy and low
financial literacy as well. Several studies supported
the role of anchor in decision-making (Furnham and
Boo, 2011). The absence of this effect on the two
extremes of financial literacy (high, low) brings an
interesting finding to the literature. One could
speculate that those with high level of financial
literacy notice the anchor but isolate its effect in
decision-making. Those with low levels of financial
literacy presumably do not notice anchor. This
finding requires further research and robustness
checks.
Our results suggest that risk aversion do not vary
with financial literacy. Risk aversion is therefore
superior characteristic and cannot be changed by the
increase of the level of financial literacy.
Table 4: Result of summaries between financial literacy
and participation in lottery using quantifiers few, about
half and most of.
high financial literacy
validity for Q
coverage
low participation
most of – 1
1
high participation
few -1
0.0081<0.75
low financial literacy
validity for Q
coverage
low participation
most of – 1
1
high participation
few -1
0
medium financial literacy
validity for Q
coverage
low participation
most of – 1
1
high participation
few -1
0.2898<0.75
LSs are suitable for mining relational knowledge
from variety of data: numbers (either crisp or fuzzy),
(weighted) categorical data, short texts. The benefit
is in low computational effort when optimization
techniques are applied (Liu, 2011), covering non-
linear dependencies and fast estimation. In order to
reveal relational knowledge, domains are divided
into several flexible granules: low, medium and high.
This ensures that we can use all data types in (1), (4-
6) considering their respective membership degrees.
3.3 Further Tasks and Opportunities
This is, according your best knowledge, first attempt
to analyse impact of financial literacy to people
perceptions and decisions as well as to analyse these
data with linguistic summaries. The framework and
mathematical formulation have been developed. In
next parts of our project, we are going to analyse
other possible combinations of attributes and their
respective granules, such as answer far from anchor,
significantly lower answers, etc.
Answers to majority of questions are linguistic
terms. Such terms do not have clear definition in
terms of set theory. Furthermore, concepts such as
around anchor value are understandable, even
though vague. Hence, we are considering building
model based on Fuzzy Functional Dependencies
(FFD) to compare with our results. Overview of
FFD can be found in, e.g. (Vučetić and Vujošević,
2012). In this way, we can cover mining relational
knowledge form questionnaires by two valuable
approaches capable to manage semantic uncertainty.
Hence, the preliminary results are going to be
confronted by approach based on FFD. Option based
on FFD is more demanding in terms of
computational effort, but is more powerful. It
compares each two values in order to reveal whether
similar values of one attribute causes similar values
of another attribute.
Our results reveal influence of financial literacy
to main threats and other attributes for Slovak
respondents. Promising aspect for research is
comparison with other countries as well as control
the effect of the socio-economic characteristics of
individuals in this relation. To reveal such answers
an international survey and research is advisable.
4 CONCLUSIONS
Financial literacy, its improvement and it drivers is a
relevant topic, which should be analysed. In this
paper, we have shown that mining relational
knowledge from well-designed questionnaires by
fuzzy logic is valuable contribution to this field.
Preliminary data mining results have shown that low
literacy causes that people have lower quality of life,
Mining and Linguistically Interpreting Data from Questionnaires - Influence of Financial Literacy to Behaviour
253

due to high level of angst, nervousness, fear and loss
of control. On the other side, high literacy does not
straightforwardly mean big improvement.
Further, we have recorded that respondents
belonging to all three categories of financial literacy
prefer low participation in lottery. It is an expected
result as the majority of the population is risk averse
and prefer sure alternative to the risky one.
Finally, result that reaction to anchoring is low in
the group of respondents with high financial literacy
and in the group with low financial literacy brings an
interesting finding, which is contribution to further
economical and data mining research.
These conclusions are preliminary. Regarding
the data mining, in the next stage of our research, we
are going to apply more powerful, but also
computationally demanding approach for revelling
flexible functional dependencies among attributes to
confront already reached results. In addition,
advisable are similar researches in other countries
with different economic situation.
ACKNOWLEDGEMENTS
This paper is part of a research grant VEGA
No. 1/0849/15 entitled ‘Economic and social aspects
of the information asymmetry in the insurance
market’ supported by the Ministry of Education,
Science, Research and Sport of the Slovak Republic.
REFERENCES
Boran, F. E., Akay, D., and Yager, R. R. 2016. An
overview of methods for linguistic summarization
with fuzzy sets. Expert Systems with Applications,
61:356–377.
Furnham, A., Boo, H. C. 2011. A literature review of the
anchoring effect. The Journal of Socio-Economics,
40(1):35-42.
Gilovich, T., Griffin, D., Kahneman, D. 2002. Heuristics
and Biases: The Psychology of Intuitive Judgment,
Cambridge University Press. Cambridge.
Hirota, K., Pedrycz, W. 1990. Fuzzy computing for data
mining. Proceedings of IEEE, 87:1575-1600.
Holt, C. A., Laury, S. K. 2002. Risk aversion and
incentive effects. American economic review,
92(5):1644-1655.
Hudec, M. 2017. Merging validity and coverage for
measuring quality of data summaries. In Information
Technology and Computational Physics. Springer
International Publishing, Switzerland (in press).
Hudec, M. 2015 Storing and analysing fuzzy data from
surveys by relational databases and fuzzy logic
approaches. In XXV-th IEEE International Conference
on Information, Communication and Automation
Technologies (ICAT 2015). IEEE.
Hudec, M., Sudzina, F. 2012. Construction of fuzzy sets
and applying aggregation operators for fuzzy queries.
In 14th International Conference on Enterprise
Information Systems (ICEIS 2012) SCITEPRESS.
Lee, C. J., Andrade, E. B. 2015. Fear, excitement, and
financial risk-taking. Cognition and Emotion, 29(1):
178-187.
Liu, B. 2011. Uncertain logic for modeling human
language. Journal of Uncertain Systems, 5:3-20.
Lusardi, A. 2008. Financial literacy: an essential tool for
informed consumer choice? (No. w14084). National
Bureau of Economic Research.
Rasmussen, D., Yager, R.R. 1997. Summary SQL - A
fuzzy tool for data mining. Intelligent Data Analysis,
1:49-58.
Tudorie, C. 2008. Qualifying objects in classical relational
database querying. In Handbook of Research on Fuzzy
Information Processing in Databases, pages. 218-245.
Information Science Reference, Hershey.
Tversky, A., Kahneman, D. 1975. Judgment under
uncertainty: Heuristics and biases. In Utility,
probability, and human decision making, pages. 141-
162. Springer, Netherlands.
Viertl, R. 2011. Fuzzy data and information systems. In
15th WSEAS International Conference on Systems.
WSEAS Press.
Vučetić, M., Vujošević, M. 2012. A literature overview of
functional dependencies in fuzzy relational database
models. Techniques and Technologies for Education
and Management, 7(4):1593-1604.
Wu, D., Mendel, J.M., Joo, J. 2010. Linguistic
summarization using if-then rules. In 2010 IEEE
International Conference on Fuzzy Systems. IEEE.
Yager, R. R. 1982. A new approach to the summarization
of data. Information Sciences, 28:69–86.
Yager, R.R., Ford, M., Cañas, A.J. 1990. An approach to
the linguistic summarization of data. In 3rd
International Conference on Information Processing
and Management of Uncertainty in Knowledge-based
Systems (IPMU ’90). EUSFLAT.
Zaleskiewicz, T., Piskorz, Z., Borkowska, A. 2002. Fear
or money? Decisions on insuring oneself against
flood. Risk, Decision and Policy, 7(3):221-233.
DATA 2017 - 6th International Conference on Data Science, Technology and Applications
254