Advanced Analytics in Central Banks:

Basic Assumptions and Preliminary Results of a Research Project

Matthias Goeken and Leni Hirdes

University of Applied Sciences of Deutsche Bundesbank, Hachenburg, Germany

Keywords: Central Banking, Advanced Analytics, Systematic Review.

Abstract: This position paper presents methodological considerations and ideas as well as preliminary results of a re-

search project. We argue that integrating design science (DS) and review research is useful for developing

artifacts that serve, for example, to share knowledge within a domain. Our project aims to support knowledge

sharing in central banking by means of systematic reviews. Such a consolidated body of knowledge can also

be considered an artifact in the sense of DS research. We demonstrate how such a review can be conducted

and what results it yields. Since it is not clear whether such an application of systematic reviews is relevant

and feasible, problem definition and objectives of the project are discussed.

1 INTRODUCTION

Central banks are knowledge-intensive organizations.

To fulfill their mandate, a critical task is to observe

and monitor markets and the economic environment,

to consider economic policy measures, to make

forecasts and to draw conclusions for potential

actions. To do so, they traditionally rely on statistics

and econometrics in connection with economic model

building. These approaches and methods are well

established and have been successfully applied in the

domain of central banking for many years.

In fact, central banks tend not to contest with other

organizations or with each other. Therefore, unlike

many players in the financial industry, they are not

dependent on gaining competitive advantages e.g., by

means of (technical) innovations. Nevertheless, they

are interested in expanding their analytical competen-

cies in order to be able to best fulfil their mandate.

And, like other players in the financial industry, they

strive to do so by adopting innovative and advanced

technologies and approaches, e.g. big data, machine

learning, AI, and data science (Bholat, 2015).

Seven years ago, a survey of central banks by the

Irving Fisher Committee (IFC, 2017) found that there

was already strong interest in big data, machine

In this paper we use “advanced analytics” as an

umbrella term to refer to big data (BD), machine

learning (ML), data science (DS), and artificial

learning and related topics in the central banking

community. In addition to the general trends fre-

quently noted for the economy as a whole (data avai-

lability, availability of superior computational power,

and more mature methods and procedures) and the

related general desire to improve analytics capabili-

ties in traditional areas, there are specific drivers that

are pertinent to the growing interest of central banks.

E.g., the Great Financial Crisis (GFC) of 2007-09 laid

bare the necessity of more disaggregated data (Doerr

et al. 2021). In this respect, macroprudential policy

has been established as a new field of activity

alongside monetary policy and banking supervision.

This is where the improved use of granular data on

individual economies for financial stability analysis

is a key task (Buch, 2019). Figure 1 illustrates the

growing interest among central banks in recent years.

It displays that the number of central bank speeches

mentioning “advanced analytics” and similar related

topics increased considerably from 2015 to 2020.

However, actual use by central banks has so far

remained limited. The IFC Annual Report 2017

identifies three reasons for this: first, a number of

operational challenges, in particular the availability of

significant (technical and human) resources to deal

with big data. Second, making adequate arrangements

for the management of data and information for

intelligence (AI). As the major parts of the paper are

about these approaches in general, a clear distinction is

not necessary in most parts of our paper.

Goeken, M. and Hirdes, L.

Advanced Analytics in Central Banks: Basic Assumptions and Preliminary Results of a Research Project.

DOI: 10.5220/0011306800003280

In Proceedings of the 19th International Conference on Smart Business Technologies (ICSBT 2022), pages 137-143

ISBN: 978-989-758-587-6; ISSN: 2184-772X

137

central banks; and third, that the use of big data for

policy purposes is not without risks, such as creating

a false sense of security and accuracy among the

audience or the public (IFC, 2017).

Figure 1: Central banks’ interest in big data (Doerr et al.,

2021).

These and other problem areas open room for

improvement. One sensible idea is to improve

knowledge sharing and avoid reinventing the wheel

multiple times. Hence, our approach is to summarize

and evaluate certain works, where experiences from

projects, case studies and applications are already

available. In this way, exchange of experience could

be strengthened – beyond the usual formats of central

bank cooperation.

For the purpose of consolidating existing and

published results, so-called systematic reviews have

become increasingly accepted as effective instru-

ments in the information systems discipline in recent

years (Webster & Watson, 2002; Wainwright et al.,

2018). Reviews are seen as powerful instruments to

enable the synthesis of prior findings, reconciliation

of inconsistent findings, and resolution of

relationships. In this realm, reference is also made to

the practical relevance of reviews.

For many reviews, the motivation is already given

by the fact that primary research exists. Here we are

faced with a domain for which it is not entirely

obvious whether a review that consolidates the body

of knowledge is meaningful, relevant, and feasible.

For this reason, the combination of review research

and design science seems appropriate.

In this paper, we first outline the rational and the

motivation for (part 2) and the methodological

background of our research project (part 3). Part 4

outlines first and priliminary results. We finish with

conclusions and and give hints for future research.

2 RELATED WORK

AI, Data Science and related topics have already been

the subject of review articles. Collins et al. (2021) ex-

amine the extent to which primary studies identify

business value of IT through AI and specific contri-

butions made by AI, and they consider the scientific

and practical implications of using AI in organiza-

tions. Abdel-Karim et al. (2021) investigate in their

review, why the “transfer of theoretical knowledge

from machine learning to applications that solve (in-

dustry-)specific problems” is only partly successful.

They do so with the goal of advancing IS theory as

well as to help practitioners in applying ML methods.

Other reviews explore the practical understanding of

AI to better identify areas of application and improve

adoption (Bawack et al. 2019).

Likewise, papers appear that investigate the “state

of the art” with a focus on methodological issues of

advanced analytics (Angra & Ahuja, 2017). Some

high-level reviews study the application of ML and

AI in general by (Warin & Stojkov, 2021).

Occasionally, but more rarely, there seem to be works

that examine the application of advanced analytics in

relation to a specific technical topic or domain. Our –

so far non-systematic – literature review e.g.

identified reviews on the application of ML and AI in

supply chain management or in the context of

manufacturing (Fahle et al. 2020; Younis et al.,

2021). However, such works are rather rare. It seems

that the focus on a domain or specific cases is much

more common in other disciplines, e.g. in medicine,

probably, because the benefits are more obvious.

We go with Abdel-Karim et al. (2021), who note

a low consideration of ML in the IS discipline.

However, like them, we assume that after remarkable

progress has been made in Computer Science from a

technical and methodological point of view, it is a

fundamental task of IS research to investigate the

transfer into companies and organizations and to

support it with appropriate methods and artifacts.

Hence, we focus on the transfer of theoretical

knowledge regarding advanced analytics to

applications. Accordingly, the focus is on sharing

knowledge for a practical purpose and action, which

is not the primary goal of existing reviews.

Since we can identify a concrete and relevant use

case in the central banking domain (see above), our

research project aims at the idea of conducting a

review of the application of advanced analytics for

this rather narrowly defined field. Last year,

Kinywamaghana & Steffen (2021) highlighted

examples of the application of the aforementioned

technologies and methods. To our knowledge,

however, there is currently no review that examines

the literature for the field of central banking

systematically and with a practical stance.

ICSBT 2022 - 19th International Conference on Smart Business Technologies

138

Phase Outcome

0. Building the Infrastructure Framework to represent the subject matter (taxonomy, ontology); classification system

1. Defining the research question Research Question

2. Searching the literature Preliminary inclusion of studies based on database research

3. Selecting the studies for inclusion Set of final eligible studies

4. Assessing the quality of included studies

and structuring of their results

Assessed and structured studies

5. Combining the results Representation of the gained and integrated results, e.g., in a 'Summary of findings'

table

6. Create a structured report Report on the findings and the evidence gained

Figure 2: Process for a Systematic Review (based on Cooper 2016; Goeken 2011).

Figure 3: DSR methodology process model (Peffers et al. (2007).

3 METHODOLOGICAL

BACKGROUND

Our research endeavour can be sensibly classified in

two ways. First, it is review research, since it ad-

dresses the stated goals (above all: synthesizing the

existing knowledge; representation of the “state of the

art” at a given point in time). Accordingly, it is based

on the now established and widely accepted process

for reviews, which is illustrated in figure 2 – phase 1

to 6. On the other hand, it is also design-oriented in

parts, since the result is intended to be an artifact that

provides support for practitioners (presenting the

“body of knowledge” for practical application and

practical usefulness). In this respect, it should also be

guided by approaches from Design Science. In this

field, the DSRM by Peffers et al. (2007) is considered

the “mostly widely referenced model” guiding the

process (Vom Brocke et al., 2020) – see figure 3.

Thus, we envisage to combine both approaches to

have the two sides of the coin in view – consolidation

of knowledge and representation of knowledge for

practical application. Due to page restrictions and

because we are presenting research in progress here,

emphasis is put in the following not on those aspects

that are already well discussed (e.g. the search

strategy) but on methodological considerations and

preliminary results of ours that are already somehow

advanced and that could lead to a fruitful discussion.

In short and certainly somewhat simplified terms,

we state that, while reviews are primarily required to

be transparent and reproducible in order to yield

purposeful and reliable results, the development of

artifacts focuses on making justified design decisions,

but also on ensuring that the final result is useful and

easy to use. Accordingly, the procedure for reviews

tends to be rather linear, while DSRM provides for

explicit feedback loops that not only allow but also

stipulate the return to previous phases, so that

continuous improvement is possible in an iterative

process. If a result is available in phase 3 (DSRM),

the process does not end, but the artifact is evaluated

after its demonstration. I.e. it is also adapted, changed

etc. after (initial) completion if necessary. Although

Advanced Analytics in Central Banks: Basic Assumptions and Preliminary Results of a Research Project

139

this is not explicitly discussed in the literature, a step-

by-step extension of the artifact might be introduced

to improve usefulness and maintain applicability of a

given artifact.

Although the definition of a research question is

relatively close to the first two phases of the DSRM

(problem motivation and definition of objectives of a

solution), a phase is added upstream here (phase 0):

Due to the fact that – compared with other disciplines,

e.g. medicine – IS lacks established nomenclatures

and accepted taxonomies of the subject matters, this

phase initially deals with the construction of a

framework (or multiple frameworks) for the research

topic at hand. The goal of this framework is to better

conceptualize the research question and potentially

decompose it in such detail as to allow for an

evolutionary approach. Due to the different nature of

this task and its importance, this is treated as a

separate phase. This construction is a design-oriented

task which itself could be based on design science

methodology accordingly (Peffers et al. 2007).

Intention of this phase is also to conceptually

atomize the research question so that the searching for

relevant literature as well as its structuring and

presentation can be supported. It also becomes

meaningful with a view to possible iterations. To

foresee multiple iterations makes sense against the

background of practicability (even in the quite

delimited domain of central banking, there are many

papers that may not be manageable due to their sheer

volume) and because in this way subject matter

experts for the different topics can be involved

(money laundering is just completely different from

financial stability).

4 SELECTED PRELIMINARY

RESULTS

In our approach, phases 0 and 1 were carried out iter-

atively. Problem identification and definition of ob-

jectives (DSRM) flow into this and lead to a high-

level research question, which is shaped with the cre-

ation of a framework (conceptual model or taxon-

omy). One focus of the research project is the moti-

vation (feasibility as well as the potentials of a collec-

tion of case studies and use cases on the application

of advanced analytics in central banks). Since central

banks traditionally have access to a significant

amount of data and statistics that are regularly col-

lected while fulfilling their mandate and inform their

decision-making processes, the project will explore

whether and to what extent such a collection would

be useful for central banks and whether it could help

to improve the decision making.

It might help central banks in their knowledge

sharing to gain a better understanding of what others

are researching or what is being done in general in the

topic area of advanced analytics. In classifying re-

search findings, a unified knowledge base could help

identify new datasets, associated methodologies, re-

lating projects, and assist central banks with analyses

and to better understand them. Thus, for example, the

assessment of the impact of policies could take place

more accurately (Tissot, 2014). At the same time, a

case study collection can serve as an overview so that

exchanges and collaborations among central banks

can be facilitated. In addition, Tissot (2014) mentions

another advantage: the added value compared to “tra-

ditional” statistics when central banks run respective

innovative projects – e.g., nowcasting and the use of

advanced analytics. In our research, beyond the liter-

ature positions, we conducted in-depth interviews that

elaborated potential benefits of a collection and we

identified relevant value dimensions. These will be

applied as objectives in a future evaluation.

Hence, we argue, that our research idea addresses

a relevant problem in the real-world environment.

From the description of the idea, we derive a high-

level research question that is at the core of the sys-

tematic review: Where in central banks (in which ap-

plication areas and departments) are advanced ana-

lytic methods (and which ones) applied to which spe-

cific issues?

The structuring and decomposition of the research

question is based on the literature, software libraries

(like Scikit-Learn and Tensor Flow), organizational

charts, library taxonomies, and interviews. Figure 4

shows an excerpt of the structuring of typical task ar-

eas of a central bank on the left. On the right, methods

and techniques of AI and ML are outlined (also in ex-

cerpts).

These frameworks allow a detailing of the re-

search question to work on it in a decentralized way

(in subprojects and by respective experts). Research

questions can now be derived from the two frame-

works by combining domain-specific areas and meth-

ods. For example, monetary policy in combination

with one or more regression methods. If necessary, it

may be appropriate to limit this further, for example,

to analyses in this area that refer to the real estate mar-

ket and examine related effects on price stability (if

studies are available at this level of detail). In either

case, the detailed research questions ultimately re-

main anchored in the high-level research question,

which allows the results to be combined at that level

(or an intermediate level).

ICSBT 2022 - 19th International Conference on Smart Business Technologies

140

Figure 4: Frameworks for decomposition of the research question.

Along the procedure shown in Figure 2, sub-re-

search questions have already been addressed in each

of several subprojects. The recommended databases

(e.g., Vom Brocke et al. 2009) were searched, alt-

hough manual searches were also carried out on spe-

cial databases and websites, for example those of the

Bank for International Settlements (BIS), due to the

special domain. After selecting and assessing the

quality of studies, a manageable number were se-

lected for deeper analysis.

Combining the results was done in “Summary of

Findings tables” (SoF). Other possibilities, e.g., sta-

tistical evaluation and combination in a meta-analy-

sis, were not considered, since mainly qualitative re-

sults are reported in the respective studies. It became

apparent that there is no “one fits all” format for the

SoF tables, i.e., that researchers have made modifica-

tions in each case, so that a consolidation of the vari-

ous tables in terms of the research question and the

research objective is a further task in the overall con-

text. The table in the appendix shows an excerpt of a

SoF table, here choosing from the framework “mone-

tary policy” on one side and “ML/AI” on the other.

The table also shows that a clear assignment to a task

area is not always possible. In various places, for ex-

ample, “monetary policy” is combined with other ar-

eas of responsibility (e.g., central bank communica-

tion).

Already with the results available, which in a way

could represent the final outcome of a Systematic Re-

view, it is now possible to switch again to the proce-

dure according to the DSRM. The SoF tables are pro-

totypes and thus, as artifacts, results of phase 3. These

can be used to carry out a demonstration and, against

the background of the objectives, an evaluation. For

the demonstration, this publication is a first step. It is

planned – in addition to an evaluation against the

background of the objectives – to involve possible ad-

dressees and stakeholders from central banks in order

to obtain their assessment of the usefulness, ease of

use and applicability. As is known from system de-

velopment, a prototype can be used as a stimulus to

make verification and validation as realistic as possi-

ble. In this respect, the reviews as partial results also

fulfil an essential purpose within the framework of

the design-oriented approach.

5 CONCLUSION AND FURTHER

RESEARCH

In this paper, a research project has been presented

that aims to demonstrate and investigate how and

whether systematic reviews can contribute to a better

sharing and exchange of knowledge. This is investi-

gated for the specific case of the central banking do-

main.

CentralBank

Tasks/Areas

Paymentsystems

Banking

supervision

Monetarypolicy

Financialstability

Cash

management

Centralbank

communication

ML/AI

Supervised

Learning

Regression

Linear

Polynomial

…

Classification

Decision Tree

SVM

…

Unsupvervised

learning

Clustering

…

Dimension

Reduction

…

Reinforcement

learning

…

Advanced Analytics in Central Banks: Basic Assumptions and Preliminary Results of a Research Project

141

The context is to identify and summarize

knowledge about the application of advanced analyt-

ics in central banks. A presentation of this knowledge

in the sense of a consolidation of the “body of

knowledge” can provide an overview of existing

work, stimulate own projects, and promote exchange

and cooperation among data scientists and developers

in central banks.

Our work also consists in a (methodological)

combination of DS and review research, as the result

should not only bring a progress of knowledge and

insight but also a very practical progress in the sense

of a case study database that should promote the ex-

change of interested organizations and participants.

In this position paper, only the basic idea has been

presented due to the fact that the methodological ap-

proach has not yet been fully defined and established,

and also in the sense that only preliminary results are

available so far. Reflections on the research process

have been presented above and it has already been

gone through several times. It is founded on estab-

lished work on review research and established ap-

proaches to design science. Against the background

of the special question pursued, a combination is

aimed at, which currently cannot yet be regarded as

well-founded in every respect. Further methodologi-

cal work is needed here. It is possible that the combi-

nation of design science and review research may

prove fruitful beyond the domain under considera-

tion.

Furthermore, beyond the still small-scale “proto-

types”, which each covers quite delimited thematic

areas, larger-scale reviews of studies and use cases

need to be carried out, so that not only the feasibility

is demonstrated, but also a concrete practical benefit.

REFERENCES

Abdel-Karim, B. M., Pfeuffer, N., Hinz, O. (2021). Ma-

chine learning in information systems - a bibliographic

review and open research issues. In Electronic Markets,

31(3), 643-670.

Angra, S., & Ahuja, S. (2017). Machine learning and its

applications: A review. In 2017 International Confer-

ence on Big Data Analytics and Computational Intelli-

gence (ICBDAC) (pp. 57-60). IEEE.

Bawack, R. E., Fosso Wamba, S., & Carillo, K. (2019).

Artificial intelligence in practice: Implications for IS

research. In Proceedings of Twenty-fifth Americas

Conference on Information Systems, Cancun, 2019

Bholat, D. (2015). Big data and central banks. Big Data &

Society, 2(1), 2053951715579469.

Buch, C. (2019). Welcoming remarks. International Semi

nar on Big Data, Building Pathways for Policy-Making

with Big Data, Bali, 26 July 2018.

Collins, C., Dennehy, D., Conboy, K., & Mikalef, P.

(2021). Artificial intelligence in information systems

research: A systematic literature review and research

agenda. In International Journal of Information Man-

agement, 60

Cooper, H. (2016). Research synthesis and meta-analysis:

A step-by-step approach (Vol. 2). Sage publications.

Doerr, S., Leonardo Gambacorta, Jose Maria Serena

(2021). Big data and machine learning in central bank-

ing. BIS Working Papers 2021,

https://www.bis.org/publ/work930.pdf

Fahle, S., Prinz, C., & Kuhlenkötter, B. (2020). Systematic

review on machine learning (ML) methods for manu-

facturing processes – Identifying artificial intelligence

(AI) methods for field application. In Procedia CIRP,

93, 413-418.

Goeken, Matthias (2011). Towards an Evidence-based Re-

search Approach in Information Systems. ICIS 2011

Proceedings. 10.

Irving Fisher Committee on Central Bank Statistics. (2015).

Central banks' use of and interest in "big data". IFC

Report.

Irving Fisher Committee on Central Bank Statistics. (2018).

IFC Annual Report 2017.

Kinywamaghana, A., and S. Steffen (2021). A Note on the

Use of Machine Learning in Central Banking, FIRE

Research Paper.

Peffers, K., Tuunanen, T., Rothenberger, M. A., & Chat-

terjee, S. (2007). A design science research methodol-

ogy for information systems research. In JMIS, 24(3),

45-77.

Tissot, B. (2018). Big data for central banks. Bali: Bank for

International Settlements.

Vom Brocke, J., Hevner, A., & Maedche, A. (2020). Intro-

duction to design science research. In Design Science

Research. Cases (pp. 1-13). Springer, Cham.

Vom Brocke, J., Simons, A., Riemer, K., Niehaves, B.,

Plattfaut, R., & Cleven, A. (2009). Standing on the

shoulders of giants: Challenges and recommendations

of literature search in information systems research. In

CAIS, 37(1), 9.

Wainwright, D., Oates, B., Edwards, H., & Childs, S.

(2018). Evidence-based information systems: A new

perspective and a roadmap for research informed prac-

tice. In JAIS 19(11), 4.

Warin, T., & Stojkov, A. (2021). Machine Learning in Fi-

nance: A Metadata-Based Systematic Review of the

Literature. In Journal of Risk and Financial Manage-

ment, 14(7), 302.

Webster, J., & Watson, R. T. (2002). Analyzing the past to

prepare for the future: Writing a literature review. In

MIS quarterly, xiii-xxiii.

Younis, H., Sundarakani, B., & Alsharairi, M. (2021). Ap-

plications of artificial intelligence and machine learning

within supply chains: systematic review and future re-

search directions. In Journal of Modelling in Manage-

ment. Vol. ahead-of-print

ICSBT 2022 - 19th International Conference on Smart Business Technologies

142

APPENDIX

Sub-results showing combination “monetary policy”

and “ML/AI” (all procedures)

Title Authors Year Central Bank Task Areas Methods and Techniques

of AI and ML

Research Question Outline / Results Research Setting

Between hawks and

doves: Measuring

central bank

communication

Tobback, Ellen;

Nardelli,

Stefano David

Martens

2017 ECB Monetary policy, Central

bank communication

Text Mining, Support-

Vector-Machines

(SVM), Latent Dirichlet

Allocation (LDA)

none explicitly

mentioned

Hawkish-Dovish (HD) indicator, which

measures the degree of "hawkishness" or

"dovishness" of the media's perception of

ECB's tone at each press conference.

Case Study

Central Bank

Communications and

the General Public

Haldane,

Andrew;

McMahon,

Michael

2018 Reserve Bank of

Australia

Monetary policy; Central

bank communication

Lineare Regression,

SVM & Random Forest

What does a high-

quality ZBK look like?

Perception of different

addressees.

There is no consensus on what high-

quality central bank communication should

look like. To shed light on this, 3 important

aspects are examined.

Case Study

Text mining for central

banks

Bholat, David;

Stephen,

Hansen;

Santos, Pedro;

Schonhardt-

Bailey, Cheryl

2015 Bank of England Monetary policy; Central

bank communication;

Financial stability

Text Mining, Supervised

Learning: Boolean

techniques, Dict.

techniques, weighting of

words, vector space

models, LDA

How can text mining

be useful for

addressing research

topics of interest to

central banks?

Discussion & step-by-step guide to text

mining, including an overview of

unsupervised and supervised techniques.

Report

The Evolving Scope

and Content of Central

Bank Speeches

Siklos, Pierre

L.; St. Amand,

Samantha;

Wajda, Joanna

2018 US Federal

Reserve and the

Bank of Canada

Monetary policy, Central

bank communication

NLP, LDA How have the use of

central bankers'

speeches and the

content of those

speeches changed

over time?

Assessment of Change.

Comparison between the use of speeches

by the Federal Reserve and the Bank of

Canada

Case Study

Topic classification of

Monetary Policy

Minutes from the

Swedish Central Bank

Cedervall,

Andreas;

Jansson,

Daniel

2018 Central Bank of

Sweden

Monetary policy, Central

bank communication

Neural Network, LDA none explicitly

mentioned

Analysis of Swedish Central Bank

protocols and collection of information

using latent Dirichlet allocation and a

simple neural network.

Case Study

Advanced Analytics in Central Banks: Basic Assumptions and Preliminary Results of a Research Project

143