Estimating the relation between digitalization and the market value of insurers

• Published date: 01 September 2021
• Author: Simon Fritzsch,Philipp Scharner,Gregor Weiß
• Date: 01 September 2021
• DOI: http://doi.org/10.1111/jori.12346

J Risk Insur. 2021;88:529–567. wileyonlinelibrary.com/journal/JORI

|

529

Received: 15 September 2020

|

Revised: 18 April 2021

|

Accepted: 22 April 2021

DOI: 10.1111/jori.12346

ORIGINAL ARTICLE

Estimating the relation between digitalization

and the market value of insurers

Simon Fritzsch |Philipp Scharner |Gregor Weiß

Faculty of Economics and Management,

Leipzig University, Leipzig, Germany

Correspondence

Gregor Weiß, Faculty of Economics and

Management, Leipzig University,

Grimmaische Str. 12, Leipzig 04109,

Germany.

Email: weiss@wifa.uni-leipzig.de

Funding information

LBBW Asset Management

Investmentgesellschaft mbH,

Grant/Award Number: PhD scholarship

Abstract

We analyze the relation between digitalization and the

market value of US insurance companies. To create a

text‐based measure that captures the extent to which

insurers digitalize, we apply an unsupervised machine

learning algorithm—Latent Dirichlet Allocation—to

their annual reports. We show that an increase in di-

gitalization is associated with an increase in market

valuations in the insurance sector. In detail, capital

market participants seem to reward digitalization ef-

forts of an insurer in the form of higher absolute

market capitalizations and market‐to‐book ratios. Ad-

ditionally, we provide evidence that the positive rela-

tion between digitalization and market valuations is

robust to sentiment in the annual reports and the

choice of the reference document on digitalization,

both being issues of particular importance in text‐based

analyses.

KEYWORDS

digitalization, Latent Dirichlet Allocation, machine learning,

market valuation

JEL CLASSIFICATION

C33, C61, G22

This is an open access article under the terms of the Creative Commons Attribution‐NonCommercial‐NoDerivs License, which permits

use and distribution in any medium, provided the original work is properly cited, the use is non‐commercial and no modifications or

adaptations are made.

© 2021 The Authors. Journal of Risk and Insurance published by Wiley Periodicals LLC on behalf of American Risk and Insurance

Association.

1|INTRODUCTION

Digitalization has already massively transformed many industries. The insurance industry,

however, has yet to take advantage of the full potential of digital technologies. This becomes

even more important as rising customer expectations, the effects of the financial markets crisis,

and the zero interest rate policy lead to an increased competitive pressure. In general, there is

no doubt on the strong impact digitalization will have on the insurance ecosystem (see, e.g.,

Cappiello, 2020). It is considered to affect the whole insurance value chain, from product

development to pricing/underwriting, sales and distribution, policy and claims management,

and asset and risk management (Eling & Lehmann, 2018). However, in contrast to other

megatrends such as urbanization or aging societies, the precise scope of digitalization is dif-

ficult to grasp. Although we know that digitalization clearly manifests itself in cloud com-

puting, Internet of Things, mobile communication, blockchain technology, artificial

intelligence, and so forth (Schmidt, 2018), evidence on the question of how to measure digi-

talization and its relation to firm outcomes is still scarce (see, e.g., Bohnert et al., 2019; Hanelt

et al., 2020; Scott et al., 2017).

In this paper, we fill this gap by proposing a new method to measure digitalization in the

insurance sector. Our method exploits the prevalence of different topics in standard annual

reports. Based on the assumption that digitally innovative insurers report their progress more

extensively, Latent Dirichlet Allocation (LDA) helps to assess the importance of digitalization

for the particular insurance company. At the same time, the method enables us to separate

digitalization from mere firm innovation. In a second step, we use our text‐based measure on

digitalization to investigate its relation with the market valuation of a large set of publicly‐listed

US insurance companies. Finally, we account for potential confounding issues related to the

construction of the digitalization measure, the reference document used for LDA, and the

sentiment in which annual reports are written.

Our results provide first evidence for a positive association between digitalization efforts

and market valuation in the US insurance sector. We find that an increase in digitalization is

strongly related to an increase in market value and market‐to‐book value of US insurance

companies. Put differently, market participants associate a more digitalized insurance company

with higher future profitability and consequently a higher firm value. Although LDA is by

design subject to some discretion, we show that our results are robust to several variations of

our model parameters. Most importantly, our findings are robust to different numbers of topics

used to structure the annual reports and to isolate digitalization from general innovation.

Furthermore, the results do not depend on the discretionary choice of the reference document

and are not confounded by annual reports' sentiment.

The topic model LDA by Blei et al. (2003) has only recently been introduced to the finance

literature.

1

In general, topic models can be used to analyze large data sets of texts that are often

unstructured (Roberts et al., 2016). These probabilistic models provide a finite set of common

topics which optimally reflect a collection of documents. By applying a topic model to a specific

document, we obtain a vector of topic loadings representing how intensively each topic is

discussed in the respective document. One of the main advantages of LDA over simple word‐

list approaches is that the topics and corresponding word distributions arise endogenously from

the data and do not have to be specified by the researcher. That is, the underlying machine

1

One of the first applications of this approach in accounting and finance is due to Huang et al. (2018), who study topical differences between conference calls

and subsequent analyst reports.

530

|

FRITZSCH ET AL.

learning algorithm determines the terms that are most important to discriminate between

documents and topics in an unsupervised fashion.

We then apply this powerful tool to the annual reports of 86 publicly listed US insurance

companies available in Thomson Reuters Datastream from 2006 to 2015 and derive a dis-

tribution of topics for each of the annual reports. This yields a low‐dimensional representation

of the document (cf. Blei et al., 2003) that we exploit to construct our text‐based measure of

digitalization. For this purpose, we compare the extent to which each topic is discussed in the

respective report to a reference document about digitalization in the insurance sector. Speci-

fically, we use the paper by Bohnert et al. (2019) since it is closely related to our work. To the

best of our knowledge, it is one of the few studies trying to establish an empirical relation

between the expression of a digital agenda and the market valuation of insurance companies.

2

More exactly, based on these topic distributions we calculate a measure of similarity between

the digitalization document and the insurers' annual reports using the Kullback–Leibler (KL)

divergence (Kullback & Leibler, 1951). This measure is then used to proxy for the extent of

digitalization in our sample insurance companies.

Our paper is related to a growing body of literature on textual analysis and machine

learning in finance. Starting with Frazier et al. (1984), Antweiler and Frank (2004), and Tetlock

(2007), researchers have studied the effect of qualitative information on equity valuations. More

recent papers (e.g., Hanley & Hoberg, 2010; Hoberg & Maksimovic, 2015; Hoberg et al., 2014;

Jegadeesh & Wu, 2013; Ke et al., 2017,2019) conduct text‐based analyses to examine a wide

variety of finance research questions.

3

Intriguingly, within the field of text analysis and ma-

chine learning, LDA is becoming more and more popular (see, e.g., Ganglmair & Wardlaw,

2017; Goldsmith‐Pinkham et al., 2016; Hoberg & Lewis, 2017; Huang et al., 2018; Lopez‐Lira,

2019). Within this growing strand of literature, our paper is most closely related to Bellstam

et al. (2020) and Lowry et al. (2019), who derive topics based on LDA and employ the KL

divergence as a measure of similarity between probability distributions.

At the same time, our paper is related to a growing body of literature on the effect of

digitalization in the insurance sector. This literature is basically centered around the impact of

digitalization on the business model of insurers (see, e.g., Cappiello, 2020; Desyllas & Sako,

2013), new forms of online marketing and sales activities (Seitz, 2017), and the overall trans-

formation of insurance companies (Barkur et al., 2007). The various facets of digitalization such

as Big Data, artificial intelligence, predictive modeling, telematics, and Internet of Things are

considered to have a tremendous impact on the whole insurance value chain. In detail, product

design and development, underwriting/pricing, sales and distributions, as well as policy and

claims management are all subject to fundamental change in the future (see, e.g., Cappiello,

2020; Meier & Stormer, 2012; Rayport & Sviokla, 1995; van Rossum et al., 2002). Numerous

opportunities like a facilitated interaction with customers via mail, chatbots, and social media

or cost reduction via automation and standardization of business processes are challenged by

few risks like the depersonalization of the insurer‐customer relationship (Cappiello, 2020).

However, there is only little empirical evidence on the relation between digitalization and the

market valuation of insurance companies.

4

Our work contributes to the current literature on

digitalization and firm valuation in several ways. First, we propose a novel approach to quantify

2

However, we also consider further reference documents in the robustness checks, for example, Cappiello (2020) and Nicoletti (2016) as well as Bohnert et al.

(2019) with the empirical study being removed.

3

An excellent review of this literature can be found in Lowry et al. (2016).

4

One of the few empirical studies in this field, as mentioned above, is Bohnert et al. (2019).

FRITZSCH ET AL.

|

531