Multiples for Valuation Estimates of Life Science Companies in Sweden

INOM

EXAMENSARBETE TEKNIK, GRUNDNIVÅ, 15 HP

STOCKHOLM SVERIGE 2019 ,

Multiples for Valuation Estimates of Life Science Companies in

Sweden

HAMPUS ERNSTSSON MAX BÖRJES LILJESVAN

KTH

SKOLAN FÖR TEKNIKVETENSKAP

Multiples for Valuation Estimates of Life Science Companies in

Sweden

HAMPUS ERNSTSSON

MAX BÖRJES LILJESVAN ROYAL

Degree Projects in Applied Mathematics and Industrial Economics (15 hp) Degree Programme in Industrial Engineering and Management (300 hp) KTH Royal Institute of Technology year 2019

Supervisors at KTH: Jörgen Säve-Söderbergh

Examiner at KTH: Jörgen Säve-Söderbergh

TRITA-SCI-GRU 2019:156 MAT-K 2019:12

Royal Institute of Technology

SE-100 44 Stockholm, Sweden

URL: www.kth.se/sci

Abstract

Market multiples are a common and simple tool for estimation of corporate value. It can express temporal dynamics and differences in markets, industries and firms. Despite their practical usefulness, some critical problems remains which continue to be debated. This thesis investigates if there exists character- istics for explaining market capitalization by market multiples within the life science industry in Sweden. The approach follows well known theory of multi- ple linear regression analysis. The results indicated only a linear relationship between the market cap and the R&D expenditures of a company. This does not mean that the other explanatory variables does not have effect on market cap only that there is no linear relationship that could be statistically proven.

Keywords: Market capitalization, valuation multiples, multiple linear regres-

sion.

Sammanfattning

V¨ arderingsmultiplar ¨ ar ett vanligt och enkelt verktyg f¨ or att approximera f¨ ore- tags v¨ arde. Det kan beskriva tempor¨ ar dynamik och skillnader hos marknader, industrier och bolag. Trotts dess praktiska anv¨ andbarhet finns en del kritiska problem som fortfarande debateras. Denna uppsats unders¨ oker om det existerar n˚ agra egenskaper f¨ or att f¨ orklara marknadsv¨ ardet med hj¨ alp av v¨ arderingsmulti- plar inom life science industrin i Sverige. Tillv¨ agag˚ angss¨ attet f¨ oljer v¨ alk¨ and teori om multipel linj¨ ar regressions analys. Resultaten visade att det endast finns ett samband mellan marknadsv¨ ardet och utgifter f¨ or forksning och utveckling f¨ or ett bolag. Detta inneb¨ ar inte att de andra variablerna inte har n˚ agon effekt p˚ a marknadsv¨ ardet, utan att det inte finns ett linj¨ art samband som kan bevisas p˚ a ett statistiskt vis.

Nyckelord: Marknadsv¨ arde, v¨ arderingsmultiplar, multipel linj¨ ar regression.

Acknowledgements

We would like to thank Per J¨ orgen S¨ ave-S¨ oderberg for great support and feed-

back throughout the process. We would also like to thank Infront for using their

system.

Contents

1 Introduction 6

1.1 Problem Statement and Research Question . . . . 7

1.2 Scope . . . . 7

1.3 Disposition . . . . 7

2 Background Theory 8 2.1 Response Variable and Market Multiples . . . . 8

2.1.1 Market Capitalization . . . . 8

2.1.2 Price to Earnings Ratio (P/E) . . . . 8

2.1.3 Price to Book Value (P/B) . . . . 8

2.1.4 Return on Equity (ROE) . . . . 9

2.1.5 Debt to Equity (D/E) . . . . 9

2.1.6 Research and Development Expenditure (R&D) . . . . 9

2.2 Market Psychology . . . . 10

2.3 Mathematical Theory . . . . 11

2.3.1 Multiple Linear Regression . . . . 11

2.3.2 Model Adequacy . . . . 12

2.3.3 Model Properties . . . . 14

2.3.4 Variable Selection . . . . 15

3 Methodology 16 4 Results 17 4.1 Model Adequacy . . . . 17

4.1.1 Leverage and Influence . . . . 19

4.1.2 Added Variable Plot . . . . 20

4.2 Multicollinearity . . . . 21

4.2.1 Correlation Matrix . . . . 21

4.2.2 Variance Inflation Factor (VIF) . . . . 22

4.3 Variable Selection . . . . 22

4.3.1 All Possible Regression . . . . 22

5 Discussion 24 5.1 Model Adequacy . . . . 24

5.2 Leverage and Influence . . . . 24

5.3 Multicollinearity . . . . 25

5.4 Variable Selection . . . . 25

5.5 Market Psychology . . . . 25

5.6 Limitations . . . . 26

6 Conclusion and Future Research 27

References 28

Appendix 29

5

1 Introduction

The life science industry can be summarized containing the following three sec- tors: pharmaceuticals, biotechnology and medical devices. Within the industry, companies often follow phase studies for developing a product. The phase stud- ies can be a process that is long and difficult to complete and even if they succeed they also have to complete an approval process by authorities which is a lengthy process. These companies often need funding and chooses therefore to be listed on a stock exchange (Exchange 2010). Depending on the market valuation more or less money can be raised. A recent study showed that when a company makes an announcement of the results from their phase trial an increase in market cap is present when the results are positive and a decrease when the results are negative (Rothenstein et al. 2011). Market valuation can differ significantly between these companies in early stages of development.

According to an article (Crean 2016), many of these companies are develop- ing new products that may generate revenue in the future. Therefore, investors must predict future market potential of the product or products of that com- pany. The potential of the product is often based on key value drives e.g. patient population, competitive landscape, government regulations and market penetra- tion rates. There are several additional challenges in the valuation of companies in early stages of development such as short financial records, rapidly changing conditions, high degrees of uncertainty, high weighted average cost of capital and lack of sufficient funding to achieve business plans within the planned time (Crean 2016). Since there is a low probability of a finished product and long time-to-market, how can differences in market valuation then be explained?

In the financial community market multiples are a common and simplified tool used for asset and company valuation. Corporate finance professionals (investment bankers, appraisers, financial analysts, private equity and leverage buy-out investors) often first use the discounted cash-flow method, and then do a cross-check with a market approach (Harbula 2009). By using financial market data combined to create market multiples for valuation. Depending on which market we are analysis these multiples will show different results affected by external factors (Harbula 2009). Maybe market multiples can be used for an overview of the market capitalization for life science companies in Sweden.

6

1.1 Problem Statement and Research Question

Since the market valuation can differ significantly it is interesting to investigate if there is any trend between these life science companies in Sweden. If such a trend is present it could be used to find undervalued or overvalued companies related to the market. The research questions are therefore,

• What are the characteristics of life science companies in Swe- den?

• Is it possible to construct a model based on these characteristics to find companies that deviate from market valuation in relation to the trend in the market?

The purpose of this paper is to find a relationship between the market value and financial ratios of life science companies in order to find value discrepancies in the market as basis for investment decisions.

1.2 Scope

We limit our study to listed life science companies in Sweden with available data. Real-time financial data was collected 26.03.2019 at 9am from Thomson Reuters Eikon.

1.3 Disposition

The continuation of this study is divided into five sections. Section 2 reviews the background theory including Response and Market Multiples, Market Psy- chology and Mathematical Theory. Section 3 explains the methodology, how the sample was selected and applied, and how the model was evaluated. In Section 4 descriptive results from model adequacy, multicollinearity and variable selec- tion are presented. Section 5 includes detailed analysis of the results and the methodology, and limitations. In Section 6 the summarized results is presented by a conclusion of the thesis followed by suggestions for further research.

7

2 Background Theory

In this section follows a description of selected variables that will be used as mar- ket multiples in the regression analysis, external factors from market psychology and mathematical theory for regression analysis.

2.1 Response Variable and Market Multiples

2.1.1 Market Capitalization

Market capitalization is the market value for all relevant issue level share types and is an important measure of financial development (Billmeier and Massa 2009). It is calculated by multiplying the company’s shares outstanding by the current price of one share.

2.1.2 Price to Earnings Ratio (P/E)

The P/E ratio is used to assign a value to a company based on it’s current share price divided by it’s earnings per share. It is often calculated by using the company information of the recent 12-months period and adjusting for stock splits if there are any. P/E ratios can be associated with current return on equity and often perform poorly when used as future evaluation of growth (Ghaeli 2017). The usefulness of P/E ratios is that they are viewed to capture a firms risk and growth and can therefore be used as a benchmark when comparing firms with similar risk and growth (Cheng and McNamara 2000). The P/E ratio is calculated by dividing the current price of a stock by earnings per share (EPS).

P rice to Earnings (P/E) = M arket V alue per Share

EP S (1)

2.1.3 Price to Book Value (P/B)

Price to book value ratio is calculated by dividing the latest closing price of a stock by its book value per share. Previous studies suggests that the P/B ratio reflects a company’s profitability, risk and growth and at the same time capturing the production efficiency (Cheng and McNamara 2000). Also, other studies have suggested that the P/B ratio can be used as a prediction of future stock returns and that it is in fact the most superior predictor of future valuation (Akademia Baru et al. 2016).

P rice to Book V alue (P/B) = Closing P rice

Book V alue per Share (2) Cheng and McNamara (2000) suggested to combine the P/E and P/BV multi- ples for the most accurate valuation results.

8

2.1.4 Return on Equity (ROE)

Stated in a report (Damodaran 2007), return on capital measures are important for analysis of how well an investment is. The measure return on equity is focusing only on the equity component of the investment. The net income from the current period is depending on the equity investment at the start of the period, it shows how well the capital was used within the company. The return on equity is calculated by dividing the net income from the end of the period with the book value of equity from the start of the period.

Return on Equity (ROE) = N et Income

Book V alue of Equity

(3) Since net income include the interest income from cash and the book value of equity incorporates the cash holding of the firm the ROE is a composite return on all of its assets - cash and operating (Damodaran 2007).

2.1.5 Debt to Equity (D/E)

The capital structure of a firm influences the operating business since it often is affected by different external parties (Dewatripont and Tirole 1994). An example of such parties are shareholders and debt-holders who provide capital to the firm and in doing so have their own opinion on how the firm should perform. Explained in a study (Dewatripont and Tirole 1994), debt-like control often produces more external interference than equity-like control since debt- holders are often called in during rough periods whilst equity-holders are more in control during periods of prosperity. Therefore, debt-holders can be viewed as acting with strict intentions and equity-holders with soft intentions. The debt- to-equity ratio not only captures the capital structure of the firm but also how above parties influence managerial decisions. In the study they also concluded that although managerial decisions not always align with external parties, they are more congruent with equity-holders than with debt-holders (Dewatripont and Tirole 1994). The debt-to-equity ratio is calculated by dividing the total debt as of the end of the fiscal period to total equity for the same period.

Debt to Equity (D/E) = T otal Debt

T otal Equity (4)

2.1.6 Research and Development Expenditure (R&D)

Research and development resulting in new knowledge, new processes and new goods is an important source of technological development (Guellec and De La Potterie 2002). Innovation and R&D has a complex linkage that is non-linear, however it is acknowledged that advances in technology are hard to achieve without intensive work. Therefore, R&D is a good quantitative measurement of how much a company dedicates its resources to hard work towards technological improvements and new innovations. Progress is essential for company growth and R&D is closely related to this and often results in new goods and services (Guellec and De La Potterie 2002).

9

2.2 Market Psychology

This section is a literature review of the book The Psychology of Investing by John R. Nofsinger (2011). It explains some of the external factors related to market psychology affecting the market valuation of companies.

Traditional finance assumes that people act rational which is a fundamental assumption behind many theories and pricing models such as arbitrage theory, portfolio theory and capital asset pricing model. However, people are known to not act rationally at all time. Behavioral finance is a study of how cognitive and emotional biases affect financial markets, financial decisions and corporations.

Prospect theory describes how investors make decisions when there is high un- certainty of the outcome. Depending on how investors chooses to frame their decisions they will be left with different outcomes. Usually investors frame their decisions based on the purchase price as a reference point in order to evaluate potential gains and losses. This however, has a high probability to strongly affect their decision. Behavioral biases comes from different sources that affect investors financial decisions. Self deception produces cognitive errors and is one source of bias where people often overestimate their own capacity, i.e. they are overconfident. Heuristic simplification is another source of bias which describes how people simplify complex analyses due to shortcomings of the human brain which highly affects one’s decision. Another source of bias that affects decision making is mood. An investors mood can have a high influence by overcoming reason and therefore be the root of bad decisions.

Another important aspect of how the market is affected by psychology is herding. When people receive information on what others think about a certain stock they often get influenced. This creates a social consensus resulting in the form of a herd. Not all investors have the skill or interest to analyze new financial information and rely therefore on the consensus of the herd. Moving with the herd creates a large problem since it enlarges psychological biases. In other words, people act on feeling instead of formal analyses. This problem has become more severe during the recent years due to quicker information flows as a result from technological development.

Misattribution bias is another factor that can influence financial decisions. It describes how current feelings and emotions has the ability to influence financial decisions, especially when risk and uncertainty is involved. For example, if an investor is in a good mood he or she is more likely of having a more optimistic view of the future and therefore be prone to investing in risky assets. Many valuation models are based on rational expectations and demand assumptions about the future e.g. growth rate, expected return and variance. In order for these traditional models to create the most accurate output one must use them in a rational and unbiased way. Therefore, when investors use quantitative methods such as present value of growing perpetuity they have to estimate both the discount rate and growth rate. This estimation is influenced by the current mood of the investor and is because of that a biased estimate. For this specific valuation model a biased estimate will produce a severely different outcome since the estimates are both in the denominator. An investor in a good mood

10

may therefore receive an overestimate whilst an investor in a bad mood might produce an underestimate, resulting in different financial decisions.

2.3 Mathematical Theory

2.3.1 Multiple Linear Regression

Multiple linear regression is a regression model that involves more than one regressor variable to explain a possible relationship. In this section necessary theory for fitting and analysing these models is discussed. Mathematical the- ory and statements are based on the book Introduction to Linear Regression Analysis by Montgomery, Peck, and Vining (2012).

2.3.1.1 Model

The model is constructed to explain a relationship between the response variable and the regressor variables. The response y, coefficients of β and regressor variables x = (x

, x

, ..., x

)

constructs the multiple regression model as

y = β

+ β

x

+ β

x

+ ... + β

x

+ ε (5) which describes the model as a hyperplane in a k-dimensional space of the regres- sor variables x

. For further discussion of the model we need some underlying assumptions of the model.

2.3.1.2 Assumptions

The five key assumptions of the model by Montgomery, Peck, and Vining (2012):

1. The relationship between the response and the regressors is linear, at least approximately.

2. The error term ε has zero mean.

3. The error term ε has constant variance σ

. 4. The errors are uncorrelated.

5. The errors are normally distributed.

Taken together, there must be a linear relationship between the response and the regressors which can be evaluated in scatterplots. The residuals are ho- moscedastic, the variance of the error terms are similar across the values of the regressor variables. Assumption 4 and 5 imply that the errors are independent random variables. Multivariate normality, multiple regression assumes that the residuals are normally distributed. As in 5 the errors are normally distributed, which is required for hypothesis testing and interval estimation.

11

2.3.1.3 Ordinary Least Squares

The beta-coefficients and the error variance σ

are unknown. To construct the model the coefficients of β need to be estimated. To visualize the ordinary least squares, from equation (5) the model can be expressed in matrix notations:

y = Xβ + ε where

y =









 y

y

.. . y









 X =











1 x

x

· · · x

1 x

x

· · · x

.. . .. . .. . . . . .. . 1 x

x

· · · x









 β =









 β

β

.. . β









 ε =









 ε

ε

.. . ε











In general, y is an n × 1 vector of the observations, X is an n × p matrix of the levels of regressor variables, β is a p × 1 vector of the regression coefficients, and ε is an n × 1 vector of random errors. The least-square estimator of β is

β = (X ˆ

X)

X

y (6)

The calculated coefficients ˆ β is the optimal for minimizing the residuals sum of squares. The model can be expressed as an equation similar to (5), that is explaining the response variable. However, to conclude something about if the model is true and reasonable than further investigation is needed.

2.3.2 Model Adequacy

The major assumptions presented so far are together the foundation for analysis of the model. It is important to always consider the validity of these assumptions and conduct analyses to examine the adequacy of the model we have tentatively entertained. There is model inadequacies that have potentially serious conse- quences. Gross violations of the assumptions may yield an unstable model and can not be used to describe the variables. This section contains methods for diagnosing violations of the basic assumptions, how it affects the model and how to construct a better model.

2.3.2.1 Residual Analysis The residuals are defined as

e

= y

− ˆ y

, i = 1, 2, ..., n (7) and describe the deviation of the fitted values from the observed values. An important assumption is that they have zero mean and approximate variance

P

i=1

(e

− ¯ e)

n − p = M S

(8)

Using different scaled residuals is useful when trying to detect extreme values or outliers. Studentized residuals is one scaling that is often used since it takes

12

into account the distance in x-space. It is defined as

r

= e

pM S

(1 − h

) (9)

Points with high influence tends to draw the fitted model towards themselves and will therefore have lower residual value compared to points that are close to the centroid. Therefore, by incorporating the distance in x-space, h

which are the diagonals of the hat matrix (13), influential observations will instead have large studentized residuals and thereby easier to detect.

Another useful scaling is the standardized residuals, which are defined as d

= e

√ M S

(10) It is scaled by the approximate variance M S

and are therefore useful when detecting outliers and or extreme values.

2.3.2.2 T-statistics

A t-test can be performed to test a hypothesis for the regression coefficients in order to examine their significance. The hypothesis is then formulated as

H

: β

= 0, H

: β

6= 0 (11) if H

is not rejected then the regressor x

can be removed from the regression.

A t-test is expressed as

t

= β ˆ

p ˆ σ

C

(12)

H

is rejected if |t

| > t

2.3.2.3 Leverage and Influence

Observations with unusual x values, i.e. are remote in the x-space, that lie on the regression fit are called leverage points. They do not affect the coefficient estimates but they do affect model properties e.g. standard errors of the coef- ficients and R

. Influential points however, are both remote in x-space and in y-space and have a large impact on the model since these points attracts the model towards itself. The hat matrix

H = X(X

X)

X

(13)

has an important role when identifying points with leverage. The diagonal elements h

of the hat matrix are a standardized measure of the distance from the i:th observation to the centroid of the x-space. Therefore, observations with large values of h

are potential leverage points. More specifically, points with h

> 2p/n are traditionally considered to be leverage points. In this

13

Cook’s distance measures influence by observing the squared distance be- tween the estimated coefficient without the i:th observation ˆ β

and the full model estimated coefficients. The distance is expressed as

D

= (M , c) = ( ˆ β

− ˆ β)

M ( ˆ β

− ˆ β)

c (14)

where M = X

X, c = pM S

and M S

=

are the usual choice of parameters. Normally, observations with D

> 1 are considered influential.

Covratio is used to measure model performance and expresses how the i:th observation improves the estimation precision. It is defined as

Covratio

=     

(X

X

)

S

(X

X)

M S

(15)

If Covratio

< 1 the i:th observation reduces the precision whereas if Covratio

>

1 it improves the precision. A leverage point will however produce large Covra- tio, therefore a cutoff value is used. Observations where Covratio

< 1 − 3p/n and Covratio

> 1 + 3p/n are to be considered influential.

Added variable plot can be used to understand the impact of each variable on the regressor. This is done by using multiple regression on y while holding all except one explanatory variable constant.(Dayal n.d.)

2.3.3 Model Properties 2.3.3.1 Multicollinearity

If the response variable and regressor variables are centered and scaled to unit length the matrix X

X becomes a correlation matrix with the simple correlation r

on the diagonal. One of the most fundamental assumptions is that the regressor variables are linearly independent, i.e. if there exists a set of constants t

, t

, ..., t

that are all not zero such that

p

X

j=1

t

X

= 0 (16)

Therefore, if some or one value of t

= 0 the rank of (X

X)

is less than p which results in the matrix (X

X)

being non existing. Furthermore, if some values of t

are close to zero near linear dependence exists and the matrix (X

X) will be ill-conditioned.

Multicollinearity causes a problem when solving the least-squares normal equations

(X

X) ˆ β = X

y (17)

If there is linear dependence between the i:th and j:th regressor variable their simple correlation r

will be close to unity which will inflate the estimation of ˆ β

and result in large variance and covariance. One way to detect multicollinearity

14

is by examining the variance inflation factor. Let C = (X

X)

. The diagonal elements of C can be expressed as C

= (1 − R

)

where R

is the coefficient of determination when regressing x

on the p − 1 remaining regressors. If there is linear dependence R

will be close to unity which leads to C

→ ∞. This causes a problem since the V ar( ˆ β

) = C

σ

will be inflated by the value of C

. 2.3.3.2 Homoscedasticity

Homoscedasticity is when the error terms, ε, have constant variance and is a key assumption for the model. The opposite of this characteristic is heteroscedas- ticity. Modest heteroscedasticity in the model have little effect for the test of significance, however it can lead to severe disfigurement of findings and weaken the analysis. (Osborne and Waters 2002)

2.3.4 Variable Selection

When constructing the model all regressor variables are included but they all may not be important for describing the response variable. Variable selection is the method to select a subset of regressor variables that in the best way describe the response. It is also an effective way to deal with multicollinearity.

It is important to find a balance meaning that as many regressor variables are included to provide more information of y and at the same time as few as possible since the variance of ˆ y increases with the number of variables. One measurement to asses the different sub models is Adjusted R

and is used to simplify the interpretation of R

. It is defined as

R

= 1 − ( n − 1

n − p )(1 − R

) (18)

for a subset of p variables. The subset model with the highest R

is the one that is optimal. Another criteria to asses the sub models is Mallow’s C

statistic. It is defined as

C

= SS

(p) ˆ

σ − n + 2p (19)

where SS

(p) = y

y − ˆ β

X

y for the subset model with p variables. Small values for Mallow’s C

are wished for.

All possible regression is the method used to review all sub models that can be constructed. By using the criterion mentioned above all models can be evaluated and the best candidate selected.

15

3 Methodology

The study is performed by selecting Life science companies in Sweden that are comparable by using Thomson Reuters Eikon. Then using multiple linear regression in order to find a model for market capitalization based on the mul- tiples. The Life science sector includes companies of the following industries:

pharmaceuticals, biotechnology and medical devices.

The data was collected at 26.03.2019 at 9am. The data sources are Thom- son Reuters Eikon, Infront, Avanza and annual reports from the companies websites. Values for all variables, Market Capitalization in million SEK, Price to Earnings, Return on Equity, Debt to Equity, Price to Book Value and Re- search and Development Expenditure in SEK were taken from the data sources.

Several sources for data is used since values for all variables do not exist on only one site. The methodology for collecting data was iterative. The primary source for data was Thomson Reuters Eikon, then Infront, then Avanza and lastly data from annual reports. Companies with missing or inconclusive data were removed from the study. The full list of companies is showed in Appendix.

Regressions analysis is made in R by following the mathematical theory.

Checking model adequacy by doing the following: normal probability plot, built in function summary() in R for summarized results of the model, and plots of residuals against fitted values. Then to investigate leverage and influence the built in function plot() in R is used and calculations to determine Covratio, Cook ´s distance and Hat values. Potentially influential observations is deleted from the model and further analysis is made. Added Variable Plot is used to see the affect from the variables on the intercept. Further the multicollinearity is investigated by the correlationmatrix and variance inflation factor. Lastly vari- able selection is made by all possible regression with R

-adjusted and Mallow’s Cp, and built in function used for summarized results.

16

4 Results

Results for the analysis in a chronological order following the methodology:

Model adequacy, Multicollinearity and Variable selection. Presented in text with associated figures and tables.

4.1 Model Adequacy

Normality assumptions is a key part of the residual analysis for model adequacy.

The normal probability plot shown in Figure 1 shows that the standardized residuals follows the pattern for normal distribution. There are some deviant observations in the tails that need further investigation.

Figure 1: Normal Q-Q plot for Standardized Residuals with marked deviant observations.

The estimations and statistical measurements for the coefficients are presented in Table 1 below. The magnitude of the standard errors varies for the variables were some are much larger than others. P/E ratio and R&D expenses are the explanatory variables with the lowest standard error whereas, the D/E ratio and the intercept have the highest. All t-values are lower than 0.85 except for R&D expenses that has a value of 17.80. The p-values for the coefficients are as a result from the t-value also very large, above 0.4, except for the coefficient of R&D expenses which is lower than 2 · 10

.

17

Table 1: Summarized results for multiple linear regression for each variable.

Coefficients Estimate Standard Error t-value Pr(>|t|)

Intercept -85.30 398.20 -0.21 0.83

P/E -0.20 0.54 -0.37 0.71

ROE 11.55 46.87 0.25 0.81

D/E -138.74 180.80 -0.77 0.44

P/B 16.06 19.00 0.85 0.40

R&D 2.98 · 10

1.68 · 10

17.80 < 2 · 10

The residuals plotted against their fitted values are shown in Figure 2 down below. The result does not show a clear trend, however a large proportion are clustered together.

Figure 2: Residual values plotted against fitted values.

The result from examining the studentized residuals is the same from the ordi- nary residuals and is shown in Figure 3 down below.

18

Figure 3: Studentized residual values plotted against fitted values.

4.1.1 Leverage and Influence

Results presented in Figure 4 shows standardized residuals versus leverage with curves for Cook ´s distance. Most of the observations are clustered with residual values between ±2.5 and leverage close to 0. The four observations outside the red dotted line have high Cook ´s distance.

Figure 4: Standardized residuals versus leverage with marked curves for high Cook’s distance.

19

A representation of the observations which fulfills the criteria of either Covra- tio, Cook’s distance or Hat values for determining influential observations are presented in Table 2. Observation 78 and 164 are the ones with the most de- viant values across all criterion with respect to the cut-off values. Furthermore, observation 1, 9 and 138 show deviations from cut-off values in two or three criterion.

Table 2: Showing selected observations with high potential influence and lever- age with represented values of Covratio, Cook's Distance and Hat values.

Observation Covratio Cook’s D Hat values

1 0.12 3.46 0.33

2 1.54 0.00 0.32

4 0.28 0.04 0.01

9 0.08 1.66 0.18

28 0.73 0.04 0.03

72 2.47 0.22 0.59

78 27.07 3.18 0.96

107 1.24 0.01 0.16

138 9.42 0.01 0.89

151 1.15 0.00 0.09

164 29.68 7.17 0.97

Further examination of the observations above was done by deleting the above observations. It resulted in an improvement of the model and were therefore excluded in the following analysis.

4.1.2 Added Variable Plot

Results for Added variable plots represented in Figure 5. The plots is showing slight positive affect from P/E, ROE and P/B. The RD have steeper inclination with a positive correlation indicating a larger impact on Market capitalization.

D/E has low negative slope and therefore, less decreasing impact on Market capitalization.

20

Figure 5: Shows Added Variable Plots for all variables.

4.2 Multicollinearity

4.2.1 Correlation Matrix

The correlation matrix is presented below in Table 3 where the off diagonals show the simple correlation between the regressor variables. No high degree of simple correlation is present since no value on the off diagonals is close to unity.

21

Table 3: Correlation matrix showing simple correlation on off-diagonals.

P/E ROE D/E P/B R&D

P/E 1.0000 -0.0541 0.0341 -0.1022 0.0746

ROE 1.0000 0.3116 0.0742 0.1145

D/E 1.0000 0.4021 0.0478

P/B 1.0000 -0.0182

R&D 1.0000

4.2.2 Variance Inflation Factor (VIF)

No value of the VIF for each of the variables exceeds 5, which can be seen in Table 4, indicating a low presence of multicollinearity.

Table 4: Variance inflation factor for regressor varibles.

P/E ROE D/E P/B R&D 1.03 1.13 1.33 1.22 1.02

4.3 Variable Selection

4.3.1 All Possible Regression

Using R

adjusted as evaluation criteria, all possible regression suggests 4 sub models and the full model to be equally good. This is represented in Figure 6 where the colored boxes represents which variables are included for that sub model, where the sub model is represented by the vertical axis.

Figure 6: Results from all possible regression with adjusted R

as criteria.

By using Mallow’s Cp instead as criteria for the all possible regression the best sub model is R&D as the only explanatory variable followed by ROE and R&D as the second best sub model. This can be seen in Figure 7

22

Figure 7: Results from all possible regression with adjusted Mallow’s Cp as criteria.

By using All Possible Regression four sub-models is presented. These models estimate, standard error, t-value and p-values is shown i Table 5. For all sub- models the the variable RD has t-values over 20. That variable also have low p-values for all the sub-models, < 2 · 10

. Another observation from the summarized data is that none of the sub-models has consistently clear figures for t-values or p-values.

Table 5: Summarized results for multiple linear regression for all sub-models from all possible regression.

Estimate Standard Error t-value Pr(>|t|) Sub-model 1

Intercept -54.46 193.80 -0.281 0.779

R&D 2.804 · 10

1.187 · 10

23.621 < 2 · 10

Sub-model 2

Intercept 85.30 219.20 0.389 0.698

ROE 182.40 131.30 1.390 0.167

R&D 2.79 · 10

1.20 · 10

23.292 < 2 · 10

Sub-model 3

Intercept 144.50 228.60 0.632 0.529

P/E 2.24 2.22 1.008 0.315

ROE 191.20 132.60 1.442 0.152

R&D 2.76 · 10

1.21 · 10

22.932 < 2 · 10

Sub-model 4

Intercept 52.36 252.60 0.207 0.836

P/E 2.43 2.24 1.086 0.280

ROE 183.00 133.10 1.375 0.172

P/B 13.61 15.78 0.862 0.390

R&D 2.77 · 10

1.21 · 10

22.919 < 2 · 10

23

5 Discussion

5.1 Model Adequacy

The assumption of normal distribution of the residuals seemed to hold based on the result presented in Figure 1. There were some observations that lied far from the line in the graph, indicating that they probably did not follow normality.

By further investigation of these observations along with the ones in the end of the tails it was concluded that they were influential or outliers that should be excluded in the model. One does have to be critical about these results since there were still some tail formation in Figure 1 even after elimination. This indicated that the model used after elimination still did not follow a perfect normal distribution but one that is sufficient to proceed with model evaluation.

Table 1 indicated that the coefficient estimates for the initial model were poorly estimated for the majority of the explanatory variables. The standard errors were severely high and the p-values were also high. The standard error is the approximate standard deviation of a statistical sample. Which means that the observations of the variables P/E, ROE, D/E and P/B included in the model are on average relatively far from the mean. High p-value indicates that the variables do not describe the response in a significant way. Therefore, the model does not hold for these variables. As can be seen in Figure 2 the residual values are clustered between 0 and 10000 which can be an explanation to why the model performed poorly. Perhaps the data set is to homogeneous and therefore, enough data to approximate larger or smaller values of market cap is missing. A good data set would result in a more widely spread plot.

However, since this is not the case these results must be interpreted critically.

To deal with the clustering, additional data could be sampled by expanding the search area to a continental delimitation.

5.2 Leverage and Influence

Some observations from the initial model were influential or outliers and were therefore removed from the data set. This resulted in a more accurate model based on the adjusted R

and the t-statistic. However, only R&D expenses had sufficient results for indicating a linear relationship with the response variable.

Furthermore, it could also be seen that P/E ratio and ROE increased in sig- nificance but not with a magnitude that is needed to conclude that there is a correlation. It is important to mention that some of these observations were in fact well known companies with more accessible data other than what was collected and are probably easier for investors to analyze since they have more historical data and a steady cash flow. The larger companies may be easier to build a model for. However, this might still be the case but at least not with a multiple regression model based on our results. A possible explanation for why some of the large companies were outliers or influential observations is that these companies are so big that the selected multiples do not capture the entirety of their business, and therefore reflects the market capitalization poorly.

24

5.3 Multicollinearity

No significant multicollinearity was present in the model a part from a relatively small simple correlation between D/E and ROE, and between D/E and P/B.

However, in combinations with the findings in Table 2 it can be concluded that no multicollinearity that would inflate the coefficient estimates is present.

As discussed in section 2.3.3 Model Properties under Multiocollinearity this is an important result for the model since the variance of the coefficients is not inflated. Therefore, multicollinearity is not the reason why the coefficient estimates standard deviations are large in Table 1 for some of the regressors.

It is more likely that it is a result from their low significance resulting in bad estimates.

5.4 Variable Selection

The variable selection of the full model made by all possible regression presented four sub-models by R

-adjusted and Mallow’s Cp. A summary of the models is shown in Table 5. The both criterion presented the same four sub-models.

For the adjusted R

criteria there is no improvement for the sub-models, they have the same value 0.82. For the variables P/E, ROE, P/B in the sub-models they have no significant statistical results to prove a good explanation of market capitalization, see Table 5. Which is almost the same for the full model presented in Table 1. Generally low t-values and high p-values, indicating that the null hypothesis that there is no correlation between market capitalization and the multiples have more support.

Furthermore, the results are showing that there is a correlation between the response and R&D expenses. It has a low standard error, high t-value and low p-value (Table 1). As seen in Figure 5, R&D expenditures has the clearest and steepest slope in relation to the other variables. The variable was also selected in both cases in section 4.3.1 All Possible Regression with criterion R

-adjusted and Mallow’s Cp. This means that depending on a firms market capitalization there is a correlation to how much they spend on research and development. Which has a logical explanation since firms often goes to the stock exchange to raise capital for further investments, and the more potential the market see in a company the more capital will be raised resulting in higher market capitalization. However, the usefulness of this simple model is not so good. The only thing that can be seen is if a company in this industry is investing more or less in R&D in relation to the other in the market. It is more of an explanation and confirmation of how the system work, and that in this industry the companies are actually investing more and more money in research and development even if they are bigger and more developed companies.

5.5 Market Psychology

Our regression model included only quantitative measurements of business per- formance but there are several qualitative factors that can influence the market

25

cap of a company. Since all the companies that were examined are listed on the stock exchange, individual and institutional investors judgment can inflate or deflate a company’s value (Lubatkin et al. 1989). This could be an explana- tion for why our model did not perform well since it is only based on some key financial ratios. An increase in market cap might not come from an increase or decrease of a certain financial ratio but as a result from a rumor or herding (Lubatkin et al. 1989). For example, when investors are pessimistic about the future the most uncertain life science companies that are associated with high risk will not be attractive investments based on the theory that people are risk averse. Furthermore, it could be a result from herding when the market is in a negative trend and investors are selling their assets resulting in a snowball effect where more and more investors sell even though the financial ratios of a company have not changed. Misattribution bias could be another factor that affects the market cap of a company when institutional or individual investors act based on current feelings or emotions (Lubatkin et al. 1989). An example of this is when a company appoints a new CEO, as a result investors tends to ex- pect higher returns when the appointed person is from outside the company and therefore, the value of a company increases. However, large companies tend to run themselves because of their organizational structure leading to limitations of the CEO’s power. In fact, a new CEO might cause disruption if fundamen- tal changes of the organizations structure is on that person’s agenda (Lubatkin et al. 1989).

5.6 Limitations

First it should be mentioned that the scope of listed life science companies in Sweden is a quiet naive delimitation. These companies are often working in the worldwide market where they have potential customers everywhere. Probably it could have been more interesting in looking at more companies, which also would increase the number of observations and make the results more reliable.

Another thing is that the hypothesis and explored variables are just a se- lection and do not represent a complete set of potential influential variables.

Except from the number of chosen variables there is limitations with selected variables. Since many of the companies in the industry are pre-revenue some multiples are not that useful. Instead of focusing on its potential and success with positive numbers, it is focused on minimizing companies losses. Somehow, this is not correct for this type of companies in an expansive phase. In early stage of development higher cost many result in better innovation and more successful company in the future.

The regression analysis that have been done includes real time market data such as the price of the stock and reported financial data. The market and industries change over time and require constant analysis. This is limiting the thesis usefulness. Which opens for an approval to investigate the change of multiples over time.

26

6 Conclusion and Future Research

This study examines if there exists any linear correlation between market cap- italization and selected multiples for life science companies in Sweden. No ev- idence for explaining the variation in the response of the variables P/E, ROE, D/E and P/B was found. However, there was a correlation between R&D ex- penditures and market capitalization. Which can be discussed not to be useful, instead more as a confirmation of reasonable thinking. Moreover, market multi- ples needs further investigation to be a useful tool for comparing and valuating companies in similar industries.

Since no reasonable correlation between the variables in the model and mar- ket capitalization, it should be interesting to expand the research for other markets with other available market multiples. Moreover, other multiples than the ones used in this study would be interesting to examine, for example growth since it could describe future success. Since multiple linear regression was used in this study and only one relationship between market capitalization and R&D expenditure could be determined it would be interesting to examine other mod- els with different methods. This could be done using the same predictor variables in order to determine if there is any model that includes these. If no relationship is found then one can conclude that they poorly estimates market cap and that new predictor variables must be selected.

Qualitative measurements and their effect on market capitalization is an- other approach that would be interesting to study. How this should be done is a difficult decision since the problem lies in collecting data. However, if this could be measured and standardized it would probably result in a good compliment to other quantitative valuation methods that exits on the market today. In this way a more covering valuation standard could perhaps be used.

27

References

Akademia Baru, Penerbit et al. (2016). “Price to Book Value, Price to Sales Multiples and Stock Price; Evidence from Nigerian Listed Firms”. In: Jour- nal of Advanced Research in Business and Management Studies 3, pp. 85–

93.

Billmeier, Andreas and Isabella Massa (2009). “What drives stock market de- velopment in emerging markets—institutions, remittances, or natural re- sources?” In: Emerging Markets Review 10.1, pp. 23–35.

Cheng, CS Agnes and Ray McNamara (2000). “The valuation accuracy of the price-earnings and price-book benchmark valuation methods”. In: Review of Quantitative Finance and Accounting 15.4, pp. 349–370.

Crean, D (2016). Valuation Methodologies for Life Science Companies. url:

https://www.linkedin.com/pulse/valuation- methodologies- life- science-companies-crean-ph-d-mba.

Damodaran, Aswath (2007). “Return on capital (ROC), return on invested cap- ital (ROIC) and return on equity (ROE): Measurement and implications”.

In:

Dayal, Mohit (n.d.). “Why Added Variable Plots”. In: Applied Statistics and Computer Lab.

Dewatripont, Mathias and Jean Tirole (1994). “A theory of debt and equity: Di- versity of securities and manager-shareholder congruence”. In: The quarterly journal of economics 109.4, pp. 1027–1054.

Exchange, London Stock (2010). “A guide to listing on the London Stock Ex- change”. In: White Page Ltd: London.

Ghaeli, M (2017). “Price-to-earnings ratio: A state-of-art review”. In: Account- ing 3.2, pp. 131–136.

Guellec, Dominique and Bruno van Pottelsberghe De La Potterie (2002). “R&D and productivity growth”. In: OECD Economic studies 2001.2, pp. 103–126.

Harbula, P´ eter (2009). “Valuation multiples: Accuracy and drivers evidence from the european stock market”. In: Business Valuation Review 28.4, pp. 186–

200.

Lubatkin, Michael H et al. (1989). “Stockholder reactions to CEO changes in large corporations”. In: Academy of Management Journal 32.1, pp. 47–68.

Montgomery, Douglas C, Elizabeth A Peck, and G Geoffrey Vining (2012). In- troduction to linear regression analysis. Vol. 821. John Wiley & Sons.

Osborne, Jason and Elaine Waters (2002). “Four assumptions of multiple regres- sion that researchers should always test”. In: Practical assessment, research

& evaluation 8.2, pp. 1–9.

Rothenstein, Jeffrey M et al. (2011). “Company stock prices before and after public announcements related to oncology drugs”. In: Journal of the National Cancer Institute 103.20, pp. 1507–1512.

28

Appendix

List of Companies included in the analysis from Thomson Reuters Eikon.

Swedish Orphan Biovitrum AB Elekta AB

Getinge AB Vitrolife AB SECTRA AB Hansa Biopharma AB Arjo AB

Biotage AB Recipharm AB

Wilson Therapeutics AB AddLife AB

Karo Pharma AB Oncopeptides AB BioArctic AB Ambea AB Xvivo Perfusion AB Medicover AB Humana AB Camurus AB Immunovia AB Orexo AB

RaySearch Laboratories AB IRLAB Therapeutics AB Cellink AB

Handicare Group AB Sedana Medical AB

Infant Bacterial Therapeutics AB

Oasmia Pharmaceutical AB Alligator Bioscience AB Calliditas Therapeutics AB Q Linea AB

Cantargia AB Dextech Medical AB SyntheticMR AB Boule Diagnostics AB Moberg Pharma AB XSpray Pharma AB Bonesupport Holding AB Bactiguard Holding AB SpectraCure AB C-Rad AB

GHP Specialty Care AB PledPharma AB Ortoma AB Swedencare AB Genovis AB

BioInvent International AB Isofol Medical AB Immunicum AB Irras AB

Vicore Pharma Holding AB Brighter AB

Paxman AB

Surgical Science Sweden AB Ascelia Pharma AB Mertiva AB Asarina Pharma AB ACTIVE Biotech AB

InDex Pharmaceuticals Hold- ing AB

Diamyd Medical AB Elos Medtech AB Hamlet Pharma AB SenzaGen AB

Phase Holographic Imaging PHI AB

Medivir AB Intervacc AB

Karolinska Development AB Stille AB

Klaria Pharma Holding AB Lidds AB

Senzime AB Enorama Pharma AB RhoVac AB

Clinical Laserthermia Systems AB

AlzeCure Pharma AB RLS Global AB Corline Biomedical AB Xbrane Biopharma AB Feelgood Svenska AB Respiratorius AB Iconovo AB

NeuroVive Pharmaceutical AB Arcoma AB

Promore Pharma AB Xintela AB

BrainCool AB Dignitana AB S2Medical AB

Scandinavian Real Heart AB Kancera AB

Sprint Bioscience AB Spago Nanomedical AB Glycorex Transplantation AB Medfield Diagnostics AB Invent Medic Sweden AB Bio-Works Technologies AB Nanexa AB

IDL Biotech AB

ISR Immune System Regula- tion Holding AB

2cureX AB Doxa AB

Double Bond Pharmaceutical International AB

Wntresearch AB Redsense Medical AB Alzinova AB SynAct Pharma AB A1M Pharma AB Cyxone AB

Hemcheck Sweden AB Integrum AB AroCell AB

Prolight Diagnostics AB Episurf Medical AB Biovica International AB Miris Holding AB

Zenicor Medical Systems AB

AdderaCare AB Follicum AB Acarix AB Kontigo Care AB Bioservo Technologies AB Ortivus AB

Ziccum AB ScandiDos AB AcouSort AB

VibroSense Dynamics AB Toleranzia AB

AcuCort AB Cereno Scientific AB Redwood Pharma AB BibbInstruments AB

Inhalation Sciences Sweden AB MediRatt AB

Micropos Medical AB NextCell Pharma AB MedicPen AB Lumito AB Prebona AB

ExpreS2ion Biotech Holding AB

Combigene AB Calmark Sweden AB Gabather AB Peptonic Medical AB Fluicell AB

SciBase Holding AB Alteco Medical AB

Karessa Pharma Holding AB ObsteCare AB

ProstaLund AB PharmaLundensis AB Eurocine Vaccines AB AlphaHelix Molecular Diagnos- tics AB

Pharmacolog I Uppsala AB Attana AB

Dicot AB PExA AB

Scandinavian ChemoTech AB Idogen AB

Annexin Pharmaceuticals AB SensoDetect AB

Aptahem AB

Chordate Medical Holding AB Cline Scientific AB

Polarcool AB Camanio Care AB LifeAssays AB

Panion Animal Health AB Quickcool AB

European Institute of Science AB

Papilly AB Curando Nordic AB Emotra AB

29

TRITA -SCI-GRU 2019:156

www.kth.se