The use of forward-looking multiples for cyclical industries

The use of forward-looking multiples for

cyclical industries

Master Thesis

MSc International Financial Management

1

The use of forward-looking multiples for

cyclical industries

Ben Koomen

Master Thesis

Abstract

Keywords: firm valuation, multiples, cyclical industries

JEL classifications: G12, G30

This paper examines whether or not cyclical industries have on average higher absolute valuation errors than non-cyclical industries using the multiple valuation approach. In order to examine this, two subsamples of cyclical and non-cyclical industries using EV/EBITDA, EV/EBIT, EV/SALES

and P/E multiples are benchmarked against each other. As the appropriateness for forecasts in combination with cyclicality is doubtful, we also examine whether or not forecasted data improves

the valuation accuracy for cyclical firms. We find significantly higher valuation errors for the cyclical subsample. Moreover, the use of forward-looking multiples only seems to be worthwhile

2

1. Introduction

“Two well-respected strategists present a clear message to investors today: Buy the stocks of companies exposed to cyclical strength in the U.S. economy. Given the recent acceleration of U.S. gross domestic to 4.1 percent, and the lowest unemployment rate since September 2008, cyclicals are significantly outperforming. The S&P North American Cyclical Sector Index has risen 19.9 percent this year, five times the return of the defensively oriented S&P 500 Consumer Staples Sector Index.1”

In the news article above, the fortunes of a cyclical firm rely for a large part on movements of macro-economic variables such as GDP growth and commodity prices. This causes the earnings and cash flows of cyclical firms to demonstrate repeating patterns of significant increases and decreases along with the cycles of economic growth (Koller et al. 2010). Even mature cyclical firms will display volatility in their earnings and cash flow patterns, as the means to account for these dependencies are limited (Damodaran, 2009). While uncertainty is to a certain extent endemic to valuation, these cycles form an extra challenge to value a cyclical firm. The challenge lies in the fact that these cycles are hard to account for. However, in practice, analysts even often seem to ignore cyclicality all together. They tend to overestimate real earnings, both when earnings are rising and when they are in decline (de Heer et al. 2000; Koller et al. 2010). Moreover, if they do take it into account, correctly predicting a firm‟s cycle is easier said than done (Damodaran, 2009). Recently reported earnings and cash flows are a function of where a firm is in its cycle, ignoring or misinterpreting these cycles, results in biased and incorrect estimates. Extrapolating those numbers into the future can lead to serious misvaluation (Damodaran, 2009).

The aim of this paper is to investigate whether or not the valuation errors2 are higher on average for cyclical firms. We will do so by using a relative valuation approach called multiples. This particular valuation approach is relevant because after the discounted cash flow method (DCF) it is in practice ranked second in terms of use, and frequently featured in academic research (Imam et al. 2008). When using multiple for valuation, firms are valued based upon how comparable firms are valued by the market. In other words, a firm is valued relatively to its closest peers. Comparability is therefore critical to the success of multiple

1 _{http://www.bloomberg.com/news/articles/2014-09-09/going-with-the-momentum-in-cyclical-stocks} 2

3

valuation. As a result, how to select the best comparable peer group is a widely studied topic (Liu et al. 2002; Bhojraj and Lee, 2002; Dittmann and Weiner, 2005; Schreiner and Spremann, 2007; Tu, 2010). While various authors stress to treat cyclicality and valuation with care (de Heer et al. 2000; Damodaran, 2009: Koller et al. 2010), it has been largely left out of the loop. This is a pity, as cyclicality may very well be one of the factors that compromise comparability in certain peer group selection practices. This problem will be discussed more in-depth in the literature review, but the basic principle of this assumption is visualized by the earnings/cashflow cycle in the figure below3.

While the contrast between the two firm cycles displayed in Figure X is perhaps a somewhat extreme, it adequately visualizes the idea how two different cycles can impede comparability between firms. The figure depicts the earnings or cashflow cycle of firm A and B. It can be seen that at the same point in time the cycle of firm A is going upwards, while the earnings of B are actually in decline. Imagine constructing a peer group of few of these firms, as they are actually incomparable, valuing them relatively to each other would leave firm A undervalued and firm B overvalued. In the literature review we will discuss why certain practices in multiple valuation are especially prone to possibly matching incomparable cyclical firms. The

4

main notion is however, that if not predicted and matched correctly, cycles can lead to peer groups with low comparability.

The only study who has empirically addressed multiple valuation and cyclicality so far is Tremolizzo (2009). While lacking a solid theoretical foundation, he did however find significantly higher valuation errors on average for cyclical firms. The major drawback is that his research did not extensively incorporate forward-looking multiples while forecasting is exactly considered one of the main hurdles when valuing cyclical firms (Damodaran, 2009; Koller et al. 2010). Hence, we expect that forward-looking multiples do not increase valuation accuracy to the same extent as they do for non-cyclical firms in the studies of Liu et al. (2002) and Schreiner and Spremann, (2007).

This study will take place in an international context as Bhojraj and Lee (2003) argue that increased global competition pushes firms to conduct business abroad. It is therefore more likely that domestic players face foreign competition in their home country, implying that comparable firms are also more likely to be found across boundaries. However, it is also assumed that national differences in legal systems and tax accounting practices may reduce comparability across firms (Doupnik and Pereira, 2007) as Liu et al. (2002) and Dittmann and Weiner (2005) found significant country effects on valuation accuracy. These studies regarding country effects do not specifically address the issue of cyclicality. The idea is however, that cyclicality can also have a profound influence as the macro-economic variables that determine firm cycles can significantly differ per country.

Because there is almost a complete lack of literature regarding both multiple valuation and cyclicality, we base our problem statement on a combination of literature concerning multiples and valuation literature concerning cyclicality. We expect to see higher valuation errors on average for cyclical industries. Moreover, literature also generally supports the use of forward looking multiples while at the same time it is hard to arrive at accurate forecasts for cyclical firms (de Heer et al. 2000; Damodaran, 2009; Koller et al. 2010). Therefore we also question whether forward-looking multiples will improve the valuation accuracy. This leads to the following two research questions:

5

The valuation accuracy will be determined for the EV/EBIT4, EV/EBITDA5 , EV/SALES6 and P/E7 multiple. These multiples are chosen because they are commonly used in practice and academic research (Schreiner and Spremann, 2007; Imam et al. 2008; Koller et al. 2010). The research questions can be answered by means of deploying parametric tests to test for differences across cyclical and non-cyclical distributions and by running a regression analysis to see whether cyclicality has a significant effect on valuation accuracy. In order to do so we will construct two subsamples of cyclical and non-cyclical industries following the classification by Boudouk et al. (1994) and Pearce and Michael (2006). Subsequently we select for these industries by using the matching two-digit NACE rev. 2 codes. In order to incorporate both trailing and forward-looking multiples, the time horizon will span the years 2013 - 2017. The base year 2013 represents the trailing multiples and the forward-looking multiples are represented by the year 2014 onwards. This finally resulted in a sample of 961 listed companies from the U.S, Canada and E.U.

The research is academically and practically relevant, because the issue of cyclicality has not yet been extensively addressed for multiple valuation. Moreover, established practices in multiple valuation such selecting peers on the basis of industry are questioned whether accurate or not by literature (Damodaran, 2009). In addition, we will test whether or not the use of forward-looking multiples will also improve the valuation accuracy for cyclical firms as this is also doubtful. Lastly, unlike most previous research, this study will take place in an international context.

This paper will continue in the following way: the next section will discuss and review the relevant literature regarding multiples & cyclical firms. This will be followed by the methodology section where data and empirical tests will be presented. Discussion and outcomes of these tests will take place in the results section. Finally this paper will end with a conclusion, implications for management and limitations of this research.

4 _{Enterprise Value/Earnings Before Interest & Taxes}

5 _{Enterprise Value/Earnings Before Interest, Taxes, Depreciation & Amortization} 6 _{Enterprise Value/Sales}

6

2. Literature Review

In order to increase the legitimacy for the use of multiples in this paper it is important to quickly discuss the various methods for corporate valuation. Firm value can either be estimated by a direct or relative method. With a direct method, such as the discounted cash flow (DCF) or dividend discount model (DDM) the value is estimated on the basis of the future expected cash flows. Relative valuation models like multiple analysis estimate corporate value on the basis of how similar firms or assets are valued in the market (Damodaran, 2009). Following a survey among UK analysts, Imam et al. (2008) concluded that it is difficult to determine why there should be a preference for one model over the other. While some preferences seem to be industry related (Barker 1999; Demirakos et al. 2004), other analysts cover similar industries with different models (Liu et al. 2002). Furthermore Imam et al. (2008) found that DCF and DDM are often used with multiples as a secondary model to triangulate results.

Koller et al. (2010) provide a sensible explanation for this. The direct discounted cash flow analysis (DCF) is arguably more accurate and flexible compared to the relative multiple valuation. However, every analysis can only as accurate as the data it relies upon. Therefore a multiple analysis – where a firm is directly compared with its peers – is often used to triangulate results obtained from other valuation approaches (Imam et al. 2008). By taking the ratio of a market price variable, such as stock price for instance, to a particular value driver such as earnings one can easily calculate a firm‟s multiple. Subsequently, based upon how comparable firms are valued by the market, a user can come up with quick estimate of firm value (Schreiner and Spremann, 2007).

2.1 Peer selection

7

assumed that industry membership itself will serve as a sufficient proxy for risk and earnings growth. Alford (1992) argues that industry selection will also mitigate most cross-firm accounting differences as industries tend to adopt similar accounting practices (Foster, 1986). Despite global harmonization efforts in accounting this is still relevant today. As potential comparability problems across industries still exist (Barth et al. 2012).

In order to select for industry membership, Alford (1992) used Standard Industry Classification (SIC) codes. He found that valuation accuracy increased when two-digit SIC codes were used instead of single digit. Further narrowing down to three-and four-digit codes did not result in significant valuation improvements. Schreiner and Spremann (2007) also constructed their peer group by means of industry classification. They found that when narrowing down industry classification from single to three-digit valuation accuracy did improve. While the results between Alford (1992) and Schreiner and Spremann (2007) slightly differ, the basic idea is the same. Narrower industry classification yields smaller but presumably more homogeneous peer groups. Up to which extent this is still worthwhile seems debated.

Instead of constructing their peer group by industry, other researchers chose to form peer groups on the basis of systematic multiples. Systematic multiples are based on firm fundamentals such as profitability, growth rate, risk proxies (Bhojraj and Lee, 2002) or size (Cheng and McNamara, 2000). Bhojraj and Lee (2002) found that the systematic multiple based peer groups outperformed peer groups formed on the basis of industry in terms of accuracy. In reaction Lie and Lie (2002) point out that peer selection on the basis of firm fundamentals will most likely involve a high degree of subjectivity as the selection criteria for a peer group may become rather arbitrary. Given the ambiguity regarding selection criteria, it is hard to establish which method is best, as the preferences seem rather arbitrary in general.

8

2.2 Many Multiples

Similar to the many roads that lead to Rome for peer group selection a vast amount of multiples exists. Being a relative valuation method, firm value is based upon how comparable firms are valued by the market. The multiple itself is derived from the median or average value from a certain (accounting) measure of the peer group. Subsequently multiplying the multiple with the target firm‟s (accounting) measure will yield an approximate firm value (Koller et al. 2010). We can make a main distinction between 1. Enterprise multiples (e.g. EV/EBITDA), 2. Equity multiples (P/E multiples) and 3. Sales multiples (EV/Sales). Multiples can range from tangible to less tangible (non-financial) metrics and can sometimes be industry specific (Imam et al. 2008; Koller et al. 2010). Similar to the use of multiples in research, there is great variety in their use among practitioners as well (Imam et al. 2008). For the results of their survey, see table 2 on page 9.

Table 1. Overview of peer selection method and multiples used in previous research.

Author(s): Peer selection method Multiples

Alford (1992) Industry, firm size, earnings growth P/E

Bhojraj and Lee (2002) Systematic multiple EV/Sales, M/B

Lie and Lie (2002) Industry EV/EBITDA, EV/EBIT,

EV/Sales, P/E

Liu et al. (2002) Industry P/EBITDA, EV/EBITDA,

EV/Sales, P/Sales Richter (2005) Growth, industry, leverage, ROE P/E, M/B

Dittmann and Weiner

(2005) Country, region, OECD

EV/EBIT

Schreiner and Spremann

(2007) Industry

P/E, M/B, P/Sales

9

In line with Koller et al. (2010) practitioners tend to prefer enterprise based multiples (Imam et al. 2008). The EV/EBITDA multiple tells more about a firm‟s value than any other as it encompasses the firm‟s growth, return on invested capital, operating tax rate and cost of capital (Koller et al. 2010). According to Koller et al. (2010) for nearly all peer groups using EBITDA instead of EBIT results in better enterprise values.

A statement backed up by Baker and Ruback, (1999), Lie and Lie, (2002) and Hermann and Richter, 2003, who found similar results in practice. Contrary to the support of enterprise multiples, Schreiner and Spremann (2007) found that the P/E multiple (equity based) outperformed enterprise multiples. Koller et al. (2010) argues that although widely used (supported by table 2) the P/E multiple has two severe drawbacks. First of all, next to the firm‟s operating performance the P/E ratio is also affected by the firm‟s capital structure. Secondly, net income (earnings) is derived after amortization and one-time events (losses or gains). A non-operating loss such as a write-off can therefore significantly lower firm earnings and thereby cause an artificially high increase in the P/E ratio. As discussed earlier, table 2 also confirms that there does not seem to be a very strong industry related preference for the most frequently used multiples. The financial industry is an exception with a seemingly overwhelming preference for P/E multiples.

When constructing multiples, the second step is to pick a set of multiples of your preference.. Multiples that are frequently featured in literature are EV/EBIT, EV/EBITDA, P/E and Price/Sales multiples (Alford, 1992; Bhojraj and Lee, 2002; Lie and Lie, 2002; Liu et al. 2002; Schreiner and Spremann, 2007). Although common in practice, empirical research offers limited evidence on the existence of industry-preferred multiples. Research by Tasker (1998) and Barker (1999) addresses for example that P/E multiples are preferred in the Table 2.

Frequency of the use of valuation methods among UK analysts (Imam et al. 2008).

Model/Industry Financial Industrial Media Retail Tech Total

1. DCF 1 10 15 11 12 49

2. P/E 15 4 4 15 7 45

3. EV/EBITDA 0 4 4 12 5 25

4. DDM 10 0 0 0 0 10

10

financial industry. However these studies do not provide conclusive evidence that these industry-preferred multiples are also the most accurate

Nonetheless, when dealing with cyclical industries Schreiner and Spremann (2007) argue that multiples should be constructed with data higher up in the income statement. As firms frequently report weak or even negative earnings multiples towards sales might me be more appropriate. Damodaran (2009) elaborates on this by stating that it is useful to use EBIT(DA) multiples when valuing cyclical firms for the following two reasons. Firstly, because operating income is less volatile than net income, the multiples will naturally remain more stable over time due to their less volatile denominator. Moreover, EBIT(DA) multiples can even be computed in the midst of a downturn, whereas P/E multiples tend to result in smaller samples as earnings become negative. Damodaran (2009) however stresses that investors ultimately only care about the earnings and cash flows. Therefore it is important to control for volatility across firms via careful peer selection at all times.

Another important facet of multiple valuation is whether to use forward-looking or trailing multiples. Research has consistently shown that forward-looking multiples lead to lower valuation errors than trailing multiples (Liu et al. 2002; Schreiner and Spremann, 2007; Tu, 2010). Koller et al. (2010) explain that forward-looking multiples are also more in line with the basic principles of valuation. I.e. firm value is represented by the value of its future cash flows, not its past or sunk costs. Moreover, forecasted data is typically normalized and thus avoids the problem of one-time events. Even though normalized, Koller et al. (2010) and Damodaran (2009) question the accuracy of these forecasts, thereby implicitly questioning the straightforward use of forward looking multiples for cyclical firms. While they not explicitly mention cyclicality in their research, Liu et al. (2002) present mixed results for certain cyclical industries when making use of forward-looking multiples. Firms in the chemicals and construction business show an increase in accuracy while firms in electrical equipment and automotive parts show a decrease when forecasts are used. Tremolizzo (2009) found a decrease for the forward P/E for cyclical companies as well as non-cyclical companies, thereby not providing direct insight on the performance of forward looking multiples for cyclical firms either.

11

found that the latter yielded the best results. The difference between the harmonic mean and simple mean is that the harmonic mean controls for heterogeneity by giving equal weights to each data point while the simple, or arithmetic, mean can be heavily influenced by outliers. Alford (1992), Kaplan and Ruback (1995), Kim and Ritter (1999) and Lie and Lie (2002) and Schreiner and Spremann (2007) all use the median multiple to estimate valuation errors. While on the other hand Bhojraj and Lee (2002), Liu et al. (2002) and Tu (2010) have deployed the harmonic mean instead. Similar to methods for peer group selection no clear preference or best practice seems to exist.

2.3 Cyclicality

The value of a cyclical company is often just as dependent on the movement of a macro variable such as a commodity price or overall economic growth as on its firm specific characteristics (Damodaran, 2009). Since (commodity) prices and economies move in cycles, a firm‟s earnings and cash flows are a directly dependent on which point in time the cycle is. Therefore even mature companies in cyclical industries will face volatility in cash flows. The volatility, thus cyclicality, can be directly tied to susceptibility to macro-economic variables such as commodity prices and economic growth. Boudoukh et al. (1994) assessed cyclicality by linking the industrial production growth rate to the aggregate production growth. They found that industrial production growth of non-cyclical industries have a low correlation with the aggregate production growth rate. Vice versa, for cyclical companies the opposite is true. Typical non-cyclical firms are in industries such as food and beverages or tobacco. Good examples of cyclical industries are manufacture of transport equipment and manufacture of electric equipment.

12

The monetarist view argues that the phenomenon called “cyclicality” is caused by (mistakes in) monetary policy rather than fluctuations in demand (Friedman, 1983) . To a large extent, the rate of consumer consumption and firm level investments are dependent upon macro-economic factors such as interest, inflation and exchange rates. These can be influenced by the central banks by increasing or decreasing the money supply thereby affecting the business cycle.

Instead of aggregate demand or monetary policy, Real business Cycle theory (RBC) suggests that cyclicality is strongly affected by random externalities (Prescott, 1983). For example (relatively) unforeseen events such as the global financial crisis of 2008 and the 11th of September 2001 have had quite an impact on the world economy. Next to the theoretical explanations by Keynes (1937), Friedman (1983) and Prescott (1983) there are many more. While not meant to contrast these three specific methods against each and pick one as best, they are displayed to show the versatile range of factors that can cause cyclicality to occur.

2.4 Peer group selection & cyclicality

As mentioned earlier, when using multiples to value a firm, its value is derived from how comparable firms are valued by the market (Damodaran, 2009; Koller et al. 2010). This implies that in order to arrive at an accurate estimate of firm value the corresponding firms need to be comparable. In an ideal world, all firms in the peer group would share have the same risk, growth and earnings characteristics (Koller et al. 2010). In practice this is unfortunately never the case. Finding comparable firms is ostensibly even more difficult for cyclical firms since we have to deal with the firm‟s cycles. These cycles can be first of all very hard to predict, let alone be exactly comparable to each other (Damodaran, 2009). , let alone be comparable to each other. We will get back to the exact specifics of comparability in the literature review.

13

group selection and the use of forecasted and/or trailing data may have a profound impact on valuation accuracy of cyclical firms.

The way a peer group is selected is important, as Damodaran (2009) points out that cyclicality is also often determined on the basis of industry. The airline and automobile industries are for instance both considered cyclical. Hence, car manufacturers like Volkswagen and Volvo are both regarded as cyclical firms. We run however the risk of tarring all the firms with the same brush, while in fact their line of business may be radically different from each other (Damodaran, 2009). Take Albert Heijn and Hornbach for example. The only thing they have in common is that they are both considered retailers. The products they mainly sell, respectively food and hardware, are however substantially different from each other. A hardware store like Hornbach is more likely to strongly feel the effects of stagnating economy than a food retailer like Albert Heijn. The direct implication of following Alford (1992) by selecting cyclical firms on the basis of industry is that we construct a peer group which may not be really comparable at all.

When constructing multiples, the general consensus is that one is able to arrive at a more accurate estimate of firm value by making use of forecasted data. This is logical, as forward-looking multiples are consistent with idea that a firm‟s value is represented by the value of its future cash flows, a main principle in firm valuation (Koller et al. 2010). The danger in doing so for cyclical firms is however that the forecasts are more likely to be wrong or positively biased. Consensus forecasts for cyclical firms often appear to ignore cyclicality completely and if they do account for it, there is a high chance of making valuation errors by misinterpreting the cycles (de Heer et al. 2000; Damodaran, 2009; Koller et al. 2010).

14

firms may shift to a very different in their cycle than their U.S. counterparts. The essence is that this may render Swedish and American firms less comparable. It is therefore important that analysts accurately predict a firm‟s current place in their cycle and subsequently place it in a peer group with firms with comparable cycles.

3.

Data and Methodology

The firms included in my dataset are retrieved from the Orbis database of Bureau van Dijk. This database contains standardized financial and non-financial data on nearly 150 million companies across the globe8.

The first step is to identify cyclical and non-cyclical industries. The industries are selected on the basis of 2-digit industry codes (Alfrod, 1992; Schreiner and Spremann, 2007). Instead of SIC a code, the Orbis database uses the NACE. Rev. 2 framework. This framework is comparable and similar in function to the SIC codes. Based on industry classifications by Boudouk et al. (1994) and Pearce and Michael (2006) table 3 depicts the cyclical and non-cyclical industries. These subsamples will allow to see if there exist significant differences in valuation accuracy between the two groups.

8

http://www.bvdinfo.com/en-gb/home

9 _{For the sake of simplicity and readability, abbreviations are sometimes used instead of full the industry names.}

This will for example lead to electric utilities instead of electricity, gas, steam, and air conditioning supply. Table 3.9 NACE Rev. 2 code Non-Cyclical Industries N NACE Rev. 2 code Cyclical Industries N 10 Manufacture of food products 94 27 Manufacture of electrical equipment 73 12 Manufacture of tobacco products 10 29 Manufacture of motor vehicles, trailers and

semi-trailers 76 16 Manufacture of wood products, except furniture 22 47

Retail trade, except of motor vehicles and

motorcycles

187

35

Electricity, gas, steam, and air conditioning

supply

179 62

Computer programming, consultancy and related

activities

285

36 Water collection,

treatment and supply 19 63

Information service

15

Data has been collected as follows, due to better availability of financial data for listed companies; the sample consists of listed firms anno 2013 only. This initially yielded a sample of 2991 listed firms from the United States, Canada and the European Union.

Furthermore it is important that all data represents the same point in time. While differences in earnings over time can be minor for stable sectors, they can be large for industries that are characterized by high growth or volatile earnings. Firm differences in fiscal year ends can therefore be problematic. To illustrate this, firms with different fiscal year ends (31/01 vs. 31/03) cannot directly compare their P/E ratios, as both ratios represent prices, earnings and macroeconomic circumstances from a different point in time. If these timing differences are not accounted for, comparing firms with different fiscal-year ends can lead to inaccurate end results (Damodaran, 2009). In order to solve for potential comparability problems, all firm data is adjusted to match the 31st of December as fiscal year end10. Firms without a known fiscal year end are therefore excluded as comparability cannot be guaranteed.

In order to avoid negative valuation multiples, firms with negative enterprise values and earnings data are removed (Liu et al. 2002; Schreiner and Spremann, 2007). The possibility of biased multiples (Damodaran, 2001) due to the omitted data will later be examined by means of the median valuation error (Lie and Lie, 2002). Lastly, firms without non-available data for key variables are excluded as well. The entire selection process resulted in a sample of 960 firms from 27 different countries (see table 4).

Table 4. Top 10 country of origin. 873 out of 960 firms are headquartered in these countries

Country # of firms Country # of firms

1. U.S. 474 6. Sweden 38 2. France 86 7. Italy 37 3. U.K. 75 8. Poland 23 4. Canada 58 9. Finland 15 5. Germany 52 10. Spain 15

10 _{EBITDA for a firm with its fiscal year ending at 31/03/2014 is adjusted in the following way:}

16

3.2 Multiples used

Based upon Lie and Lie 2002, Liu et al. (2002), Schreiner and Spremann, (2007), Imam et al. (2008), Damodaran (2009) and Koller et al. (2010) the following multiples are used for analysis.

1. EV/ EBIT (Enterprise multiple)

2. EV/ EBITDA (Enterprise multiple)

4. EV/ Sales (Enterprise multiple)

3. Price/Earnings (Equity multiple)

The reasons for this are threefold. First of all, these multiples are frequently featured in prior research (See table 1, page 9) which will allow for direct comparison and benchmarking. Secondly EBITDA and the P/E ratio are frequently used by practitioners and hence increase the practical relevance of this research (Imam et al. 2008).

Lastly, as hypothesized by Schreiner and Spremann (2007) and Damodran (2009) the enterprise multiples can be of help to analyze cyclical companies because they are constructed of variables higher up in the income statement and therefore show less volatility than multiples based on net income.

The multiple of a company is calculated by dividing its 2013 market value variable, in this case enterprise value or share price, by a corresponding value driver. Take for instance the EV/EBITDA multiple for a random firm. Dividing its 2013 enterprise value by a trailing value driver (EBITDA 2013) this results in a so called trailing EV/EBITDA multiple. Alternatively, if the firm‟s 2013 enterprise value is divided by a forecasted value driver (EBITDA 2014) this results in a forward-looking EV/EBITDA multiple. How the multiples are exactly constructed can be found in Appendix I.

17

cyclical firms have higher valuation errors and also whether or not the use of forecasted data improves the valuation accuracy of cyclical firms.

The descriptive statistics in table 5 indicate that the calculated multiples follow a non-normal distribution If the assumption of normality is violated we cannot simply use a parametric tests to check for differences across groups (Keller, 2008). If we would like to compare two or more non-normally distributed populations the Kruskal-Wallis test can be applied (Keller, 2008). This is a test that checks whether values are equally distributed or ranked across groups. Table 5 also shows us that there is a substantial decrease in standard deviation over time between the forward and trailing multiples. Therefore we should also expect see differences in valuation accuracy as we progress from trailing to forward-looking multiples.

The Kruskal-Wallis tests are conducted on the aggregated industry multiples. These aggregated industry multiples are computed by picking an appropriate measure of central Table 5. Sample Descriptive Statistics of company multiples. * indicates significance at the 1% level

for the Shapiro-Wilk test. Significant p-values indicate a non-normal distribution.

18

tendency. This can either be the simple mean, median or harmonic mean (not displayed here). Table 5 shows that the values of the median are consistently lower than those of the simple mean. The structurally higher scores of the mean indicate the presence of large positive outliers, which causes an upward bias. The median tends to control for this by picking the „middle‟ observation, and is therefore not „dragged upwards‟. The harmonic mean also accounts for outliers by assigning equal weights to each data point thereby reducing outlier bias. Whether the mean is affected by outliers depends on the sample size and relative amount of outliers. In this case, the median or the harmonic mean as measures of central tendency is highly preferred.

Academic literature also seems to generally accept the median (Alford, 1992; Kaplan and Ruback, 1995; Lie and Lie, 2002; Schreiner and Spremann, 2007) and harmonic mean (Bhojraj and Lee, 2002; Liu et al. 2002; Tu, 2010) as means for firm valuation. By removing the top 3% top and bottom observations (Bhojraj and Lee, 2002; Ditmann and Weiner, 2005) we have already sufficiently controlled for outliers. Therefore the harmonic mean should yield no clear advantage over the median. Hence the latter will be used.

While the Kruskal-Wallis tests will allow us to further inspect whether cyclicality or industry classification is relevant in explaining differences between median multiples. We first apply the test to see whether or not our sample is subject to country effects. Appendix II displays the median multiples for the top 10 countries of origin. The outcome of the Kruskal-Wallis tests implies that there are statistically significant differences across countries. The downside of these tests is however that they do not specifically tell us which countries are significantly different from each other.

19

This is most likely due to the fact that a fairly large portion of our sample firms are located in the U.S. Therefore countries with less sample firms like Spain or Finland are likely to have an insufficient amount of firms for certain industries. As a result, when a national industry is represented by less than 5 firms its multiples need to be discarded. This is because peer groups smaller than 6 to 4 firms are generally regarded as sub-optimal by literature (Bhojraj and Lee, 2002; Schreiner and Spremann, 2007). Instead of improving comparability across peer groups it seems to have actually diminished it. This might very well be due to the influence of variables such as firm size. The presence of a large multinational firm can have extreme influence on smaller peer groups and thereby reduce accuracy. Hence, we disregard the country of origin as a criterion for peer group selection for our research.

Table 6 displays median multiples for the cyclical and non-cyclical subsamples. The cyclical subsample yields slightly lower EV/EBITDA and P/E multiples, but significantly lower EV/EBIT and EV/SALES multiples. Cyclicality seems to be relevant in explaining differences in the median multiples as the Kruskal-Wallis test also indicates that for the latter Table 6. Non-cyclical multiples vs. cyclical multiples The forward-looking multiples are

pooled together in order to limit the table size. * Indicates significance at the 1% percentage level for the Kruskal-Wallis test. This implies that the distributions of the median multiples are not the same across the two cyclical & non-cyclical subsamples.

Trailing EV/EBITDA Forward EV/EBITDA Trailing EV/EBIT Forward EV/EBIT Non-cyclical _{9.01 8.05 13.88* 13.04*} Cyclical _{9.41 8.41 13.00* 11.50*} Pooled 9.29 8.30 13.46 12.16 Test statistic 1,021 0,963 -2,573 -3,477 Kruskal-Wallis p-value 0,307 0,336 0,010* 0,001* Trailing EV/Sales Forward

EV/Sales Trailing P/E Forward P/E

20

two the differences are also statistically significant between the two subsamples. Lastly, Appendix III shows the median EV/SALES multiples for each year of observation per industry. Similar data for the EV/EBITDA, EV/EBIT and P/E multiples can be found in Appendix IV, V and VI respectively.

The P/E and EV/EBITDA multiples display no significant differences between the cyclical and non-cyclical subsample. For the EV/EBIT multiple non-cyclical firms appear to have slightly larger multiples. However, they stand in sharp contrast with the differences found for the EV/SALES multiple. For the non-cyclical sample the median multiples are almost twice as high. This indicates that the enterprise values for these non-cyclical industries are relatively high compared to their revenues. This is confirmed by Appendix VII, that displays the median statistics of the enterprise values and value drivers used to calculate the median multiples. Firms in the tobacco industry and electricity and gas industries have extraordinary large enterprise values. They are large in absolute value but also relatively to their sales figures. For the retail industry it is exactly the opposite. For retail firms the median sales are even slightly larger than their enterprise value. These differences in firm size to sales examples explain why the median EV/SALES multiples is so radically different for the two subsamples.

21

3.3 Estimation of Value

Each median multiple represents the industry multiple at time t. To calculate the valuation errors the industry multiples are multiplied with the denominator at time t which will return the estimated corporate value at time t. For the EBITDA and EBIT multiples this resulted in an estimated enterprise value while the PE multiple yields the value per share of common equity.

The valuation errors (VE) are calculated by applying the same formula as Lie and Lie (2002), As can be seen in the formula below, it takes the natural logarithm of the ratio of the estimated value (EV) and the market value (MV).

Formula: )

The errors obtained from the formula will result in positive valuation errors (overvaluation) and negative valuation errors (undervaluation). The main idea is to establish valuation accuracy not whether a firm is specifically over or undervalued. In order to assess valuation accuracy per industry, the median valuation errors are calculated. If the multiples are unbiased, the median errors would simply lead to zero.

In order to solve for this, prior research (Kaplan and Ruback, 1995; Lie and Lie, 2002; Bhojraj and Lee, 2002) obtained the absolute valuation errors. This is simply done by translating a relative value, either negative or positive, to an absolute value. To illustrate this, two relative valuation errors of -7% and +7% will both lead to two absolute valuation errors of 0.07. By calculating the median absolute valuation error we can examine the valuation accuracy of each industry multiple at time t. Similar to the multiple analysis a Kruskall-Wallis test is deployed in order to test differences across industry distributions.

22

3.3 Regression analysis and error determinants.

If valuation accuracy differs for cyclical and non-cyclical differs we can check this by performing a regression analysis by incorporating a dummy variable for cyclicality. In order to validly test this we also need to control for variables that also possibly influence the valuation errors. Derived from research concerning asset pricing theory and peer group selection (Fama and French, 1992; Bhojraj and Lee, 2002; Liu et al. 2002; Schreiner and Spremann 2007; Koller et al. 2010) this led to the following list of control variables:

1. Size (SIZE) 5. Research & development (R&D)

2. Profit margin (PM) 6. Market-to-book ratio (MB)

3. Leverage as proxy for risk (LEV) 7. Return on assets (ROA)

4. Firm beta as measure of risk (β)

(Information on how these variables are constructed can be found in Appendix I.)

3.3.1 Size

Alford (1992) did not find significant increases in valuation accuracy when he controlled for size, earnings and leverage. However, other authors beg to differ (Lie and Lie, 2002; Bhojraj and Lee, 2002). The idea behind size as a determinant of accuracy is that larger firms are more diversified in terms of location, projects and products. Therefore they are likely to be more stable over time and also exhibit less cyclical behavior. Hence, size is incorporated in the analysis in order to control for this.

Table 7. Median valuation errors. The median serves as an indicator to which

extent the valuation multiples are biased. If unbiased the median valuation error will approximately lead to zero.

EV/EBITDA EV/EBIT EV/SALES P/E

2013 -0.012 0.007 -0.052 -0.003

2014 -0.016 -0.003 0.012 0.003

2015 0.000 -0.021 -0.062 0.006

2016 0.039 -0.017 -0.068 0.009

23

3.3.2 Risk, Profitability and Return on assets

Similar to more complex valuation approaches, multiple valuation is based on the premise that firm value is positively related to a firms expected future earnings and negatively to its risk (Liu et al. 2002). Based on a firm‟s risk and earnings profile, multiple valuation assumes that the market is efficient in setting its prices (Schreiner and Spremann, 2007). In order to account for this, the multiples in this research are constructed on an industrial basis. However, Herman and Richter (2003) and Schreiner and Spremann (2007) argue that there is not a strong theoretical foundation why firms from the same industry should always have similar profitability and risk profiles. Therefore identifying comparable firms based on these explicit measures could improve results (Schreiner and Spremann, 2007).

We can control for this by adding risk and profitability variables in our regression analysis. Leverage is chosen as proxy for risk because capital structure can decrease the reliability of equity multiples. Whereas, the EBITDA and EBIT multiples will be unaffected by capital decisions the P/E ratio is not. Under an efficient market two identical firms with different capital structures will have different P/E ratios. Under the multiple valuation approach we would conclude that firm A is undervalued while firm B is overvalued. This is however incorrect as the market prices the equity of both firm‟s correctly.

24

3.3.3. R&D and Market-to-book ratio.

Strongly related to behavioral finance, firms with a large part of their value in intangible assets or R&D are hard to value as only a small portion of the value can be derived from the assets itself. In those cases it seems high stock prices are possibly based on R&D and the uncertain growth opportunities it represents (Lie and Lie, 2002). Moreover, Lie and Lie (2002) argue that R&D expenses can reduce the current earnings to such an extent that earnings based multiples become a bad predictor of value. While the relationships are not entirely made clear in prior research, the bottom line is that R&D can however lead to misvaluation. Hence we incorporate R&D as a control variable and we expect R&D to negatively influence valuation accuracy.

R&D driven firm value is closely connected to market-to-book ratio, the small portion of the value that can be derived from the assets itself will automatically lead to high market-to-book ratio‟s. In a survey conducted Wall street professionals, Bloomfield and Michaely (2004) discovered that investment bankers found MB a strong indicator of mispricing and risk. A high market-to-book ratio was often seen as a sign of overvaluation while lower ratio‟s as a sign of the opposite. The survey outcome generally confirms the empirical findings of Fama and French (1992) who established a significant conjunction between firm size, market-to-book ratio‟s and security returns for a non-financial sample. Combining the results of these studies led us to believe there might be a relation between R&D, market-to-book ratios and valuation accuracy. Overly optimistic or pessimistic markets can drive the stock prices further away from their fundamental values and multiple valuation does not take these effects into account.

3.3.4. Dummy variables

25

3.4 Model specification

Including all the variables from section 3.4 in our regression this leads to the following equation:

This regression model is estimated using the ordinary least squares (OLS) technique. Because all the multiples are estimated based on the enterprise value anno 2013 this results in 5 annual cross-sections for each multiple. The EV/Sales multiple has been dropped, due to reasons that will be discussed later. All in all, this leads to 15 cross-sections in total. Moreover, it has to be noted that VE, represents the relative valuation error. The reason behind this is that the variables will either lead to overvaluation (positive sign) or undervaluation (negative sign), something we cannot determine when using the absolute errors.

3.5 Diagnostic tests

The OLS estimation technique has a number of assumptions in order to be validly conducted (Brooks, 2008). If the model does not control for these assumptions the user may yield invalid conclusions (Brooks, 2008). The first assumption is that the average value of the errors is zero. However, if a constant is included in the regression – which is - this assumption will never be violated (Brooks, 2008).

The second assumption for OLS is homoskedasticity. This assumes that the error terms have the same variance. If variance among the errors is different, they are said to be heteroskedastic. With cross-sectional heteroscedasticity OLS is still consistent but potentially suboptimal for drawing inferences (Brooks, 2008). For the above equation the White‟s test for heteroskedasticity is significant, thereby implying hetereskedastic residual variance for the sample. The White‟s cross-section robust standard error option in Eviews is used to control for this (White, 1980; Brooks, 2008).

26

various independent variables. The highest correlation of 0.29 between beta and size is well below the 0.70 threshold. The others correlations are sufficiently small and can reasonably be ignored. As mentioned earlier, in order to avoid the problem of perfect multicollinearity by including dummy variables, at least one dummy variable per category needs to excluded (Brooks, 2008).

4. Results

The results relating to two main research questions will be discussed first. Thus first, do cyclical industries have higher valuation errors than non-cyclical industries? And secondly, do forward-looking multiples increase the valuation accuracy for cyclical firms? Thereafter, we will discuss additional findings such as the overall performance per multiple, and possible differences across industries etcetera. Lastly the results of the regression analysis will be discussed.

4.1 Valuation Accuracy

Table 8 on the next page displays our main results. It shows the median absolute valuation errors for the cyclical and non-cyclical subsamples. The results are largely in line with our initial expectations. The Kruskal-Wallis tests show, that on average the median valuation error is substantially lower for the non-cyclical subsample for the EV/EBITDA, EV/EBIT and P/E multiples.

27

Moreover, due to the extended time horizon, it becomes more of a level playing field, as the chance of inaccurate forecasts now also becomes more likely for non-cyclical firms as well. This seems supported by the significance levels that decrease over time, as we finally find insignificant results for both subsamples in 2017. This implies that the two distributions are no longer significantly different from each other. In regard to whether or not the use of forward-looking data significantly increases the accuracy we find interesting results.

Table 8. Kruskal-Wallis tests of absolute valuation errors for the cyclical and non-cyclical subsamples.

* or ** indicate significance at the 1% or 5% percentage level for the non-parametric Kruskal-Wallis test. Significant p-values reject the null hypothesis that distributions are the same across industries.

28

Table 8 shows that the accuracy of the EV/EBITDA multiple increases over time for both subsamples. The EV/EBITDA multiple does not seem to be affected by misinterpretation of cycles or overly optimistic forecasts. Table 9 shows tests results of forward vs. trailing multiples. It shows that as well for the cyclical as the non-cyclical subsample, the increase in accuracy for the EV/EBITDA multiple is deemed significant. Hence, for the EV/EBITDA multiple we can positively answer the second research question by stating that the use of forecasted data does increase the accuracy for cyclical firms.

The only side note here however is that accuracy decreases as we move from 2-year forward-looking multiples towards 3-and 4-year forward forward-looking multiples. This is also most likely attributable to diminishing number of observations and possible inaccurate forecasts.

For the EV/EBIT multiple, the 1-year forward looking multiple brings a slight decline in accuracy for both samples. The exact reason why is unknown. Moreover, the seemingly improved performance if we progress to later year forecasts do not hold for the EV/EBIT multiple. The recorded valuation errors for the cyclical subsample show a decline in error but are deemed not significant as we move from trailing towards forward-looking data. Hence, we

Table 9. Forward vs. Trailing multiples. Results of the Kruskal-Wallis tests for the

forward & trailing multiples.*,* ,*** indicate 1% ,5% and 10% level significance.

Cyclical Firms

Test

statistic Sig. Non Cyclical

29

cannot say that the forward-looking multiples produce more accurate results. The interesting thing is however that the opposite is true for the non-cyclical subsample, where the tests indicate that the use of forward-looking does lead to improved accuracy. Therefore we negatively answer the second research question for the EV/EBIT multiple as the seemingly increased accuracy of forward-looking data is statistically not supported. With respect to the first research question we do see similar results to the EV/EBITDA multiple. The difference between the cyclical and non-cyclical subsamples for the trailing multiple amounts to 8.37% in favor of the latter. Looking at the forward-looking data the differences also decrease as we shift towards later-year forecasts. But as mentioned before we cannot derive statistically significant conclusions from them.

As can be seen from its error values the EV/Sales multiple (Appendix XI) is somewhat a unique case. The extremely high valuation errors are seemingly subject to how a multiple is constructed. Therefore we skip this multiple in discussing the differences between cyclical and non-cyclical firms as we cannot draw sensible conclusions based on its error values. The recorded valuation errors for EV/Sales will be discussed later when the overall accuracy per multiple is assessed. Up till then, its values should be discarded.

Lastly, non-cyclical firms also outperform the cyclical firms for the P/E ratio although by a much slighter margin. For the trailing P/E multiple and 1 & 2-year forecasts the difference is approximately 4 to 5%. The most substantial difference is observed for the 3-year-forward looking multiple, with a difference of 7.67%. According to Koller et al. (2010) high P/E ratios represent high growth prospects or earnings. This means that a firm has high market capitalization as opposed to earnings or vice versa. For both subsamples the decrease in median multiples for each forecasted year led to more accurate results. This might be due to the fact that an insufficient amount of estimates were available for high growth prospect industries such as IT (Koller et al. 2010). Along with the P/E, the number of observations significantly drops over time for firms in the computer programming industry. Resulting in less inflated P/E‟s due to the presence of (unrealistically) high multiples for growth firms.

30

EV/EBIT multiple. The test results indicate that the minor improvements in accuracy over time are insignificant. Similar to the EV/EBIT the opposite applies to the non-cyclical subsample with statistically significant improvements in accuracy for forward-looking multiples

To sum up, in general the results seem to be affirmative for the first research question. For the second research question only the test results for the EV/EBITDA multiple were significant. The reason as to why remains vague. A plausible explanation is that both EV/SALES and EV/EBITDA, who both have significant test results, are the furthest up in the income statement. While EBIT might for example be subject to write-offs and P/E to changes in earnings, EBITDA and SALES are easier to predict, also in times of economic distress (Schreiner and Spremann, 2007). However, this is only a mere guess, the exact reasons remain unknown.

Due the low amount of research conducted on cyclical firms and multiple valuation, there are only two studies which allow for partial benchmarking of our findings. Tremolizzo (2009) investigated the accuracy of relative valuation for cyclical firms. Similar to this study, he found that cyclicality generally leads to a decrease in estimation accuracy for EV/EBITDA, EV/EBIT and P/E multiples. Yet besides a 1-year forward P/E, his research was performed by only analyzing trailing multiples. Similar to him, we also found a decrease for the 1-year forward-looking P/E. .

The second study that allows for partial benchmarking is research by Liu et al. (2002) who incorporate a tremendous amount of industries in their study. The industries are ranked according to pricing errors with the highest observed rank being 17 and the lowest 1. Here, lower rankings equal lower average pricing errors. While the industry descriptions do not exactly match nor do they explicitly classify industries as cyclical, the results can still be seen as a confirmatory indicator for our research question, as the cyclical Automotive (16), Retail (12), Electrical Machinery (12) and Computer industries (12) are ranked incredibly high in comparison to non-cyclical industries such as Gas (1) and Water (1) utilities.

31

observations. Firms in IT intensive sectors seem to yield higher valuation errors on average firms from other industries. However, this is with two cautionary notes. For EV/EBITDA for instance, the information services industry actually have significantly lower valuation errors than other industries. The difference between the two IT oriented industries may very well be explained by the different industry size. The industry size of 15 falls in sharp contrast with the industry size of 250 for computer programming. Similar to information services, the tobacco and water sectors have small sample sizes and more or less similar errors as well. These smaller industries probably form more homogenous industry group which increase comparability and result in representative multiples. That a small industry size does not necessarily has to lead to a more homogenous group is proven by the manufacture of wood industries. It scores consistently higher errors than most industries despite its small sample size. Closer inspection of this industry reveals that some company multiples are not accounted for by the current outlier method. While not necessarily faulty, they do skew the results. Indicating that selection on the basis of size like Tu (2010) might prove worthwhile.

The second note concerns the fact that the Kruskall-Wallis test indicates differences in distribution on an industry basis for the whole sample. In order to discover which industry distributions are specifically different from each other, post hoc tests can be performed.11 Hence Dunn (1964) post hoc tests with Bonferroni corrections are applied. The Bonferroni correction adjusts the significance level order to avoid Type I errors. A Type I error occurs when you declare a result statistically significant when it is not. The chance of a Type I error increases with every additional pairwise comparison (Dunn, 1961)12. A well-known problem for two-stage mean comparisons is that positive Kruskal-Wallis tests do not necessarily lead to significant pair wise comparisons (Gabriel, 1969). The Kruskal-Wallis test is technically a comparison of the mean or median ranks.

For all the pairwise comparisons this leads to 15 instances for which the distribution between Computer programming were significantly different (Appendix XVI). While the post hoc tests merely taps into the distribution of the medians and not the actual values itself, Appendix VII shows that the industries indeed differ quite a lot in terms of their median statistics for enterprise value and value drivers. The high P/E ratio‟s combined with relatively low earnings per share for 2013 indicate other measures such as the growth or intangibles must incorporated in order to arrive at such a high multiple.

32

While not the main focus of this paper, these results are still worth mentioning and perhaps provide a stepping stone for feature research.

Now that the two main research questions have been answered it is time to discuss the performance of each of the chosen multiples. Let us first examine what seems to be the ugly duckling of the four, the EV/Sales multiple. In line with Liu et al. (2002), the EV/Sales multiples are the worst performers. Normally this would suggest that sales do not properly reflect profitability until expenses have been deducted. However, as mentioned before, something fishy seems to be going on here. The size of the absolute valuation errors is so substantial, that the construction of the multiple seems to be flawed all together. In Appendix XI, firms in the tobacco industry and electricity and gas industries present extraordinary large enterprise values relatively to their sales figures.

It seems the EV/Sales multiples are considerably affected by large firms in terms of size relatively to their sales. If the number of observations per industry decreases the errors actually increase, which are subsequently not controlled for by correcting for outliers. This is contrary to the idea that narrowing down classification codes, thus creating narrower samples, can increase comparability due to increased homogeneity (Schreiner and Spremann, 2007). If we contrast the errors of the tobacco, water utilities and information service industries against those of others it can be said that fewer observations per industry can also actually increase the valuation error.

The large valuation errors are not unique in literature. Bhojraj and Lee (2002) also report an absolute error of 55.0% when selecting peers on the basis of industry and size. However, when their systematic multiple is used the error significantly decreases to 35.0 %. Liu et al. (2002) also find valuation errors of around 70.0% for one fifth of their sample, but their observed median values (35.0%) are also significantly lower. The different peer selection method and lower absolute valuation errors of both studies indicate that the peer selection solely on the basis of industry might not be suitable for EV/Sales multiples in the first place.

33

the opposite. The differences between Schreiner and Spremann (2007) and our results can very well be due to (random) differences in sample composition. For instance, some industries may have a preference for certain multiples. As Lie and Lie (2002) point out, firms with highly liquid assets such as financial institutions tend to prefer P/E multiples for example. These findings are supported by Imam et al. (2008) who state that analysts also share this preference.

Furthermore, similar to Dittman and Weiner (2005), we again observe country differences in the valuation errors (Table XIII). Which countries yield significantly higher or lower errors will be discussed in the results of the regression analysis. Lastly, similar to Liu et al. (2002), Schreiner and Spremann, (2007) and Tu, 2010 our results show a consistent decline in absolute valuation errors for the non-cyclical subsample when later-year forecasts are used instead of earlier-year forecasts. Ostensibly unique to this study is that inclusion of 4-year forward looking data. However, the 4-year forward-looking multiples consistently increase the valuation errors. Hence this might be the reason that so far only 3-year our forecasts have been used so far.

4.5 Regression Analysis

In this section we will discuss the results of the regression analysis. The analysis allows us to triangulate whether or not cyclicality significantly influences the valuation errors. Moreover, we can also establish which specific countries and industries play a role too. The variables that have been defined by prior literature will be used as control variables. Due to the extremely inaccurate valuation results the EV/Sales multiple is not included. The regression coefficients of the remaining three multiples can be found in table 10 on the next page.

34

Table 10. OLS Regression coefficients. *, ** and *** indicate significance at the 1% , 5% and 10% level

respectively. * for the R2 Statistic indicates a significant value of the F-test which indicates whether the model has significant predictive capability.

Panel A

EV/EBITDA Intercept Cyclical PM ROA SIZE MB LEV Beta R&D Adj. R

2 2013 0.359* -0.199** 0.052 0.000 -0.016 0.000* 0.000 0.095 0.044* 0.154* 2014 -0.188 -0.123** 0.257 0.000 -0.001 0.000* 0.000 0.057 0.037* 0.123* 2015 -0.164 -0.122** 0.313 0.000 0.008 0.000* 0.003* 0.053 0.034* 0.119* 2016 -0.288 -0.112 0.397 0.001 0.029 0.000* 0.030** -0.043 0.031* 0.096* 2017 0.061 -0.157 0.252 0.001** 0.012 0.000** 0.006 -0.167 0.022* 0.08** Panel B

EV/EBIT Intercept Cyclical PM ROA SIZE MB LEV Beta R&D Adj. R

2 2013 0.368 -0.301* -0.263* -0.001 -0.017 0.000** 0.000 0.116 0.047* 0.099* 2014 0.018 -0.257* -0.260* 0.000 0.014 0.000* 0.000 0.083 0.038* 0.091* 2015 -0.134 -0.238* -0.408** 0.000 0.032** 0.000* 0.003 0.032 0.036* 0.114* 2016 -0.384 -0.137 -0.090 0.000 0.056** 0.000* -0.031 -0.070 0.030* 0.119* 2017 0.242 -0.231* -0.426 0.000 0.010 0.000 -0.011 -0.149 0.023* 0.067** Panel C

P/E Intercept Cyclical PM ROA SIZE MB LEV Beta R&D Adj. R

35

Size is not proving to t have much significant explanatory power for the valuation errors. It is hard to derive any definitive conclusions, as the sign‟s indicate that size is generally neutral towards over or undervaluation. This is possibly in line with Alford (1992) and Cheng and McNamara (2000) who found that valuation accuracy increased with size, i.e. implying the indifferent signs of approximately 0. Size only yields explanatory power for the 3-and 4-year forward-looking EV/EBIT multiples, where it seemingly leads to overvaluation. This could be an indication of less risk that is perceived with large firms. Because as size increases firms become more diversified in terms of projects, products and locations and this should in general lead to more stable cash flows or earnings (Lie and Lie, 2002). The significance for size holds for all multiples up till 2015. As more forward-looking data is used the overall predictability of the models is compromised as indicated by in the adjusted R2‟s as well. This is an observation that applies to virtually all multiples.

In accordance with Bhojraj and Lee (2002) R&D also yields explanatory power. As they point out, higher R&D expenditures may lead to understated current profitability relative to possible future profits. Here, R&D expenditures seem to lead to a consistent degree of overvaluation. The possible explanation for this might lie in behavioral finance, as Lev et al. (2005) found that investors tend to fixate on reported R&D measures. They detected undervaluation of conservatively reporting firms and overvaluation of aggressively reporting firms. This evidence is consistent with the phenomena of cognitive bias in behavioral finance. The influence of R&D in our sample might very well be due to the large portion of R&D intensive firms in the cyclical subsamples, such as the automobile and IT industries.13

A similar line of reasing could be applied to the profit margin. However, the profit margin mainly stands out by the absolute size of its coefficient. Moreover, for our sample it leads to significant undervaluation. This is exactly contrary to what one would expect, as higher profits should generally lead to a more optimistic firm view. With that said the coefficients for the EV/EBITDA multiple seem to be more appropriate. These are however not statistically significant. The exact reason for these outcomes is hard to tell. The reason why profit margin as a variable is kept in de model is because it increased the overall explanatory power.