The Relation between Labor Market Conditions and the
Predictability of the CSV. A Cross-Country Analysis.
DYLAN DE MOTS
Master Thesis
MSc Finance Asset Management
Thesis supervisor: E.Eiling
University of Amsterdam, Amsterdam Business School, Amsterdam, the Netherlands.
Abstract
Recent literature shows that a proxy for sectoral labor reallocation (CSV) has very strong significant
power for predicting US stock market returns. This paper contributes to the existing literature by
testing the predictive power of the CSV across countries other than the United States. More
importantly, we shed a light on the predictive power of the CSV under different labor market
conditions. We conclude that the predictive power of the CSV varies across countries with different
labor market conditions. Specifically, the in-sample predictive power seems to disappear for countries
with high industry concentrates and high unemployment rates. Also the out-of-sample predictive
power of the CSV seems to decrease with rising levels of industry concentration. Further we find the
predictive power of the CSV to be significant for countries with high net migration rates. Our
sub-analysis shows that CSV is the best proxy for sectoral labor reallocation as well as a good stock return
predictor for Canada and the United States.
Keywords: Sectoral labor reallocation, stock return predictability, labor market conditions
JEL Classification: G12, G17, J01
Statement of Originality
This document is written by Student Dylan de Mots who declares to take full responsibility for the
contents of this document.
I declare that the text and the work presented in this document are original and that no sources other
than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the
work, not for the contents.
1Table of Contents
1.
Introduction ... 4
2.
Literature ... 6
3.
Data ... 10
3.1
CSV ... 10
3.2
Data description ... 11
4.
Methodology ... 14
4.1
In-sample and out-of-sample: A cross-country analysis ... 15
4.2
Quintile regression analysis ... 16
4.3
Time series predictive power ... 16
4.4
CSV and unemployment growth ... 17
5.
Results ... 17
5.1
In-sample and out-of-sample: A cross-country analysis ... 18
5.2
Quintile regression results ... 22
5.3
Time series predictive power ... 23
5.4
CSV and unemployment growth ... 26
6.
Robustness checks ... 30
7.
Conclusion and discussion ... 32
8.
References ... 36
9.
Appendix ... 38
1.
Introduction
Recent literature finds a proxy for sectoral labor reallocation (CSV) to be a very strong predictor with
respect to stock market returns. Moreover, Eiling et al. (2016) find that sectoral labor reallocation
outperforms any other predictive variable based on US data. This paper contributes to the existing
literature by extending the research of Eiling et al. (2016) to other countries in the world. With every
country facing its own economic conditions, this study will open a new dimension in the literature by
combining different labor market conditions and the predictability of the CSV on stock returns. This
research should be of interest for academic research and investors since stock return predictability is a
widely investigated topic in field of asset pricing. Since this proxy for sectoral labor reallocation
seems to outperform any other predictive variable; it is appealing to look at predictive power of the
CSV for other countries with different labor market conditions.
Popular factors in studies conducted on stock return predictability are mainly distinguishable
in macroeconomic variables and financial variables. Examples of studies done on stock return
predictability using macroeconomic variables are among others Campbell & Vuolteenaho (2004) and
Lettau & Ludvigson (2001). These studies use inflation rate and consumption-wealth ratio respectively
as explanatory variables. Examples of financial variables are dividend-price ratios (among others
Fama & French, 1988), price-earnings ratios (Campbell & Shiller, 1988) and the net payout yield
(Boudoukh et al., 2007). Although some papers find slightly contradicting results about size of the
predictive power of a variable, they all argue that a small fraction of the stock market return can be
explained by other variables. Fama & French (1988, p. 4) state: “The common conclusion, usually
from tests on monthly data, is that the predictable component of returns, or equivalently, the variation
through time of expected returns, is a small fraction (usually less than 3%) of return variances.”
More recent studies focus on a new area when it comes to stock return predictability. Next to
macroeconomic- and financial variables studies show that human capital is a very strong predictor for
stock market returns. In particular, sectoral labor reallocation seems to outperform any other variable
when it comes to stock return predictability (Eiling et al., 2016). Sectoral labor reallocation measures
the need for labor reallocation between different industries. In reality, cross-industry labor
reallocations do not work out frictionless. Workers often have to acquire new skills or make other
adjustment costs which are very time consuming. There are adjustment costs involved for the
employers as well which depress stock return growth. The need for sectoral labor reallocation is the
consequence of the demand for labor being greater in one industry relative to other industries. This is
often the case when one industry starts outperforming another industry which distorts the labor market
equilibrium at that point. Based on this theory, studies constructed a variable which measures the
dispersion in cross-sectional industry stock returns [Brainard & Cutler (1993) and Eiling et al. (2016)].
Rather than having the same trend as the stock market returns, sectoral labor reallocation has the
power to predict future stock market returns. The intuition behind the predictability lies in the labor
adjustment costs. Since labor adjustment costs are time consuming, the need for sectoral labor
reallocation is not directly fulfilled after the industry demand shock. Thereby, the need for sectoral
labor reallocation continues to have impact on firm value a few months after the industry demand
stock.
To generalize the results of Eiling et al. (2016) to data samples other than the United States,
we need to take into consideration that economic characteristics differ among each other. Studies show
that labor market conditions have impact on labor market efficiency and thereby on predictability of
stock returns. In this research we hypothesize the impact of different labor market conditions across
countries on the predictive power of the CSV. We analyze countries based on their demand for skilled
labor, the role of industry concentration, the role of international migration and the effect of bad labor
market conditions due to high unemployment rates. Our sample consists of 26 countries in the sample
period from January 1973 to December 2017. The countries are selected based on the data available in
Thomson Reuters DataStream Database. The CSV variable is constructed based on 10 industry returns
as available in DataStream. This contrasts Eiling et al. (2016) where the CSV measure is constructed
from 49 industry returns as available on the Kenneth French's website. Although our CSV measure is
based on fewer industries, we show it is still able to significantly predict future stock market returns.
This CSV measure gives us the opportunity to compare different countries and economies with each
other. We perform a cross-country analysis, followed by a quintile regression analysis where we sort
the countries according to the labor market conditions in the sample period. Next, we check for
changes in time series predictive power for different labor market conditions. This analysis is
performed using panel data as well as data at the country level. At last we perform a sub-analysis
where we examine the predictive power of the CSV with respect to unemployment growth. All in all
our research can be summarized by the research question: How is the Stock Return Predictability of
the CSV related to Labor Market Conditions? A Cross-Country Analysis.
The results show that CSV significantly predicts future stock market returns in countries with
low unemployment rates, while it does not in countries with high unemployment rates. Due to the
increased willingness to work in countries with high unemployment rates the effect of sectoral labor
reallocation weakens. Also, the predictive power of the CSV is declining with the level of industry
concentration of a country (in-sample as well as out-of-sample). Further we find that increased search
costs due to high net migration rates increase the need for sectoral labor reallocation, therefore CSV
significantly predicts future stock market returns in countries with high levels of international
migration. We confirm the findings of Eiling et al. (2016) that CSV is a good predictor with respect to
future stock market returns in the United States. However, our results suggest that the predictive
power of the CSV varies a lot for different countries with different labor market conditions.
The remainder of this paper is organized as follows: Section 2 presents an overview of
previous studies conducted in the field of stock return predictability and sectoral labor reallocation.
Section 3 provides an overview of the data used. Section 4 explains the methodology used to conduct
the research. Afterwards, in Section 5, the results of the study are presented. In section 6 we perform
some robustness checks. At last we conclude the findings, discuss limitations of the paper and bring
forward suggestions for further research.
2.
Literature
Numerous studies have been conducted on stock market return predictability and the link to the labor
market over the last years. Chen & Zhang (2011) show that labor market variables as payroll growth
and net job creation rates in manufacturing have significant predictive power when it comes to future
stock market returns. Another labor market related variable is the labor reallocation shocks between
industries. A widely used proxy for labor reallocation shocks is the dispersion in cross-sectional
industry stock returns as argued by Brainard & Cutler (1993). High industry cross-sectional volatility
leads to a need for sectoral labor reallocation among the workers. Ideally workers should move
frictionless from the underperforming industry to the outperforming industry. Belo et al. (2014)
suggest that frictions in the labor market can significantly affect stock market returns via asset prices.
High reallocation costs will be part of the firm value and therefore will generate lower future returns.
Next to impeding stock return growth, Lilien (1982) states that sectoral labor reallocation
leads to future aggregate unemployment growth. Workers fired in underperforming industries are not
directly re-hired in the outperforming industries. This leads to – temporally – unemployment. Lilien
(1982) finds that this cause of unemployment was the main driver for increasing unemployment rates
in the United States during the seventies. This relationship between sectoral labor reallocation and the
rise of unemployment growth is also known as the sectoral shifts hypothesis. Prakash (1986) extends
the research of Lilien (1982) by accounting for the impact of oil shocks. This research shows that
dispersion of sectoral labor reallocation is mainly driven by cross-industry impact of oil shocks.
Accounting for these oil shocks, the dispersion in labor reallocation does no longer predict
unemployment growth. A more recent study building on Lilien (1982) is done by Fortin & Araar
(2010) who investigated the impact of sectoral shifts on the labor market in Canada. Their results
show that unemployment changes are mainly driven by monetary actions and export of Canadian
goods and services. However, they also state that dispersion across industries is a potential driver for
short-term unemployment changes.
A main driver of sectoral labor reallocation is the development of technology. A study
conducted on the labor market in Hungary is done by Kézdi (2002). Due to rapid increases in
technology, low-skilled jobs start to disappear while on the other hand the demand for high skilled
labor rises. This causes a change from inter-sectoral reallocation between low-skilled jobs to sectoral
labor reallocation to industries with a high demand for expertise. Since experts are limited, their wages
rise due to the ‘skill premium’ they are able to negotiate. Another consequence is the increasing
rigidity in the short run supply of workers. Low-skilled workers need to acquire new skills and other
adjustment costs which are costly and time consuming.
So far we discussed labor mobility in the light of sectoral labor reallocation. Another potential
factor hindering labor mobility could be geographical mobility. Geographical labor mobility reflects to
what extend workers are hindered by geographical factors to relocate to find the perfect job. Country
borders, legislation and low market integration are examples of factors reducing the mobility of labor.
Individual countries and their economies can differ significantly from one another. Factors such as
climate, culture and technological development determines the main industries in a country. Hence,
reallocating between different industries may go hand in hand with relocating between geographical
areas.
Ahn et al. (1999) study the willingness to move for work and unemployment duration in
Spain. They find the willingness of workers to move to be increased when the unemployment benefits
are low and when they are in need of monetary incentives. Another study of Anh et al. (1999)
investigates the relation between human capital resources and area characteristics to international
migration in Vietnam. They show that migration plays a role in relocating human capital to
labor-scarce areas.
Rosenbloom (1990) performed research on the labor market integration in the United States.
Altough the sample used is already some time ago (1870-1898), the results are still meaningful for
present-day research. He finds that most of the markets in the United states were nationalizing, except
for the labor market. The labor market faced huge differences in wages across different states. These
wage differentials caused a lack in labor mobility and therefore labor demand growth. This study
concludes that a lack of labor market integration impedes (sectoral) labor reallocation. Johansson et al.
(2002) find increased time distances between municipalities to diminish the extent of labor market
integration. Diminished time distances result in an overall labor market size increase because of
integrating markets. These findings are in line with Rosenbloom (1990).
2.1
Hypothesis derivation
Inspired by the findings of Ahn et al. (1999) we examine countries where labor market conditions are
very bad in general due to high unemployment rates. As Ahn et al. (1999) argue, people are more
flexible when it comes to changing from job when they are are in need of monetary incentives and/or
when unemployment benefits are low. In other words, they are more willing to move, study or make
other adjustment costs to get a better job. When the workers are willing to make adjustments,
employers have to incurr less adjustment- and reallocation costs when hiring new personnel. Hiring of
new workers will have lower impact on firms costs and therefore lower impact on firms' future returns.
Therefore we expect to observe less severe impact on market returns due to sectoral labor reallocation.
This leads the first hypothesis:
H1: The CSV has stronger predictive power in countries with low unemployment.
As already mentioned, Eiling et al. (2016) conducted research on the predictive power of the CSV in
the US. However, not all countries in the world have such a developed and diversified economy as the
US. Many countries in the world are heavily reliant on one important sector. For instance, the
economies of Russia and Malaysia are mainly driven by the oil industry. When many residents of a
specific country work in the same industry, the demand and supply of labor should also be very high
in that particular sector. Therefore, we expect less reallocation- and adjustment costs in those countries
when the demand for labor increases. This results in a reduced impact on market returns. All in all this
leads to the second hypothesis where we expect reduced predictive power of the CSV in countries with
higher industry concentration rates.
H2: The CSV has less predictive power in countries with higher industry concentration.
Similar to the argumentation that not all countries’ economies are so diversified as the economy of the
United States, not all countries are as developed as the United States. A highly developed country goes
hand in hand with increased demand for technological skills and narrow – but in depth -expertise. Jobs
in other sectors such as manufacturing are often less complicated and less specialized, which makes it
easier to adjust when a worker moves to this sector. We expect that industry specific shocks have a
bigger impact on stock returns in countries which have a greater demand for skilled labor. As proxy
for skilled labor demand we use the market value of the technological sector as fraction of the total
industry market value.
H3: The CSV has stronger predictive power in countries with greater demand for skilled
labor.
Ahn et al. (1999) show that people relocate to find work. People will find it easier to relocate between
countries with highly integrated labor markets. When markets are highly integrated, countries will be
able to attract (skilled) labor more easily. Hence, a country with high immigration rates is valued by
workers as an attractive country to work in. On the other hand, highly integrated markets lead to
increased search costs for firms. A larger pool of worker available takes more time, energy and money
to screen and find the perfect candidate to fulfill the vacancy. Thus, the effect of international
migration can be argued in both directions. This raises the question; does international migration play
a role here? If so, in which direction? As separating variable we will use the net migration rates which
measure the difference between immigrants and emigrants per 1000 people.
H4: International immigration has an impact on the predictive power of the CSV.
3.
Data
3.1
CSV
The variable of interest in this research is the cross-sectional volatility of industry specific returns. The
CSV measures the differences in industry returns. In other words, it measures the differences in
demand for labor across industries. When there is high cross-sectional volatility within a country,
labor needs to be reallocated over the industries. Outperforming industries are in need of more labor,
while on the other hand underperforming industries need to fire workers. As consequence workers
from underperforming industries need to reallocate to outperforming industries. This is also known as
sectoral labor reallocation. In efficient markets workers should be able to move frictionless between
industries. In practice however, there are barriers which impedes workers to move freely such as
distance to work, lack of knowledge or legislation. To overcome these barriers firms and workers need
to make adjustment costs before vacancies can be fulfilled. These adjustment costs lower firm value
and thereby the potential future returns. In this sense, an increased CSV is related with lower stock
market returns.
The CSV variable is, among others, proposed by Brainard & Cutler (1993). We follow their
methology by constructing the CSV measure based on industry-specific returns instead of total
industry returns. By doing so, we do not take into account aggregate shocks which do not lead to
sectoral reallocation of labor. The CSV variable in this study is based on the industry returns of the 10
industries as available in the DataStream database (expressed in US dollars). We run a regression with
the excess industry returns for each industry i in country j as dependent variable to construct the
industry-specific returns. The independent variable is the excess market return of the past 36 months
of country j.
Using the estimated constant (
and the fitted residuals
obtained from regression (1),
we construct the industry-specific returns
):
From this point onwards we deviate from Brainard & Cutler (1993). While they construct the
cross-sectional standard deviation based on employment weights, we follow Eiling et al. (2016) using an
equal weight approach across the 10 industries. This allows us to put our findings in perspective with
the findings of Eiling et al. (2016).
Finally, we construct the CSV as the cross-sectional standard deviation of industry-specific
returns. Mathematically summarized as follows:
where:
3.2
Data description
Thomson Reuters DataStream provides all information with respect to CSV construction, labor market
variables and other predictive variables. Data retrieved from other databases will be explicitly
mentioned. Since we need to have a sufficient amount of data to perform the out-of-sample analyses,
we use a sample period of 44 years from January 1973 to December 2017. However, not all data is
available for all countries in this time period. For each country we retrieved the maximum range of
data available within this sample period. Since we find ourselves somewhat limited in data for other
countries than the United States, we select the sample of 26 countries based on data availability. The
full sample of countries is listed in Appendix A, Table 1. All data is expressed in US dollars, which
allows us to compare results across countries. Further we compute the monthly industry returns as the
percentage difference of the industry total return index. For the monthly market returns we use the
total return index of the main index for each country. Both the excess industry- and market returns are
constructed by subtracting the risk free rate. As proxy for the risk free rate we use the one month
T-Bill rate from the Kenneth French's website. We download the market value of each industry per
country to construct the proxy for industry concentration. The industry concentration variable is
constructed of the percentage of each industry’s market value compared to the total market value of all
industries. Afterwards we square all the percentages and sum them according to the
Herfindahl-Hirschman Index (HHI) as proposed by Shugart (2008). Squaring all percentages weighs larger
industries more heavily. Using the HHI also allows us to observe the relative sizes of the industries.
As proxy for demand for skilled labor we use the relative market value of technology. This variable
consists of the market value of technology in fraction of the total market capitalization. For the
benchmark predictors we us the monthly inflation rates based on the CPI index of each country and
the logarithm of the dividend price ratio. We choose for these predictors because these data are widely
available across all countries in our sample as opposed to other predictive variables. Further we
download the monthly unemployment rates from DataStream as well. The net migration rates are
retrieved from the database of the United Nations Statistics Division.
Table 1 shows the descriptive statistics for the predictive variables and the labor market
conditions. The CSV varies between 0.001 and 6.48 with an average of 0.22. The average monthly
excess market return is 0.76% and with a maximum of 79.76%. The huge crash of -63.42%
corresponds to the market return in Russia in September 1998, as consequence of the Russian financial
crisis. The maximum excess market return corresponds to March 1973 in Hong Kong. In this month
the Urban Council selection was held. The relative market value of technology compared to the total
market value of all industries varies between zero and 78.47%. Where the zero means that there was
no market for technology at that time in that country. In the late 70's Hong Kong had the highest net
migration rate with almost 18 migrants per thousand population per year. The net migration rate of
Greece was the lowest of our sample in the period 2010-2015 with -2.86. Furthermore, the average
unemployment rate is almost 7 percent and varies between 0.90% and 27.96%. The latter value
corresponds once again to Greece (July 2013). The inflation rate reports the yearly inflation rate on a
monthly basis with an inflation rate peaking in 1994 in Brazil (the end of the hyperinflation period).
All variables roughly contain between 9,000 and 11,000 observations, leaving us with enough data to
work with. The only exception is unemployment rate data, which contains nearly 8,000 observations.
Unemployment data is not available for all years and countries, especially the periods which are a
longer time ago.
Appendix A, Table 1: Panel A, shows the mean values of the CSV variable and the different
labor market variables per country. The labor market conditions differ substantially among countries
in the sample. I.e. compared to the United States, Greece and Spain suffer from much higher
unemployment rates over the sample period. Similarly, the economies of China and Russia are highly
concentrated whereas the economy of the United States is much more diversified with an HHI of 0.13
(compared to 0.41 and 0.46 respectively). Similar interpretations can be made for the development of
technology within a country and the size of international migration. Next to the different labor market
conditions, this table also shows the variety of the average CSV value across countries. Missing data
for a particular country is indicated by a blank cell in the table. To get a feeling for how the labor
market variables relate to the CSV, Panel B of Table 1 in Appendix A shows the pairwise correlation
matrix of the predictive variables and the labor market conditions. The CSV variable shows the
highest correlation with the industry concentration rate (0.18). The matrix also reports some
correlation between the relative market value of technology and the HHI (-0.13). This negative
correlation shows that the development of technology lag in countries with highly concentrated
industries. All other labor market variables show no or limited correlation with each other.
Table 1 : Summary statistics
This table presents descriptive statistics such as mean, median, standard error, minimum, maximum, and the number of observations (N) for each variable across all countries. HHI represents the industry concentration rates, and technology represents the market value of the technology sector relative to the total industry capitalization rate of the country. The sample contains 26 countries as listed in Appendix A, Table 1. The sample period over which the data is collected runs from January 1973 to December 2017.
Mean Median SE Min Max N
CSV 0.22 0.18 0.002 0.001 6.48 9,879
Excess market return 0.76% 0.73% 0.001 -63.42% 79.76% 11,075 Unemployment rate 6.96% 6.50% 0.046 0.90% 27.96% 7,893
HHI 0.23 0.21 0.001 0.11 0.75 11,075
Technology 5.30% 0.71% 0.001 0.00% 78.47% 11,075
Net migration rate 2.62 1.87 0.029 -2.86 17.77 10,471 Inflation rate 0.05% 0.03% 0.003 -0.06% 17.41% 9,974 Dividend Price ratio 2.99 2.83 0.014 0.00 13.43 11,059
4.
Methodology
We analyze the relationship between the predictability of the CSV and labor market conditions by
combining a few commonly used methodologies in the literature. First, we run an in-sample and out-of
sample analysis (as argued by Eiling et al., 2016) at the country level and test for the impact of the
different labor market conditions on the predictive power of the CSV. Next, we run quintile
regressions where we sort the different countries in quintiles based on the labor market variables. This
results in more observations per regression compared to the cross-country analysis, since we group the
countries. Thereby we (partially) eliminate other country-specific effects. Further, we test for time
series predictive power in panel data and at the country level. At last, we test the predictability of the
CSV to future unemployment growth and link this predictive power to predictability of the stock
market returns.
Eiling et al. (2016) find that CSV has the highest predictive power for the 12 month horizon.
Therefore, our analysis focuses on the 12 month horizon as well. We incorporate the other time
horizons (k=1, k=3, k=24 and k=36) in the robustness section. In this way we check for consistency of
our results over other time horizons.
4.1
In-sample and out-of-sample: A cross-country analysis
We run a cross-country in-sample regression to obtain the relationship between the CSV and excess
market returns. In this model we regress the CSV on the continuously compounded excess market
return for all countries (j) and all specified time horizons (k):
In this model we make use of Newey & West (1987) standard errors with k-1 lags to account for serial
correlation.
Once we find the predictability of the CSV in-sample, we test whether it also has a significant
forecasting power out-of-sample. We use the first 20 years (240 months) of the data to forecast the
k-period ahead monthly excess returns. We choose for an expanding window rather than a rolling
window based on our assumption that investors make their decisions based on all information
available. For example; the k-period ahead monthly excess returns for t=242 are based on the previous
241 months. Expressing this method in mathematical terms leads to follow regression model:
Afterwards we compare the expected k-period ahead monthly returns with the realized returns using
the out-of-sample R-squared:
The last step involves testing for the impact of the labor market variables on the forecasting power of
the CSV. We regress the in-sample and out-of-sample R-squared on the labor market variables. Hence,
we run regressions with one observation per country j:
4.2
Quintile regression analysis
The next step of our analysis consists of a quintile regression analysis, in which we test the impact of
the labor market conditions on the in-sample predictive power. We sort the countries by each labor
market condition in five different portfolios based on quintiles.
The first quintile contains the countries
with the lowest level of that specific proxy (e.g. countries with lowest unemployment rates) and the
fifth quintile the countries with the highest level of that proxy (highest unemployment rates.). The
quintile selection is based on the mean value of the labor market condition over the period 10-1993 to
12-2017. Therefore, the analysis neglects predictive power due to time varying changes. We choose
the sample period such that for every country the data is available for every month.
2In this way all
countries are analyzed under the same global business cycles. We run one panel regression per quintile
in which we regress the CSV variable on the 12 month ahead excess market returns CSV. We account
for serial correlation by using Newey & West (1987) standard errors with k-1 lags. To bring our
results in perspective, we run the same analysis for the two benchmark variables; inflation rate and the
log dividend-price ratio.
4.3
Time series predictive power
Since the quintile analysis neglects the time-varying component, we extent our analysis by testing for
time series predictive power. To test for the impact of the labor market variables over time we include
the labor market variables in the regressions. We test for predictive power over time using panel data
with country fixed effects. Followed by the same regressions for each country separately to isolate the
effect cross-country:
where:
are the predictive variables; CSV, inflation rate and log dividend-price ratio.
are the different labor market variables.
Again, we make use Newey & West (1987) standard errors (k-1 lags) to account for serial correlation.
2
4.4
CSV and unemployment growth
According the sectoral shifts hypothesis, sectoral labor reallocation leads to higher unemployment
rates in the future. In this section we test the empirical validity of this theory by running a regression
model of the CSV measure on future unemployment growth (7). Unemployment growth for horizon k
is defined as
.
If the empirical results are perfectly in line with the theoretical frame work, we expect that countries
for which the CSV measure significantly predicts the k-month excess market returns are the same
countries for which the CSV measure significantly predicts unemployment growth.
5.
Results
Eiling et al. (2016) conducted their research using the CSV variable based on a cross-sectional
volatility of 49 industry specific returns. However, these specific data are only available for the United
States. In this study we analyze the impact of different labor market conditions over different
countries. Ideally, we need the same number of industries per country. In this study we use the 10
industries as available in DataStream. Since we use a reduced number of industries (and therefore less
specific defined industries) to construct the CSV from, the CSV variable will be less accurate in
measuring cross-sectional volatility between industries. Therefore, the CSV variable based on 10
industries is expected to be weaker with respect to predicting market returns compared to the CSV
variable based on 49 industries.
Table 2 shows a replication of the research done by Eiling et al. (2016) for the in-sample
analysis of the United States, adjusted for the time span 1973-2017, and the results of our study. The
results show that the CSV variable constructed from 49 industries indeed has a stronger predictive
power than the CSV variable based on 10 industries. The R-squared values are higher and the results
are more significant for every horizon, except k=36. This is in line with our expectations. Although the
results are overall less significant compared to the CSV variable based on 49 industries, we still find
significant results at the 1 percent level for k=12, k=24 and k=36 and at the 5 percent level for k=3. As
already mentioned, the CSV measure constructed from 10 industries is less accurate (captures less
cross-sectional volatility between industries) than the CSV constructed from 49 industries. Signals that
appear “noisy” for the long-term horizon are the drivers for short term price fluctuations. These signals
might not be included in the CSV based on 10 industries while they are in the 49 industries CSV. This
is also in line with the higher significance of the 10 industry CSV for the longest time horizon of k=36.
The R-squared values for the 10 industries CSV are only slightly lower (and for k=36 even higher).
Hence, while maintaining the predictive power of the CSV, we are now able to construct the CSV
variable for 26 different countries when we use the CSV variable based on 10 industries. This allows
us to compare between different labor market conditions.
Table 2: Comparison with Eiling et al. (2016).
This table includes the replicated in-sample results from the study by Eiling et al. (2016) and the results of this research for the United States. The CSV variable of Eiling et al. is based on 49 industries and run for the sample period 1973-2017, while the CSV during this research is based on 10 industry specific returns as available on DataStream and the same sample period. The output reports the beta-coefficient, the t-ratio and the in-sample R-squared. The standard errors are adjusted for serial correlation using Newey & West (1987) with k-1 lags. * p<0.1; ** p<0.05; *** p<0.01 k=1 k=3 k=12 k=24 k=36 49 Industries beta -0.07 -0.31*** -1.25*** -2.12*** -2.50*** t-ratio -1.57 -3.16 -4.14 -4.13 -3.22 R²IS 0.56% 3.38% 12.69% 19.27% 17.49% 10 Industries beta -0.05 -0.25** -1.27*** -2.45*** -3.32*** t-ratio -0.97 -2.27 -3.57 -3.37 -3.37 R²IS 0.25% 1.89% 10.24% 17.49% 18.55%
5.1
In-sample and out-of-sample: A cross-country analysis
For the in-sample analysis the k-month ahead excess market return is used as dependent variable and
the CSV as independent variable. Table 3 shows the regression output for the in-sample analysis. The
table presents the results for countries for the 12 month horizon. For a more extensive table including
the regression results for all horizons see Appendix B, Table 2. The regression results in table 3 show
the CSV is significant for 5 of the 26 countries and has an out-of-sample predictive power in 17
countries. This suggests that indeed some factors have impact on the predictive power of the CSV. For
some countries we notice that the value of the out-of sample R-squared is negative. This means that
the very naive measure of the average excess market returns of the past k months does a better job of
predicting 12 month ahead market returns than the CSV. The out-of-sample R-squared for the United
States in the 12 month horizon equals 18.98%. This value is slightly higher than the out-of-sample
R-squared obtained by Eiling et al. (2016). They found an out-of-sample R-R-squared of 14.88%. The
difference might be due to the different sample periods or caused the fact that the 10 industries CSV
ignores the smallest changes in cross-sectional volatility industry returns leading to not prediciting
small (noisy) changes in market returns. This leads to a stronger predictive power overall.
Furthermore the values of the R-squared are rising over k (Appendix B, Table 2). However,
this is (partly) due to the economectrics behind the regressions. Also, the sign of the beta coefficients
differ across countries. A possibible explanation could be that other variables are active which
influence market returns and are correlated with the CSV. Countries for which the CSV shows a
negative relationship with the market returns, the out-of-sample R-squared shows more favorable
results than for countries for which the CSV shows a postive relationship. In total 33 regressions show
significance of at least 10 percent. Of which 21 regressions show a negative relationship between the
CSV and the excess market returns and 12 regressions a positive correlation. From the sample with the
negative relationship, 62% manages to do a better job at forecasting future market returns than just
taking the market average over the past k months. This result is in contrast to the group which showed
a positive relationship between the CSV and future market returns. In this group only 25% of the
results show a positive R-squared. These results suggest that the theory of a negative relationship
between sectoral labor reallocation indeed holds with respect to out-of-sample predictability.
All in all, we conclude that the predictive power of the CSV varies a lot between countries.
Specifically, the CSV has the strongest out-of-sample predictive power in Canada, Taiwan and the
United States for the 12 month horizon.
Table 3 : Cross-country regression results CSV for k=12
This table shows the in-sample regression results as well as the out-of-sample R-squared with the CSV as explanatory variable and the 12 month ahead excess market return as dependent variable. All regressions are run for each country separately. This table only shows the regression results for k=12. See Appendix B, Table 1 for the regression results for all time horizons. A negative R2 OOS implies that the historical average of the past 12
month excess market returns as predictor has a higher predictive power than the CSV variable. The standard errors are adjusted for serial correlation using Newey & West (1987) with k-1 lags. * p<0.1; ** p<0.05; *** p<0.01. beta t-ratio R2 IS R 2 OOS Argentina 0.930** 2.111 10.85% -1.23% Australia 0.590** 2.334 3.42% 5.27% Brazil -0.300 -0.788 0.61% 1.38% Austria 0.333 0.325 0.35% 5.10% Canada -0.161*** -3.213 2.10% 8.27% Chile 0.354 0.546 0.62% 14.80% China 0.212 0.659 1.33% -2.10% Finland 0.123 0.331 0.11% -19.42% France -0.321 -0.836 0.68% 0.92% Germany -0.362 -1.483 2.55% 3.27% Greece 0.003 0.031 0.00% 1.60% Hong Kong -0.074 -1.004 1.76% 3.36% India -0.163 -0.577 1.18% 1.06% Italy -0.351 -0.475 0.50% -1.38% Japan 0.037 0.161 0.02% -3.01% Korea -0.315 -1.165 1.15% -2.27% Malaysia -0.139 -1.081 0.58% 6.53% Netherlands 0.080 0.417 0.19% -0.88% Norway 0.011 0.046 0.00% 3.62% Russia 0.025 0.662 0.12% Spain -0.082 -0.171 0.05% -4.45% Sweden 0.190 1.238 1.55% 3.14% Taiwan -0.589* -1.876 6.17% 15.69% Turkey 0.062 0.263 0.20% 26.27% United Kingdom -0.255 -1.060 1.34% 1.84% United States -1.271*** -3.571 10.24% 18.98%
As argued before, and supported by the results of table 3, we know the predictive power of the CSV
indeed differs across countries. In this section we test for changes in predictive power of the CSV
measure (both in-sample and out-of-sample) due to different labor market conditions in different
countries. We run four separate regressions with the labor market variable of interest on the
out-of-sample squared, followed by four regressions with the labor market variable on the in-out-of-sample
R-squared.
The results do not show any statistical significance (except for the impact of the HHI on the
in-sample predictive power). We think these insignificant results are partially due to the limited
amount of 26 observations (the number of countries). However, as shown in Appendix A: Figures 1-8
we observe some correlation between the labor market conditions and the predictive power of the
CSV.
Appendix B, Table 1 shows the pairwise correlation matrix between the out-of-sample-, in-sample
R-squared values and the different labor market variables. The results show that labor market conditions
have a greater impact on the in-sample predictive power than the predictive power out-of -sample. The
correlations between the out-of-sample R-squared and the labor market conditions are very low and
not in line with our hypothesis, except for the demand for skilled labor. The in-sample predictive
power however, shows higher correlations with the labor market conditions. The in-sample predictive
power shows a negative correlation with the unemployment rate and the industry concentration rates in
the form of the HHI of approximately -0.3. For the relative market capitalization rate of technology we
observe some intermediate positive correlation with the in-sample R-squared. The net migration rates
seem to have no correlation with the in-sample predictive power while it has a quite large negative
correlation with the out-of-sample R-squared.
Appendix A, Figures 1-4 show the out-of-sample regression results with the labor market
variables as independent variables. None of the regression results show statistical significance. This
result could be partly due to the limited number of observations. In contrast to the out-of-sample
regression results, the relationship between the in-sample predictive power and the labor market
conditions are all in line with our hypothesis. The industry concentration rates show a negative
relationship with the in-sample predictive power of the CSV which is significant at the 10 percent
level (Appendix A, Figure 6). The results also show a negative relationship between unemployment
rates and the predictive power of the CSV (not statistical significant). A reason for the lack of
significance could be the direct effect of the unemployment rate on future stock market returns. This
effect might overrule the predictive power of the CSV. In the analysis of time-varying predictive
power we will show that unemployment rates have indeed an impact on market returns. Figure 7
shows an upward trend; showing that the out-of-sample performance of the CSV increases when there
is an increased demand for skilled labor. Figure 8 suggests that the increased search costs of
companies slightly outweigh the larger pool of suitable workers for the job. An increase in net
migration rates increases the predictive power of the CSV.
Although no conclusions can be drawn from these results due to the limited amount of
observations, the results give an insight in which direction the predictive power CSV moves in relation
with different labor market conditions.
5.2
Quintile regression results
The quintile regression results are shown in table 4. The results with respect to the CSV measure are
reported in columns 1 and 2. The results of the inflation rate and dividend-price rate are shown in
columns 3 & 4 and 5 & 6 respectively. The CSV variable is significant at the 5 percent level for
countries with low unemployment rates while for countries with high unemployment CSV has no
predictive power at all. The results also report the CSV to be a strong predictor for countries with
diversified economies. On the other hand, for countries with high industry concentration rates the
predictive power of the CSV is insignificant and the beta coefficient even has a different sign than
expected. This result supports the theory that friction of sectoral labor reallocation is lower when
countries have highly concentrated industries. In countries with diversified economies the predictive
power of the CSV is improved since there is more variation in cross-sectional industry returns due to
the impact of the extra adjustment costs that have to be made.
The predictive power of the CSV is neither significant for countries with highly developed
technology sectors, nor for countries with lagging technology industries. At last, international
migration seems to play a role when it comes to market return predictability of the CSV measure. For
countries with high net migration rates, the CSV measure predicts the 12 month ahead excess market
returns at the 0.01 percent level. While for countries in the lower quintile the CSV is not significant at
all.
High net migration rates imply that more people – and potential workers – are moving into the
country than moving out. This leads to an increased pool of workers available. Since the CSV does a
better job at predicting excess market returns with high net migration rates, the cross-sectional
dispersion in industry returns is greater for high net migration rates. This implies that search costs
outweigh the larger pool of workers available, such that the adjustment costs of the firms are
increased.
Columns 3 till 6 in table 6 show the results for the other predictive variables; inflation rate and
the log dividend-price ratio. The results show that the predictive power of the CSV varies under
different labor market circumstances compared to the other predictive variables. The dividend-price
ratio is significant at the 10 percent level (or very close to the 10 percent level) for all quintiles. Hence,
it seems to perform well regardless of the labor market conditions. The inflation rate does not change
in the same direction of the CSV variable. These results indicate that the labor market conditions do
not impact other predictive variables in the exact same way.
5.3
Time series predictive power
To capture the time-varying component of the labor market conditions on the predictive power of the
CSV variable we include the labor market conditions in the analysis as explanatory variable. We first
test for time series predictive power in panel section, followed by a cross-country analysis. Table 5
shows the regression output from the regression analysis with the labor market variables included as
dependent variable as well as an interaction term between the CSV variable and the labor market
variable [equation (7)]. Column 1 shows no significant relationship between the unemployment rates
and the predictive power of the CSV. The direct effect of unemployment rates, significant at the 0.01
percent level, seems to overrule the CSV variable. Neither column 2 nor column 3 show a significant
change in the predictive power over time in the CSV when industries become more concentrated or
when the demand for skilled labor rise. Column 4 shows significant evidence for increasing time series
predictive power of the CSV for rising net migration rates. Supporting the theory that increased search
costs causes firms to make higher adjustment costs. Columns 1, 2 and 3 report the expected – negative
– sign for the direct relation between the CSV and the 12 month ahead excess market return. However,
for column 4 this sign flips to positive. This strengthens the result that predictive power of the CSV
varies for different levels of international migration.
Table 4 : Quintile regression results CSV
This table shows the regression results between the CSV as explanatory variable and the 12 month ahead excess market returns as independent variable. All regressions are run with country-fixed effects and the sample period runs from October 1993 to December 2017. The odd columns show the regression results ran in the lower quintiles of the labor market variables and the even columns the results of the upper quintiles. The standard errors are adjusted for serial correlation using Newey & West (1987) with 11 lags. Technology represents the market value of the technology sectors relative to the total industry capitalization. * p<0.1, ** p<0.05, *** p<0.01
Dependent variable: 12 month ahead excess market returns
CSV Inflation rate Log(DP)
(1) (2) (3) (4) (5) (6) Lower quintile Upper quintile Lower quintile Upper quintile Lower quintile Upper quintile Unemployment rate beta -0.187*** 0.050 -0.066 -0.005 0.438*** 0.116
(-3.028) (0.743) (-0.917) (-0.046) (5.182) (1.638)
Country FE Yes Yes Yes Yes Yes Yes
R² 3.10% 0.20% 0.40% 0.00% 11.70% 1.30%
HHI beta -0.187*** 0.023 -0.019 -0.208*** 0.746*** 0.119
(-2.578) (0.359) (-0.108) (-2.746) (6.906) (1.633)
Country FE Yes Yes Yes Yes Yes Yes
R² 2.80% 0.10% 0.00% 3.40% 19.8% 1.30%
Technology beta 0.049 -0.015 -0.027 -0.369*** 0.075 0.213***
(0.823) (-0.217) (-0.528) (-3.636) (1.101) (2.915)
Country FE Yes Yes Yes Yes Yes Yes
R² 0.20% 0.00% 0.10% 4.90% 0.50% 4.10%
Net migration rate beta 0.001 -0.182*** -0.071** -0.041 0.195*** 0.547*** (0.023) (-2.665) (-1.447) (-0.530) (2.856) (6.226)
Country FE Yes Yes Yes Yes Yes Yes
R² 0.00% 2.90% 0.50% 0.20% 2.80% 19.20%
The predictive power over time of the inflation rate does not change for different net migration rates,
but it does for all other labor market conditions. The predictive power of the dividend-price ratio
seems to change over time for different levels of industry concentration and international migration.
These results suggest that time series predictive power of other predictive variables behave different
from the CSV measure. The labor market conditions seem to have an impact on the CSV measure
specifically, rather than acting as aggregate shocks for all predictive variables.
Table 5 : Time series predictive power: Panel data
This table shows the results of the following panel regression:where:
are the predictive variables; CSV, inflation rate and log dividend-price ratio. are the different labor market variables.
The table shows the beta coefficients for each explanatory variable with t-statistics in parentheses below. All regressions are run with country-fixed effects. The standard errors are adjusted for serial correlation using Newey & West (1987) with 11 lags. Column (1) show the results with as labor market variable the unemployment rate, column (2) the HHI, column (3) the market value of the technology sector relative to the total industry capitalization, and column (4) the net migration rates. * p<0.05, ** p<0.01, *** p<0.001
Dependent variable: 12 month excess market return
(1) (2) (3) (4) CSV -0.052 -0.058 -0.011 0.032 (-1.125) (-0.936) (-0.403) (0.789) Unemployment rate 0.185*** (2.875) HHI -0.054 (-1.148) Technology -0.065 (-1.003)
Net migration rate 0.059
(1.268) CSV*Unemployment rate 0.036 (0.618) CSV*HHI 0.041 (0.629) CSV * Technology -0.057 (-1.224)
CSV*Net migration rate -0.081*
(-1.737)
Country FE Yes Yes Yes Yes
R² 1.60% 0.10% 0.60% 0.30%
Table 6 shows the results for time series predictive power for each country separately. This table
presents the beta coefficient and the corresponding Newey West t-ratio of the interaction term between
the CSV measure and the labor market variable (
. Increased willingness to work - as consequence
of higher unemployment rates - is expected to decrease the predictive power of the CSV. Hence, we
expect
. Similarly, we expect the predictive power of the CSV to weaken for countries
with high industry concentration (
Increased demand for skilled labor is expected to lead
(
. At last,
indicates that increased search costs outweigh the
benefits of the larger pool of workers available due to high international migration. A positive
suggests the exact opposite.
The results vary a lot across countries. For some countries the labor market variables have a
significant effect on the predictability of the CSV. While for other countries the labor market variables
have no effect at all. For instance, for Japan, China and the United States the time series predictive
power of the CSV changes for all labor market conditions except net migration rates. And for other
countries (i.e. Australia, Brazil and France) the time series predictive power of the CSV is not
influenced by any of the labor market conditions.
5.4
CSV and unemployment growth
A positive significant relation between the CSV and the future unemployment growth enforces the
idea that CSV is a good proxy for sectoral labor reallocation (Lilien, 1982). Table 7 reports the
cross-country regression output between the CSV measure and unemployment growth for the 12 month
horizon. The table presents the beta coefficient, the t-ratio based on Newey West (1987) standard
errors (with 11 lags) and the in-sample R-squared. Blank cells indicate that no unemployment data was
available for that country in the sample period. We observe mixed results when it comes to the
predictive power of the CSV to unemployment growth. Ideally, countries for which the CSV shows
negative significant predictive power for the 12 month ahead market returns horizon (Table 3, Panel
C), are the same countries for which the CSV positively and significantly predicts unemployment
growth. CSV has the expected negative sign and shows significance in forecasting stock returns for
Canada, Taiwan the United States. For Canada and the United States the CSV is able to forecast
unemployment growth as well. Based on in-sample results, in 21 countries the CSV is not able to
predict 12 month ahead excess market returns, neither does the CSV have a positive relationship with
future unemployment growth in 17 of these 21 countries (and 2 countries which have no
unemployment data). This suggests that for these 17 countries and their economies the CSV might not
be a good proxy for sectoral labor reallocation. With respect to the benchmark paper we confirm the
results of Eiling (2016) that CSV is a good proxy for sectoral labor reallocation in the United States
and significantly predicts excess market returns and future unemployment growth.
Table 6 : Times series predictive power: Cross-country
The output reported below consist of the regression analysis with the labor market variables included as interaction term, in addition to the CSV variable, on the 12 month ahead excess market returns. The regression
is run for each country j separately:
Where Z represents the labor market variables UN, HHI, Technology and net migration rates.
UN represents the unemployment rate per country and technology stands for the relative market value of the technology sector compared to the total market value of all industries. The output below reports the beta-coefficient ( and the t-statistic of the interaction term of each regression. Significance indicates that the specific labor market variable has an impact on the predictive power of the CSV over time. The standard errors
are adjusted for serial correlation using Newey & West (1987) with 11 lags. Blank cells indicate missing data for the labor market variable of interest. * p<0.1, ** p<0.05, *** p<0.01
UN HHI Technology Net migration rates
t-ratio t-ratio t-ratio t-ratio
Argentina 0.12 1.34 5.33 0.99 -2.81* -1.72 Australia 0.02 0.09 -2.15 -0.53 438.60 0.70 0.09 0.69 Brazil 0.00 -0.01 -0.76 -0.33 0.29 1.16 Austria -0.35*** -2.37 -27.89 -0.96 103.18 0.13 16.24 0.15 Canada 0.00 -0.04 -3.71 -0.59 -0.36 -0.11 -0.27* -1.78 Chile 0.89*** 2.78 -1.25 -0.10 39.70 0.19 -1.41 -0.47 China -6.63** -2.24 3.17** 2.15 -130.7*** -3.27 -3.04 -0.97 Finland -0.06 -0.45 -4.63** -2.15 -2.68* -1.91 0.39 0.71 France 0.16 0.86 -1.08 -0.11 -7.13 -0.95 -0.30 -0.33 Germany -0.10 -0.89 6.68 0.68 -15.37* -1.94 -0.02 Greece 0.05 0.79 -1.16 -1.29 325.95*** 4.03 -0.12** -2.24 Hong Kong -0.11*** -3.69 -1.39 -0.80 -3.07 -1.35 -0.04*** -2.98 India -13.89*** -2.99 -7.34*** -4.16 4.97*** 4.29 Italy -0.54** -2.01 2.09 0.43 82.07 0.70 -0.22 -0.90 Japan -0.63*** -4.18 67.76** 2.35 -17.79** -2.52 -1.10 -0.82 Korea -6.65 -0.92 2.37 1.12 Malaysia 0.39 1.57 -0.21 -0.08 -381.11 -1.52 0.06 0.30 Netherlands -0.27 -1.10 3.81 0.64 -6.17 -1.70 0.21 0.90 Norway -0.27** -2.44 -2.86 -0.43 -3.01 -0.23 0.11 1.00 Russia 0.23 1.19 -0.11 -0.07 -0.61 -1.17 Spain 0.10 1.25 -36.86 -0.85 -3.74 -0.17 -0.14** -2.38 Sweden -0.57*** -3.11 -7.68 -1.42 -3.78** -2.00 0.06 0.54 Taiwan 0.52** 2.40 -8.80 -1.41 4.47** 2.00 0.25 1.63 Turkey 0.11 0.22 -9.63** -2.02 -127.75 -0.95 -0.11 -0.51 United Kingdom 0.15** 2.29 -39.65* -1.67 0.75 0.09 -0.33** -2.34 United States 0.36** 2.47 -41.00*** -3.66 -12.22*** -3.24 -0.08 -0.49
Table 7 : Forecasting unemployment growth using the
CSV measure
This table shows the results of the following regression for each country j:
Where:
.
All regressions are run for each country separately for the data available in the sample period January 1973 to December 2017. The regression results for k=1, k=3, k=24 and k=36 can be found in the Appendix B, Table 4. Blank cells indicate missing data on unemployment for that country. Standard errors are adjusted for serial correlation using Newey & West (1987) with 11 lags. * p<0.1, ** p<0.05, *** p<0.01 Panel C: k=12 beta t-ratio R2 IS Argentina -4.987*** -2.986 21.37% Australia -1.446 -1.348 1.38% Brazil -0.455 -0.766 0.86% Austria -9.651* -1.733 11.67% Canada 0.798*** 2.772 2.03% Chile 3.967 0.706 1.95% China -0.224*** -3.103 17.43% Finland -2.828* -1.716 4.86% France 0.198 0.214 0.05% Germany -0.398 -0.370 0.25% Greece -4.187 -1.227 2.50% Hong Kong 0.241 1.039 1.90% India Italy -2.990 -0.894 2.26% Japan 0.074 0.280 0.05% Korea Malaysia 0.103 0.561 0.93% Netherlands 0.168 0.150 0.06% Norway 0.106 0.207 0.03% Russia 0.383*** 5.019 6.08% Spain -1.553 -0.316 0.19% Sweden 1.037** 2.550 6.61% Taiwan -0.240 -0.499 0.23% Turkey -3.625 -1.375 5.17% United Kingdom 0.385 0.447 0.11% United States 3.971* 1.928 2.41%