Fiduciary Asset Management, Portfolio Congruence and Herding
Behavior
-
A Spatial Statistical Investigation of Dutch Pension Funds
Master Thesis M.Sc. Economics
-
Monetary Policy & Banking In terms of a thesis internship at
De Nederlansche Bank
Submitted by: Jan Gierkes Student number: 11374047
1. Supervisor: Damiaan Chen (Ph.D.) 2. Supervisor: Ward Romp (Ph.D.) Amsterdam, 15th of August
Abstract
In this thesis I show empirically that overlapping asset management structures lead to higher congruence in pension funds’ portfolios and proof statistically that mutually managed pension funds engage in fiduciary herding. I do so by transferring the concept of spatial econometrics to the asset management structure of pension funds and estimate spatial correlations of asset allocations for mutually managed pension funds. My results suggest that mutually managed pension funds have a higher congruence, in their asset allocation, relative to all Dutch pension funds. This congruence tends to increase, the riskier the asset is categorized. I find evidence that similarities in investing in certain assets are more dependent on mutual asset management that divesting assets. This does also account for asset reallocation due to regulatory arbitrage. Furthermore, I find evidence that pension funds which share a similar asset manager occasionally trade in the same direction at the same time. These results indicate that asset managers might align their investment strategies for different pension funds, for which they are in charge of.
Table of Content
1. Introduction 4
2. Theoretical Background and Hypotheses 6
2.1 Herding Behavior 6
2.2 Asset Management Delegation 8
2.3 Asset Allocation and Management Mandates 9
2.4 Performance Assessment and Compensation Structure 11 2.5 External Environment and Information sets 12
3. The Dutch Pension System 13
4. Data and Descriptive Statistics 15
5. Methodology 18
6. Empirical Results 22
6.1 Congruence in Asset Weights 23
6.2 Time Narrative 26
6.3 Herding Behavior 35
6.4 Possible Consequences and Policy Implications 37
7. Limitations 38
8. Conclusion 40
9. Reference List 42
10. Appendix 45
1. Introduction
Pension funds have become one of the largest investor groups over the past decades, globally managing assets of more than 36 trillion USD (OECD, 2016), which makes them highly significant for developments in financial markets. Not only because of their sheer size, pension funds’ asset allocation and investment strategies are crucial from a macro prudential perspective: Multiple studies have shown that pension funds tend to follow each other in their asset allocation and might herd in their investment decisions (e.g. Broeders et al. 2016 or Blake et al. 2017).
Herding behavior, defined as trading in the same direction at the same time (Nosfinger and Sias, 1999), might pose a significant threat to financial stability: On the one hand, clustered movements in asset allocations among institutional investors can contribute to excess volatility in markets, by driving asset prices away from their fundamentals (e.g. Avery and Zemsky, 1998). On the other hand, congruent compositions of pension funds’ portfolios might pose hazards in terms of potential contagion. If asset allocation is aligned due to herding across pension funds, their portfolios are more vulnerable to macroeconomic shocks (Kioyataki, 2002). An investigation of herding tendencies and congruent portfolios among pension funds is therefore of key importance from a macro-prudential perspective.
As mentioned by the Bank of England, pension funds tend to outsource large parts of their asset management (Haldane, 2014). This is especially the case in the Netherlands, as described in more detail in Chapter 3. The asset management industry itself is dominated by a handful of corporations. Consequently, different pension funds are managed by mutual third parties. Theoretically, asset managers should act more or less mechanically, in line with their management mandate, which differs across pension funds. Yet structural and economic circumstances might lead to an alignment in investment decisions beyond different mandates. This gives rise to the question: If asset allocations of mutually managed pension funds are not independent, are they aligned by a common asset manager?
Asset managers find them self in a narrow playing field. After the recent financial crisis funding ratios of pension funds plunged to historically low values, exerting pressure on
asset managers to perform well and close the funding gap. At the same time, historically low interest rates complicate long term investments. This might lead not only to a constraint demand for certain assets but could also foster herding behavior among asset managers. (Blake et al., 2017)
Moreover, asset managers are evaluated in comparison to market and peer group benchmarks. In order to keep their reputation, asset managers might align their investment decisions and follow similar benchmarks for different portfolios for which they are in charge of (Scharfstein and Stein, 1990). It is likely that they rely on mutual information sets, share a common risk appetite and therefore they might align their asset allocations accordingly (Froot et al. 1992)
The goal of this thesis is twofold: First, I investigate the effects of mutual asset management on congruence of pension funds’ portfolios. Thus, I test if congruence in pension funds’ asset allocation coincides with similarities in their asset management structure. Second, I examine if mutually managed pension funds trade in the same direction at the same time and consequently herd due to overlapping asset management structures. Thereby I cover a static and dynamic perspective on asset managers’ influence of pension funds’ asset allocations.
I do so by choosing a spatial econometric method which is largely used in quantitative geo science. In this context however, I treat pension funds as spatially close when they share a common asset manager. This gives me the opportunity to group pension funds according to their asset management. Next, I calculate spatial correlations in their asset allocations by constructing a parameter called Moran’s I for different asset classes, risk categories and time periods. The idea is simple yet effective: Moran’s I is set up as a parameter, which measures the correlation of asset allocations of mutually managed pension funds, relative to the whole sample of pension funds. This allows me to measure allocation congruencies and herding tendencies, which are related to overlapping management structures, across assets and from a time narrative perspective.
This thesis provides a contribution to academic literature in two ways: First, the empirical relation between asset management delegation and herding behavior is, to the
best of my knowledge, relatively new to academic research. Moreover, I contribute to the empirical literature on herding behavior by methodically transferring the concept of spatial correlation to the management structure of pension funds.
The thesis is structured as follows: In chapter 2 I expound the theoretical context, which serves as a background for the constructed hypotheses. Chapter 3 introduces the institutional framework of the Dutch pension system. Chapter 4 describes data and presents descriptive statistics. The methodology is described in Chapter 5. While Chapter 6 illustrates my empirical results and their interpretation, Chapter 7 provides some limitations of my methods. Lastly, Chapter 7 presents the conclusion of this paper.
2. Theoretical Background and Hypotheses
2.1 Herding Behavior
Herding behavior in the financial sector has gained attention in academic literature over the last years. Both theoretical and empirical literature address herding behavior on a large scale. Almost every big economic institution has published academic research on herding behavior. Besides the IMF ( e.g. Bikhchandani and Sharma, 2001), the World Bank (Raddatz and Schmukler, 2011), the Bank of International Settlements (e.g. Nirei et al. 2012) and the FED (e.g. Cai et al. 2012), research on herding behavior has recently also been published by the Dutch central bank (Broeders et al. 2016). One of the main reasons why this phenomenon is studied extensively is the possible implications for financial stability. Institutional herding behavior is of particular interest, since clustered investment decisions can have a severe impact on financial markets. Chan and Lakonishok (1995) have shown that institutional herds can impact market prices by driving assets away from their fundamentals and create considerable externalities. If institutions follow each other in their asset allocation, their portfolios are likely to be congruent. Higher congruence raises potential hazards of contagion (Broner and Gelos, 2003).
Several studies state that institutions are more likely to herd than individual investors. Froot et al. (1992) claim that institutions receive and react to the same exogenous signals. Benarjee, (1992) suggests that institutional investors are more informed about each other’s investment decisions than individuals and may consequently herd. Empirical studies on pension funds’ herding behavior are less frequent in academic literature than studies of mutual funds. This might be traced back to the scarcity of pension fund asset allocation data (Blake et al., 2017). One of the first studies which coped empirically with herding behavior of pension funds was published by Lakonishok, Schleifer and Vishny (1992). Their empirical research suggests that pension funds herd, as they in- and divest simultaneously in similar asset classes. More recent literature of herding behavior finds evidence that herding behavior is especially pronounced among subgroups. According to Blake et al. (2017), pension funds with similar size or funding structure are more likely to follow each other in their investment decisions than funds which differ in their institutional structure.
Herding behavior among pension funds can arise due to different reasons. Three main motives emerged in theoretical literature over the past years: First, pension funds may restructure their portfolios due to external shocks in a similar way. Changes in macroeconomic circumstances can lead to similar asset reallocations. Consequently, they rely on similar information about market developments and possible returns of asset categories. A similar exposure to this information causes identical asset allocation decisions (Froot et al. 1992). Second, pension funds might restructure their portfolios in a common way to changes in regulation. If certain asset classes are valued differently in regulatory frameworks, pension funds may restructure their portfolios in the same manner, to make use of regulatory arbitrage (Sias, 2004). Third, portfolios might be restructured according to investment decisions of pension funds peer groups. Asset managers might follow each other in their asset allocation due to the fear of underperformance relative to different pension funds or asset managers (Scharfstein and Stein, 1990). Broeders et al. (2016) find empirical evidence, for all three types of herding among Dutch pension funds.
In addition to the three motives mentioned above, it would be useful to add a structural dimension to the analysis of herding behavior among pension funds: This thesis aims to
examine the impact of overlapping asset management structures for multiple pension funds. Therefore, main institutional features are highlighted in the following section and serve as a background for my hypotheses.
2.2 Asset Management Delegation
In contrast with wide spread misconceptions, pension funds are generally not in charge of managing their assets in a daily operational manner. Instead, asset management tends to be delegated to third party investment corporations, which are responsible to invest in assets in line with the pension funds overall investment objectives. This process, called fiduciary asset management, has gained worldwide importance over the last decades and evolved to be the primary model of asset management among institutional investors and especially pension funds (Haldane, 2014).
In particular, fiduciary management is a common practice among Dutch pension funds. According to the European Fund and Asset Management Association, Dutch institutions are ranked fifth in Europe, with a total amount of approximately €1,2 trillion assets under management (AUM) (EFAMA, 2017). While almost all small pension funds delegate their portfolio management to investment companies, some of the larger Dutch pension funds have created subsidiaries, which are in charge of their asset management. Initially, these subsidiaries exclusively serviced their parent pension fund. However, over time these subsidiaries also began operating for other pension funds, which originally had no link to the parent company of the fiduciary asset manager.
Pension funds delegate their asset management for different reasons. First, asset management has become more complex and opaque. A rise in range and complexity of financial instruments requires skilled specialists in order to cope with return and risk modeling adequately. Second, the regulatory framework for Dutch pension funds narrows the financial scope and necessitates investment expertise. Furthermore, market volatility and a potential rise in risk and return uncertainty enhance needs for sophisticated dynamic asset allocation strategies (Xia, 2001).
In addition, large asset management corporations benefit from comparative advantages in terms of their economies of scale. In financial literature, it is commonly accepted that financial investors need an ample size to manage assets efficiently and sustainably (e.g. Latzko, 1999)
Asset management corporations operate in a highly dense market, which is dominated by a handful of investment companies. Therefore, it is likely that one asset management corporation is in charge of multiple pension funds.
2.3 Asset Allocation and Management Mandates
In general, asset allocation and investment decisions of pension funds are moving in a limited scope, which have to be in line with prudential principles of Dutch pension law. In theory, adequate principles of risk modeling, implementing appropriate hedging strategies and ensuring asset-liability matching, narrow the range in which asset managers can in- and divest in assets and restructure the pension fund’s portfolio (OECD, 2006).
The overall investment policy is commonly set by the governing board of the pension fund, clarifying pension funds objectives and a corresponding asset-mix, the so called strategic asset allocation (SAA). The SAA mirrors the long run investment strategy of the pension fund and defines the composition of different asset classes within the pension fund’s portfolio. Weights for different asset classes are chosen to match the pension funds’ liabilities in terms of their maturity profile and risk exposure. In general, weights of asset classes are allowed to move within a certain bandwidth, but should return to the defined and fixed mean in the long run (Bikker et al., 2010).
As mentioned above, third party asset managers are in most cases in charge of day to day asset management tasks, which are predefined in the asset manager’s mandate. The mandate specifies investment objectives and boundaries which should be in line with the pension funds overall SAA and is generally advised by the pension fund's consultant. Theoretically, strategic asset weights cannot be influenced by the asset
predefined weights ( Sections 13 and 14 of the Decree on the Implementation of the Pensions Act, DNB 2015). Yet, the Dutch regulatory framework provides an explicit open standard for outsourcing activities of pension funds. Section 34 of the pension act obliges the pension fund to ensure that the asset manager complies with rules and regulations for overall investments. These should be sound and ethical and in line with the interest of the pension funds’ beneficiaries (Prudent person rule, section 135 of the pension act, DNB 2015).
Nonetheless, in reality actual asset allocations can move away from the SAA. Deviations of the strategic asset allocations can be explained in two ways: First, market movements are passively changing the pension fund’s portfolio, as assets are valued by time varying market prices. If the actual asset allocation moves too far from the SAA, e.g. due to changes in asset valuations, pension funds are forced to rebalance their portfolios. According to Bikker et al. (2010), 39 percent of changes in asset allocation among pension funds can be traced back to rebalancing strategies. Rebalancing typically means selling winners and buying losers, in order to restore predefined asset weights and is therefore sometimes referred to negative feedback trading. (EIOPA, 2016)
Second, pension funds can deviate intentionally from their SAA in reaction to returns of different asset classes. This kind of asset rearrangement is called tactical asset allocation (TAA) or momentum trading. Assets are actively swapped into different asset classes, in expectation to gain additional market returns. In practice, the asset manager might be in charge of TAA decisions and free to select deviations in a predefined range. Within the asset classes, asset managers are also responsible for in- and divesting on the level of securities (Blake, 2017).
2.4 Performance Assessment and Compensation Structure
Moreover, it is important to understand that asset management is an agency activity. As assets are managed on behalf of the pension fund and their beneficiaries, asset managers are assessed in terms of their investment performance. Under-performance in the short run and significant deviations from the defined mandate are most of the time the reason for replacements of asset managers (Financial Times, 2014). Nonetheless, asset managers are often assessed rather on their relative performance against their peers, than on their absolute returns. Thus, asset managers might have an incentive to relate on either approved strategies for different portfolios or strategies of their peer group (Blake et al., 2017). Most literature suggests that asset managers rely on stable and unspectacular investment strategies to maintain reputation, keeping old and gaining new customers (Prendergast and Stole, 1996). Multiple empirical studies on pension funds’ asset management suggest that asset managers peg their strategies around the peer group’s average fund, which appears to be an index matcher, indicating that pension funds tend to herd around the market benchmark (Blake et al., 2017).
In addition, compensation structures of asset managers contribute to an alignment in investment strategies: Fees paid to fiduciary managers are mostly dependent on the overall amount of managed assets, less on the performance of these assets (Del Guercio and Tkac, 2002). Blake (1999) finds evidence that neither overperformance is compensated exceptionally, nor that underperformance leads to financial penalties. Consequently, imitation of assets managers is rather rewarded than punished and managers are urged to perform close to the market’s average or simply follow market benchmarks in their investment strategies. Deviations from proved strategies are therefore unlikely and an alignment in security selection and asset allocation for different portfolios seems reasonable.
Due to this theoretical background, this section concludes with the following hypotheses: Asset allocations are more congruent for pension funds, which share a mutual asset management (H1).
2.5 External Environment and Information sets
In general, Dutch pension funds operate under a similar external environment and asset managers face identical problems in order to respond to changes in economic circumstances or changes in the regulatory framework. More concretely, the fiduciary managers are exposed to similar developments on security markets, receive the same market signals and face similar risks on their asset and liability sides (Bikhchandani and Sharma, 2001). Asset allocation and investment strategies have to counter these risks, while generating returns and ensuring a long term asset liability match.
Herding among pension funds asset allocation arises when investment strategies for commonly managed pension funds are similar. As information sets, underlying asset investment decisions are likely to be highly related within one asset management cooperation (Froot et al. 1992), pension funds may be induced to implement common investment strategies, when managed by the same third party.
In theoretical literature, this kind of herding is commonly related to an information motive. Bikchandani and Sharma (2001) define herding driven by a change of economic or regulatory fundamentals as “spurious herding”, as it is a consequence of common exposure to external developments and not an intentional replication of investment strategies without an economic reason.
Herding driven by changes in fundamentals is fueled by a rise in index and dynamic asset allocation strategies. According to Haldane (2014), active asset management is increasingly replaced by automated allocation algorithms, encountering changes in economic circumstances. Homogeneity in these dynamic models are predictably leading to an alignment of asset allocation decisions.
This leads to the following:
H2: Mutually managed pension funds react in a similar way to changes on economic and regulatory circumstances.
3. The Dutch Pension System
The Dutch pension system consists of three pillars, which contribute to the retirement scheme in the Netherlands. The general old age pension (Algemene Ouderdomswet, AOW) is the first pillar and is a governmental flat pension scheme, available for all Dutch citizens. The second pillar contains pension plans which are related to employers (Ponds and van Riel, 2007). Employers save collectively for their employees to guarantee a company based retirement scheme. These savings are managed by pension funds. The third pillar consists of private savings, which severe as a coverage for individuals, in addition to the public pension.
Pension funds can generally be subdivided into Defined Benefit (DB), Defined Contribution funds (DC) or hybrid schemes. While this analysis focuses only on Dutch DB pension funds, this distinction is nevertheless important, as the scheme structure plays an important role in pension funds’ portfolio and asset management. As one can derive from its name, DC pension plans are arranged as predefined and more or less fixed contributions from the employee, employer or both. At retirement, the employer receives the amount accumulated and invested over the working period, either as an annuity or lump sum. DC plans are conceptually simpler but raise uncertainty for the employee about his retirement scheme (Bodie et al., 1988). In the Dutch pension system, DC plans still play a minor part. In contrast, DB plans calculate the retirement benefits of their employees ahead of the retirement and guarantee (to some extent) payments thereafter (ibid).
The Dutch DB pension system is an especially interesting case to study herding behavior, since it shows some unique patterns which could foster herding:
First, the Dutch pension funds underwent a process of significant market concentration. The number of pension funds shrunk from over 900 to just about 300 active funds over the last fifteen years (DNB data). Market consolidation affected especially smaller pension funds, which in most cases merged with other pension funds (IPE, 2015). Simultaneously, the amount of total assets held by pension funds tripled within the same time span to approximately €1200 billion, leading to an allocation concentration which is almost nine times higher than in 2001 (DNB, 2017). Moreover, the pension market is dominated by a handful of pension funds, the five largest funds manage over fifty percent of total pension assets in the Netherlands. ABP, the pension fund for the public sector and PfZW, responsible for employees of the healthcare sector, dominate the Dutch pension market, with a combined share of 40% (ibid.)
Second, the financial crisis in 2008 and the following European sovereign debt crisis changed pension funds' playing field dramatically. Most Dutch pension funds suffered significant declines in their funding ratios (the ratio of total assets and liabilities). In the outbreak of the financial crisis, the average funding ratio plunged from 144 percent in 2007 to 95 percent at the end of 2008 and only slowly recovered to 103,75 percent in 2017 (DNB data). At the same time, interest rates are, mainly as a consequence of quantitative easing programs of central banks, unfavorably low. This adds additional pressure on constrained long term investors, like pension funds.
Third, the Dutch pension system is large on an absolute and also relative scale. In 2016 total pension assets represented 196% of the Dutch GDP or over €80.000 per inhabitant (Towers Watson, 2017). Therefore, the Dutch pension funds constitute a substantial part of the Dutch financial sector. An analysis of potentially clustered asset allocation decisions and congruence in pension funds' portfolios are therefore important from a macro prudential perspective.
4. Data and Descriptive Statistics
The data used in this analysis is provided by the Dutch Central Bank (DNB). Dutch Pension funds are obliged to report their portfolio structure in terms of actual and strategical asset allocation, as well as corresponding returns of these assets to the DNB. Among 302 pension funds existing in Netherlands in 2017, 212 reported their asset allocation and corresponding returns of different asset classes on a quarterly basis. Pension funds are subdivided by the DNB into four different classes (T1-T4), which state their size, measured in participants, the amount of assets and liabilities and hence also their system relevance for the Dutch pension sector. While pension funds in class T1 are ‘special cases’ due to restructuring or liquidation, class T2 constitutes of 180 small and active pension funds. Pension funds grouped in class T3 are medium sized and count for 27 pension funds. The five biggest pension funds are categorized as T4.
In my analysis I use quarterly asset allocation and asset return data of defined benefit (DB) pension funds, which are classified T2 or higher, from the first quarter in 2012 until the fourth quarter in 2016. I cover pension funds asset allocation of equity, subdivided into mature and developing markets and allocation of fixed income assets. Fixed income assets are grouped according to different credit risk categories, based on Fitch ratings and covering bonds rated AAA, AA, A, BBB and lower than BBB. Fixed income assets incorporate sovereign bonds as well as corporate bonds, mortgages and index-linked bonds. After elimination of pension funds for which quarterly data is not available over the complete time horizon, a balanced panel of 188 pension funds remains for the empirical analysis. The balanced panel consists of 33.840 observations for pension funds asset allocations, 11.280 observation for returns (total returns, fixed income returns and equity returns) and 32.148 observations for quarterly differences in the asset allocation.
In my analysis I cover approximately 85 percent of total assets held by pension funds in the panel. The remaining 15 percent are assets in real estate, hedge funds, microfinance and other alternative assets. These assets are excluded from the congruence and herding analysis since observations are too little within one group of mutually managed pension funds to infer valid estimations.
Data of pension funds fiduciary asset management structure also stems from the DNB database. In contrast to asset allocation and asset returns, data of the asset management structure is not available on a quarterly basis. The database only provides information on the current fiduciary asset management corporation and fiduciary asset managers are labeled as the main asset manager. Pension funds are grouped in my analysis according to this information. Since data of the asset management structure is time invariant, I cover a rather short time period of 20 quarters. Thereby I make sure not to exceed the average duration of a fiduciary asset management contract, which is on average 4 years, (DNB data, Blake, 2017). This minimizes estimation biases, since it is likely that the asset management structure has changed for time periods belonging to the distant past. A more detailed discussion of (data) limitations is provided in chapter 7.
Descriptive Statistics
As pictured in Table 1, fixed income assets account for the majority of assets held the pension funds of this sample. On average, 57 percent of total assets are related to fixed income securities. The majority of these fixed income assets are riskless bonds with a credit rating of AAA. As displayed in the Appendix, however, the relative amount of AAA assets decreased from 32 percent in 2012 to 21 percent in the end of 2016, while the cross sectional variance of AAA weights decreased simultaneously. Bonds with a credit rating of BBB and lower than BBB constitute a minority of pension funds asset weights. While BBB bonds account for 8 percent of total assets on average, bonds with a credit rating lower than BBB represent just 4 percent on average of total assets, measured as asset weights (the relative share of one asset category in comparison to total assets in pension fund’s portfolio).
Assets in equity account for 27 percent of total assets on average in the given panel. Equity of assets related to mature markets constitute a majority with a weight of 23 percent. The relative share of equity stayed relatively stable and dropped sharply in the beginning of 2015.
Actual Allocation Observ. Mean St. Dev Min Max 95%CI LB 95%CI UB as weights of Total Assets
Fixed Income 3760 0,58 0,13 0,10 1,00 0,33 0,84 AAA Bonds 3760 0,24 0,12 0,04 0,74 0,01 0,47 AA Bonds 3760 0,12 0,08 0,00 0,80 0,00 0,28 A Bonds 3760 0,07 0,06 0,00 0,73 0,00 0,18 BBB Bonds 3760 0,08 0,05 0,00 0,35 0,00 0,18 <BBB Bonds 3760 0,04 0,04 0,00 0,26 0,00 0,11 Equity 3760 0,27 0,11 0,00 0,69 0,05 0,48
Equity Mature Markets 3760 0,23 0,10 0,00 0,69 0,03 0,42 Equity Emerging Markets 3760 0,04 0,03 0,00 0,36 0,00 0,09
Table 1. Average asset weights (N=188, T=Q1 2012 – Q4 2016)
As stated in the introduction, the asset management industry is highly concentrated and hence it is plausible that different pension funds delegate their asset management to a mutual third party. And indeed asset management structures of Dutch pension funds are overlapping to a large extent: Figure 2. illustrates the relation of pension funds and their asset management structure. The biggest clusters of mutually managed pension funds are displayed as gray boxes, in which every vertical line represents one pension fund. Single points display pension funds which are not categorized to a distinct asset manager, as either their asset management structure was not reported to the DNB, multiple asset managers were in charge of their asset management or these pension funds did not outsource their asset management to third parties. 66 out of 188 pension funds are not grouped due to a common asset management structure. Nonetheless they are included in the sample, to relate asset allocation of mutually managed pension funds to all pension funds, regardless of their asset management structure.
Figure 1. Sketched pattern of asset management structure of Dutch pension funds
Gray boxes: pension funds with mutual asset management, single dots: pension funds not grouped in management cluster.
5. Methodology
“Everything depends on everything else, but closer things more so” Tobler’s first law of geography (Tobler, 1970)
Testing for herding behavior has become increasingly popular in academic literature with reference to the financial sector, but the definition of herding and accordingly the methods that test for herding differ to a large extend. The spatial econometric approach, chosen in this paper, is commonly used in geostatistics and quantitative regional science, but transferred to the analysis of the pension funds asset management context. Spatial econometrics are used when distance has an explanatory effect and its omission would lead to biased or inconsistent estimations. Here, the spatial relation between pension funds is covered by treating pension funds as spatially close, if they share a mutual fiduciary asset management corporation. This transformation gives me the possibility to measure the correlation of asset allocation decisions due to overlapping asset management structures in a simple but powerful way:
Spatial Structure – Determining the Spatial Weight Matrix
In the first step, I define the spatial relation of pension funds due to information of their asset management structure. Following the method of Anselin and Bera (1998), I assume that the spatial structure is known a priori. Therefore, I can abstain from an extensive estimation of the true spatial distances between the observed pension funds. These assumptions are valid in this context, since pension funds state their responsible asset management parties in their annual reports, which are collected in the underlying dataset for this analysis. To some extent, the determination of the spatial relations is nonetheless arbitrary and a more detailed discussion of methodological limitations is provided in Chapter 7. Spatial relations according to their management structure are displayed by a spatial weight matrix Μ. This matrix is constructed in a binary form, where every element of the matrix mirrors overlapping asset management structures of pension funds 𝑖 and 𝑗 due to mutual investment corporations: Zeros indicate no mutual asset management party and ones indicate similar asset managers.
Spatial Correlation – Moran’s I
In a second step, I analyze if correlation of returns and asset allocations coincide with congruence in the pension funds asset management structure. I do so by constructing a spatial correlation parameter called Moran’s I (Moran, 1950). Moran’s I resembles the classical correlation coefficient of Pearson, but instead of measuring the correlation of two variables, Moran’s I gives correlation of two entities (pension fund 𝑖 and 𝑗) for one univariate variable 𝑥. The correlation is weighted by the spatial distance 𝑤𝑖𝑗 between two entities. Finally, Moran’s I is multiplied by the sum over all entities 𝑁 divided by the sum over all 𝑤𝑖𝑗 called 𝑊.
𝑀𝐼𝑡𝑘 = 𝑁 𝑊
∑ ∑ 𝑤𝑖 𝑗 𝑖𝑗(𝑥𝑖𝑡𝑘− 𝑥̅̅̅̅)(𝑥𝑡𝑘 𝑗𝑡𝑘 − 𝑥̅̅̅̅)𝑡𝑘
∑ (𝑥𝑖 𝑖𝑡𝑘− 𝑥̅̅̅̅)𝑡𝑘 2
Hence, Moran’s I ranges between -1 and 1 and indicates whether the variable of interest 𝑥 is spatially correlated, or in other words spatially dependent. Whereas values of 1
indicate a perfect spatial correlation, values of -1 indicate perfect spatial dispersion. Values of 0 suggest random spatial dependencies.
Perfect dispersion perfect spatial correlation spatial randomness Figure 2. Illustration of spatial correlation
In the context of an empirical investigation of herding behavior among Dutch pension funds, Moran’s I allows the grouping of pension funds according to their asset management structure and simultaneously calculates if asset allocations 𝑥 are correlated within one asset manager.
Regarding a robust measure of herding due to overlapping management structure, I calculate spatial correlations of actual asset allocations for different asset classes 𝑘 as well as spatial correlation of portfolio returns. I thereby capture cross sectional correlations of a broader range of asset categories, as well as indirectly possible correlations of individual securities.
In the original form, Moran’s I measures only cross sectional spatial dependencies. For my analysis I extend the parameter with a time dimension. This allows me to observe if spatial dependencies in asset allocation change over time and hence extends the analysis with a time narrative. This extension makes sense in so far as it reveals if asset managers react in a common way for different portfolios to changes in macro-economic circumstances or changes in regulation. Consequently, it indicates if herding and congruence in asset allocations are pronounced over a particular time horizon.
In my analysis I focus on spatial correlation of the most dominant asset classes on pension funds’ portfolios. These contain fixed income and equity and a breakdown for different risk categories. Thereby I cover 85 percent of all assets of the pension funds sample portfolios. Furthermore, I decompose the asset classes into different credit risk
categories, to investigate if spatial correlations differ in that regard. More concretely, I analyze equity allocations for mature and emerging markets and bond allocation for ratings of AAA, AA, A, BBB and lower than BBB securities.
Testing for significance of spatial correlation
Another benefit of Moran’s I are its distributional properties. Cliff and Ord (1981) showed that Moran’s I is asymptotically normally distributed, when the number of entities exceeds fifty. Therefore, can be drawn inference if spatial correlation is significant by a simple t-test. I.e. transforming Moran’s I to the standard normal distribution and comparing its p-value to a predefined confidence level.
𝐻0: 𝑠𝑝𝑎𝑡𝑖𝑎𝑙 𝑟𝑎𝑛𝑑𝑜𝑚𝑛𝑒𝑠𝑠 𝐻1: 𝑠𝑝𝑎𝑡𝑖𝑎𝑙 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛
Congruence and Herding
The influence of overlapping fiduciary asset management structures can be analyzed from two different perspectives. One perspective is the influence of mutual asset management on pension funds’ congruence in their asset allocation, the other perspective incorporates pension funds herding behavior, in the sense of trading in the same direction at the same time (Nosfinger and Sias, 1999). Methodically, these two perspectives can be captured by a) calculating spatial correlation for actual asset allocation at one point in time to measure congruence in mutually managed portfolios. And b), by taking the differences between two points in time, it states the direction in which pension funds trade. Hence replacing the actual allocation of a certain asset with its first difference in Moran’s I shows if mutually managed pension funds trade in the same direction at the same time and consequently herd.
Over the long run, spatial herding should lead to spatial congruence in pension funds´ asset allocation. Since I analyze a rather short period and pension funds “start” with different asset allocation weights in their portfolio, spatial correlation and spatial herding do not have to move in the same direction. In comparison between herding and
congruence, multiple scenarios can occur. To understand how congruence and herding intertwine, it might be helpful to think of an example: If pension funds i and j share a common asset manager and start with a fixed income weight of 0.2 and 0.3 respectively and both increase their fixed income weight by 0.1 spatial correlation of differences would arise (if pension funds with different asset management would not trade in this direction), and consequently they would herd. Their congruence would be less effected. Another possibility could be the following scenario: the two pension funds start with the same weights as in the afore mentioned example. But now only pension fund i increases the weight by 0.1 (all other pension funds stay the same). Now one would see an increase in spatial correlation of their congruence, but no increase in spatial correlation of differences, which measures their herding behavior.
For the interpretation of Moran´s I, it is important to understand that the spatial correlation is measured relative to the whole sample of pension funds. Due to the spatial weights, Moran´s I measures correlation which states a value relative to all pension funds regardless of their asset management cooperation. This has to be taken into consideration when interpreting values of spatial correlation. Low values of spatial correlation can therefore have two possible implications: Firstly, there is no fiduciary herding or congruence in mutually managed pension funds or, secondly, herding due to similar asset managers is “crowded out”. If all pension funds have similar asset weights or move in the same direction, the influence of mutual asset management is negligible.
6. Empirical Results
In this section I discuss the results of the analysis based on Moran’s I. Section 6.1 discusses spatial correlation of asset weights, allowing me to analyze congruence in pension funds allocations of mutually managed pension funds. I present spatial correlation estimations across different asset classes and of their related credit risk categories. Section 6.2 discusses the development of congruencies over time (Q1 2012 – Q4 2016) allowing me to analyze clustered asset reallocations due to changes in economic and regulatory circumstances in the medium run. In section 6.3 I analyze spatial correlations of quarterly changes in pension funds´ asset allocation. As described in Chapter V) this allows me to measure if mutually managed pension funds
trade in the same direction at the same time1. Section 6.4 points out possible consequences and suggests policy implications which might be derived from this analysis.
6.1 Congruence in Asset Weights
First and foremost, spatial correlation of actual asset weights are positive over the analyzed time horizon, with only few insignificant exceptions. As stated in Table 1 and Table 2, estimations of Moran´s I are located on average above zero for every asset and credit risk class, while periods of spatial randomness constitute exceptions. Moran’s I moves on average between 0.15 for equity allocation and 0.16 for fixed income allocations. These results suggest that pension funds which share a similar asset manager have a higher correlation in their asset weights and hence a higher congruence in their asset allocation. Consequently, Hypothesis 1 from the theoretical background is verified.
A more detailed comparison of Moran’s I on a return and asset allocation level gives evidence if congruence is rather observable on a general asset class or more detailed security choice.2 Comparing spatial correlation of fixed income returns and allocation shows a relatively small difference, both indicating a spatial correlation on average of around 0.16. Yet, spatial correlation of fixed income returns reaches a maximum of 0.46, whereas spatial correlation of equity allocation only reaches a maximum of 0.26. This indicates that occasionally congruence in securities is larger than in the asset class for fixed income assets. Yet, testing for significant differences in spatial correlation means show, that the difference between spatial correlation of fixed income returns and fixed income assets is not significant. Hence one cannot conclude that congruence due to mutual asset management is higher in their security choice than on a general fixed income asset class level.
1Distinct quarterly spatial correlation values for allocation of different asset classes and their returns is provided Appendix.
A slightly different picture can be drawn in comparing average spatial correlation of equity returns and equity allocation. T-test show significant differences in their mean, so that one can conclude that spatial correlation of returns is significantly higher on a 1 percent level. This result goes along with the amount of time periods, in which Moran’s I is significantly larger than zero. Whereas Moran’s I of equity returns is larger than zero in 15 out of 20 time periods, with 95 percent confidence, Moran’s I of equity allocation is only in 12 out 20 times larger than zero. These results suggest that mutually managed pension funds have a larger congruence in their equity securities than in their equity class.
Mean Variance Observ. > 03 Min Max
Total Returns 0.109 0.0074 13 -0.036 0.245
Returns Fixed Income 0.169 0.0107 17 0.045 0.461
Returns Equity 0.174 0.0144 15 -0.063 0.374
Allocation Total FI 0.156 0.0028 19 0.062 0.263
Allocation Total Equity 0.1488 0.0073 12 0.029 0.282
Table 2. Moran’s I, Returns and Actual Allocation (Q1 2012 – Q4 2016)
A decomposition of fixed income and equity assets into different risk categories gives a further indication of asset allocation congruencies due to mutual asset management differs for different risk levels. Table 2 shows average values for spatial correlation for fixed income assets, subdivided by their credit risk rating and for equity assets for developed and emerging markets. With the exception of Bonds rated AA, spatial correlation values tend to be higher for fixed income assets with a higher credit risk. A similar picture can be drawn by comparing values of Moran’s I for equity allocation in mature and emerging markets. While spatial correlation of mature market asset allocation is on average 0.162, Moran’s I for equity in emerging markets shows a higher value of 0.199 These results suggest that congruencies in asset allocations of
mutually managed pension funds tend to be higher for assets with a higher credit risk exposure, relative to all Dutch pension funds.
Average Moran's I Variance Observations > 04 Min Max
Bonds AAA 0.141 0.002 18 0.078 0.232 Bonds AA 0.243 0.009 19 -0.029 0.386 Bonds A 0.101 0.002 14 0.033 0.226 Bonds BBB 0.130 0.004 16 0.049 0.314 Bonds <BBB 0.206 0.005 20 0.104 0.343 Equity Mature M. 0.162 0.008 13 0.037 0.271 Equity Emerging M. 0.199 0.003 19 0.019 0.263
Table 3. Moran’s I, Risk Classes, Actual Allocation (Q1 2012 – Q4 2016)5
This observation is statistically proved by t-tests for differences of spatial correlation means: Testing for significant differences of the average Moran´s I, indicates that congruence in AAA Bonds is significantly lower for bonds with a credit rating less than BBB on a 99 percent confidence level. This also applies for a comparison of bonds with credit rating A and BBB bonds, as well as bonds rated lower than BBB. Moreover, with a probability of 99 percent, equity of mature markets show a significantly lower spatial correlation as equity in emerging markets. Spatial correlation for AA rated bonds constitute an exception in the comparison of congruencies for different risk levels. The average Moran´s I peaks out with a remarkably high value of 0.32. This might be explained by economic developments, which are discussed in section 6.2.
The tendency of higher congruence in riskier assets might be explained from a broader perspective: In general, pension funds use their fixed income assets to hedge against interest rate risk. A majority of fixed income assets are held within riskless bonds and
4With a probability of 95 percent, refers to the confidence interval. Min and Max refer to the estimate of Moran’s I 5Deviances from average spatial correlations of total equity and total fixed income assets, in comparison to the sum of different risk categories originate from the fact that total equity and total fixed income assets contain also assets
variance across all pension funds in the sample is rather small. A relative low spatial correlation for AAA bonds due to congruent asset managers shows that alignment in riskless fixed income assets is relatively independent from the asset management structure. In contrast, allocation of riskless bonds is only weakly dependent on mutual asset management and more aligned across all Dutch pension funds. This relation is changing as credit worthiness of fixed income assets is decreasing. Allocation of fixed income assets with higher risks of default and simultaneously higher yields go along with higher correlation within mutually managed pension funds. On the one hand, these empirical results support the assumption that pension funds outsource their asset management due to third parties expertise to manage assets which involve higher risks. And moreover, asset managers consequently align the asset allocation in riskier assets for different pension funds in a common manner. On the other hand, pension funds might also choose a common asset manager which matches their risk profile best. Thus pension funds with a similar risk profile might also choose a common fiduciary asset manager.
6.2 Time Narrative
Second, Moran’s I shows an interesting variation over time. Although the time horizon is rather small, peaks in spatial correlation can answer the question, if asset managers might implement common investment strategies in order to react to different economic or regulatory circumstances. However, a decomposition of explanatory variables in the context of spatial correlation is difficult, since the correlation indicates how mutually managed pension funds ‘move’ together, but it does not include predicting or explanatory, independent variables. Accordingly, an interpretation why Moran’s I changes over time has to be treated with caution. Nonetheless, spatial correlation varies over time for different asset classes, providing room for further economic interpretations.
Portfolio Restructuring
The empirical results show that over the analyzed time horizon pension funds restructured their portfolio to a large extent. Therefore, pension funds’ asset class weights stayed rather stable on average, but within one asset class shifts in credit risk categories are observable. Whereas the total share of fixed income assets increased from the beginning in 2012 to the end of 2016 only about one percent, movements for risk levels of fixed income assets are more pronounced: In particular pension funds across the analyzed sample decreased their riskless fixed income assets, rated AAA, from 32 percent in the first quarter of 2012 to approximately 20 percent on average at the end of 2012.
This significant drop in riskless fixed income assets goes along with a decline of variance across all pension funds in the given sample and a relatively low spatial correlation value of 0.14 for riskless fixed income assets on average. These results suggest that a discharge of fixed income assets, rated as risk free, is clustered rather on a general level across all pension funds, than dependent on the fiduciary asset manager. Regarding investment decisions of riskless fixed income assets, a decline in cross-sectional variance shows an alignment of pension funds in their weighting of AAA bonds, but asset reallocation among pension funds is only slightly stimulated by mutual asset managers in that regard.
Figure 3. Reallocation and risk shift from AAA to AA bonds
Figure 4. Spatial correlation comparison AAA and AA fixed Income assets.6
A different picture can be drawn with reference to fixed income assets with slightly higher risk categories. Particularly bonds rated AA show an exemplary pattern for asset allocation alignment due to mutual asset managers. On average, pension funds tripled their allocation of fixed income assets, rated AA, from the third quarter in 2012 to the third quarter in 2014. In contrast to triple AAA rated bonds, the overall cross sectional variance rises significantly to values five times higher than in 2012, indicating that a
6 For reasons of clarity, the confidence interval of Moran´s I is not display in graphs which compare two different credit risk categories. A figure of the single time series of Moran´s I, including the 90 percent confidence interval is to be found in Appendix A 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Evolution of Fixed Income Assets, as Porfolio
Weights
AA Bonds AAA Bonds
-0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Moran's I
shift towards AA bonds differs across pension funds. Simultaneously, Moran’s I doubled from 0.2 in 2012 to approximately 0.4 in 2014. A rise in spatial correlation indicates that a shift towards slightly riskier bonds is dependent on the fiduciary asset manager. Consequently fiduciary asset managers seem to align their investment strategies in AA rated fixed income assets for different pension funds’ portfolios, which are under their management.
The analysis of AAA and AA rated fixed income assets show that asset reallocation among Dutch pension funds may take place on two different levels: Dutch pension funds divest riskless fixed income assets which generate low yields on a clustered but more general level. Simultaneously shifting the investments towards riskier assets is dependent on the fiduciary asset management. Hence one can conclude that divesting in the medium run rather takes place across pension funds, regardless of their asset management, whereas an alignment in investing towards alternative, more risky assets is dependent on pension funds overlapping asset management structures7.
The shift in asset allocation towards riskier assets can be linked to economic developments and pension funds’ urge to adjust their asset allocation strategies.
Especially the pronounced peak starting in the first quarter of 2013 on fixed income returns indicates common reaction to exposure in market developments. A rise in spatial correlation of fixed income returns, can be traced back to common investment decision of asset managers in riskier bonds, as spatial correlation of returns and asset allocation of bonds rated below BBB are rising simultaneously. The cease of the European sovereign debt crisis at the end of 2012 and a simultaneous deterioration in term structures might have caused clustered changes in risk aversion and a common restructuring of riskier fixed income assets among mutually managed pension funds. A closer look into asset allocation data indeed reveals, that pension funds moved in the beginning of 2013 systematically towards riskier fixed income assets. The risk free discount rate reached its historic low in the second quarter of 2012 and spatial
7 The general movement away from low return/risk assets incorporates all pension funds, regardless of their asset management structure. The question which alternative assets are chosen in order to achieve a reasonable return / risk exposure is then again relatively more dependent on the fiduciary asset manager. Higher spatial correlations indicate
correlation of allocation for bonds rated lower than BBB increased significantly in the preceding periods. This pattern might be explained by a reaction of asset managers, to move from lower to higher risk bonds for different pension funds’ portfolios in terms of a “search for yield”. A decline in returns on risk free rated bonds and low funding ratios can be seen as the driving forces to restructure pension funds’ portfolios towards riskier but higher yield fixed income assets. Fiduciary asset managers might seek to achieve the nominal returns they had been generating when interest rates were higher. Nonetheless, spatial correlation of risky and riskless fixed income assets declined from second quarter of 2014 onwards, indicating that bond allocation heterogeneity in mutually managed pension funds is decreasing. In addition significant difference between AAA bonds and bonds rated lower than BBB is falling simultaneously.
Figure 5. Fixed Income Actual Allocation - Risk Class Comparison
Changes in Regulation
The most recent and severe turning point in the regulatory framework for Dutch pension funds was the implementation of changes to the financial assessment framework (FTK). At the end of 2014 changes in indexation rules and recovery plans where adopted, which lead to necessary changes in the pension funds attitude in risk assessment (Hoeckert and Troost 2015). The agreement stated that funds “will have to determine and disclose their attitude to risk, which must be agreed with the social
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 12q 1 12q 2 12q 3 12q 4 13q 1 13q 2 13q 3 13q 4 14q 1 14q 2 14q 3 14q 4 15q 1 15q 2 15q 3 15q 4 16q 1 16 q 2 16q 3 16q 4 Mo ra n 's I
partners” (ibid.), compulsorily leading to changes in risk assessment and hence the pension funds investment policies. One central aspect of the new regulatory frame work (nFTK) was a onetime possibility to change their investment risk profile, under the premise that the pension fund’s funding ratio lies above the minimum required capital (MVEV). Reports of the DNB indicate that regulatory changes in the end of 2014 where indeed decisive: At least 41 out of 300 existing pension funds switched their risk profile after the implementation of the new FTK (DNB Database).
Spatial correlation shows two striking aspects, which can be linked to changes in regulatory environment in 2015. First, Moran´s I of total equity allocation drops significantly in the beginning of 2015, to a spatial correlation value of around 0.05, while the lower bound of its confidence interval turned negative. This drop is remarkable, since spatial correlation stayed rather stable from 2012 until 2014, highly significant above zero and on average above 0.2. While equity allocation in mature market shows a similar picture, equity in emerging markets is, with the exeption of one quarter, not as much affected. An extreme drop in the significance for spatial correlation indicates that valid inference for spatial heterogeneity in equity asset allocation due to mutual asset management cannot be drawn. In other words, one cannot reject the Null hypothesis that equity allocation follows a spatial random pattern across all Dutch pension funds, regardless of overlapping management structures, after the introduction of nFTK. Hence, congruence in mature equity allocations within mutually managed pension funds is not confirmable, after changes in regulation.
This might be explained due to a rise in variance of equity allocation across all Dutch pension funds. A closer look into data shows, that on average Dutch pension funds restructured their portfolio concerning their equity asset holdings, as a reaction to changes in regulation. From the last quarter of 2014 to the first quarter of 2015, average equity assets, measured as percentage of their total assets, drops sharply by five percent. This decline is remarkable, since the share of equity within pension funds´ portfolio stayed relatively stable of around 28 % in the previous periods. Simultaneously, the variance of equity holdings across all pension funds doubles within only one quarter. These results suggest that on the one hand pension funds shifted their equity allocation on average away towards alternative assets, but on the other hand this
process of restructuring differs significantly among all Dutch pension funds. An insignificant spatial correlation indicates, that the equity restructuring process might lie outside of the fiduciary managers’ mandate and might rather be determined by the pension funds changes in strategic policy.
Figure 6 and 7.: Changes after nFTK introduction in Q1 2015. Spatial correlation of equity in mature markets and evolution of cross sectional mean in equity weights and changes in cross sectional variance.
-0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Moran's I Equity Mature Markets
MI 90 % Confidence Interval 0.4% 0.6% 0.8% 1.0% 1.2% 1.4% 1.6% 1.8% 2.0% 15.0% 17.0% 19.0% 21.0% 23.0% 25.0% 27.0% 29.0% 31.0% 33.0% 12Q 1 12Q 2 12Q 3 12Q 4 13Q 1 13 Q 2 13Q 3 13Q 4 14Q 1 14Q 2 14Q 3 14Q 4 15Q 1 15Q 2 15Q 3 15 Q 4 16Q 1 16Q 2 16Q 3 16Q 4 Vari an ce Me an
Equity Mature Markets Cross Sectional Comparison
The second major change in the regulatory framework dates back to mid-2012. The introduction “September Package” introduced a so called “ultimate forward rate” (UFR), which is used to determine pension funds’ long term liabilities. The change incorporated a switch from a ‘swap curve interest rate’ to a pre agreed fixed interest rate to discount long term liabilities. As pictured in Figure 6, the interest rate increased for liabilities with a maturity of 20 years and longer, which consequently led to a decrease of liabilities present value. A decline of pension funds’ liabilities increased their funding ratio, which determines pension funds’ abilities to increase their risk profile.
Figure 8. Changes in discount rates, September Package in orange
As the market interest rate is typically characterized by higher volatility, the UFR also increased pension funds certainty in discounting long term assets. This change led to a decline of interest rate risk and therefore influenced asset allocation decisions of pension funds.8 By how far the September package led to clustered congruence in asset allocations of mutually managed pension funds shows the analysis of spatial correlation:
While spatial correlation of equity allocations stays rather stable, both in emerging and mature markets, Moran’s I of fixed income assets changes in the fourth quarter of 2012, possibly as a consequence due to regulatory changes. Spatial correlation of riskless
8 One can assume that the introduction of the UFR had predominantly an influence on pension funds decision to reduce durations of their fixed income assets. Nonetheless an increase in the discount rate certainly influences
bonds, rated AAA, drops from 0.15 to 0.10 in the third quarter and recovers to its initial value in the fourth quarter but Moran’s I for riskier fixed income assets is increasing significantly: Especially bonds with a credit rating of BBB shows high values for spatial correlation, relative to periods before the introduction of the September Package, as pictured in Figure 10.
Figure 9. Spatial correlation of BBB bonds
Hence one can conclude that mutually managed funds congruence in BBB rated fixed income assets increased after changes in regulation. Noticeably, spatial correlation of BBB rated bonds converges to values close to zero in the following time periods. This might stem from two different reasons which cannot be determined distinctively: Either differently managed pension funds aligned their allocation of BBB bonds in response to mutually managed pension funds in a time lagged manner, or mutually managed pension funds changed their allocation weights of BBB bonds, independent from their fiduciary asset manager. Yet, in both cases asset allocation congruence of pension funds which share a common fiduciary asset manager is declining.
-0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Moran's I BBB Bonds
MI 90 % Confidence Interval
6.3 Herding Behavior
As mentioned in Chapter V, the spatial correlation does not only allow one to measure congruence in mutually managed pension funds´ portfolios, it also serves as a method to analyze if pension funds trade in the similar direction at the same time, when they share the same fiduciary management. This can be done by taking the first differences of quarterly asset weights and estimating their spatial correlation values. It is worth recalling that Moran´s I measures a spatial correlation relative to the whole sample. Hence, spatial correlation states a value which can be interpreted as fiduciary herding additional to possible herding behavior across all pension funds, regardless of their responsible asset management corporation.
In comparison to spatial correlation of actual weights, Moran´s I of quarterly differences in asset weights shows three striking dissimilarities: First, average values of Moran’s I for quarterly differences in asset allocation are positive, but significantly smaller than spatial correlation values for absolute asset allocations. T-tests for differences in Moran’s I means indicate that the average Moran’s I is significantly higher for absolute values in comparison to means of spatial correlation of their first differences. These results are valid on a 99 percent confidence level and count for a comparison for assets of every credit risk class. As displayed in Table 4, these results are accompanied by the amount of time periods in which Moran’s I is significantly greater than zero, with a probability of 95 percent. Whereas spatial correlation values of absolute asset allocations are positive over almost the entire time horizon, positive spatial correlation values of first differences constitute rather exclusions. This accounts especially for assets which contain higher credit risks.
Second, a tendency of a higher average spatial correlation for riskier assets is, in contrast to spatial correlation of absolute asset allocations, not verifiable. This can be seen in multiple comparisons between fixed income assets for different credit ratings and also a direct comparison of fixed income and equity assets. On average, spatial correlation values for quarterly differences in equity allocation show notably small values and only few time periods of significantly positive spatial correlation values. In
contrast, fixed income assets rated AAA and AA are significantly positive over the majority of points in time.
Third, for the majority of asset categories, variance of spatial correlation over time is higher for quarterly differences than for spatial correlation of absolute allocation values. If, however spatial correlation of differences is significantly observable, Moran’s I reaches values which are in most cases on the same level and exceed in some cases also spatial correlation values of absolute asset allocations. This accounts for example for spatial correlation of Bonds rated BBB, which reaches a maximum value of 0.47, is however not observable for bonds with a credit risk rating lower than BBB.
Average Moran's I Variance Observations > 0 Min Max
Delta Bonds AAA 0.122 0.018 11 -0.076 0.471
Bonds AAA 0.141 0.002 18 0.078 0.232 Delta Bonds AA 0.134 0.018 13 -0.168 0.333 Bonds AA 0.243 0.009 19 -0.029 0.386 Delta Bonds A 0.041 0.003 4 -0.036 0.148 Bonds A 0.101 0.002 14 0.033 0.226 Delta Bonds BBB 0.101 0.012 9 -0.010 0.471 Bonds BBB 0.130 0.004 16 0.049 0.314 Delta Bonds <BBB 0.106 0.026 8 0.082 0.071 Bonds <BBB 0.206 0.005 20 0.104 0.343 Delta EQ Mature M. 0.026 0.003 3 -0.055 0.123 EQ Mature M. 0.162 0.008 13 0.037 0.271 Delta EQ Emerging M. 0.070 0.009 8 -0.076 0.252 EQ Emerging M. 0.199 0.003 19 0.019 0.263
Table 4. Comparison of first difference in asset allocation (fiduciary herding) and absolute allocation (fiduciary congruence).
