Selecting Peer Groups

In mutual fund analysis, peer group construction often plays a key role.  Many ways exist to construct peer groups.  In this article, we focus on a way of constructing peer groups for board or 15(c) reporting.   These peer groups can often be difficult to construct given the number of variables or dimensions a board needs to review for 15(c) purposes.  A small list of these factors includes but are not limited to: performance, management expenses, advisor fees and total expenses.

A significant part of the 15(c) process is contract renewal.  As contract renewal involves fees and performance data, a board may want a peer group that is consistent across fee and performance comparisons.  The peer group needs to be multi-faceted – you can turn it one way when looking at advisor fees and turn it another way when comparing performance.

To demonstrate how this multi-faceted peer group can be constructed, we will compute all the distance measures from each fund to every other fund in Lipper’s large-cap core equity (LCCE) classification.  To demonstrate the effectiveness of using distance as a way of forming peer groups, we construct a peer group for the American Funds American Funds Investment Company of America (ticker symbol AIVSX).  We choose this particular fund purely for the reason that it is the largest fund in the LCCE classification based on AUM.

We use the following factors to compute the distances between all the LCCE funds and AIVSX: annual returns for full year 2010, 2011, 2012; AUM for end-of-year 2012; advisor expenses for the full-year 2012, management fees for full year 2012 and total expenses for full-year 2012.

The method used to compute distances between funds is dissimilarity indices.  The reader is referred to http://cran.r-project.org/web/packages/vegan/vegan.pdfl for a brief introduction to dissimilarity indices and their uses.

In the figure below we show the closest funds (by distance) to AIVSX:

Figure 1: Peer Group for AIVSX

AIVSX is the fund marked in blue.  The funds in green are the best “fit” peer group for AIVSX while those in yellow are close companions (you could call them “cousins” to AIVSX).  It bears noting that all the remaining funds 237 LCCE funds (primary share class only) are a significant distance away from AIVSX and would not be the best funds to include in the peer group.

If we remove the two American funds from our green-colored list, there are three “true” peers to AIVSX:  Davis New York Venture, Managers AMC Yackman Fund and the Fidelity Strategic Advisors Core Fund.

The author draws the readers’ attention to the similarity and dissimilarity in the annual returns across 2010, 2011 and 2012.  If you look purely at the value of the returns in 2010 and 2012, you might choose the funds marked in yellow as the best peer group members, given the similarity of their returns to AIVSX.  2011 is a different story.  You might choose some of the green and yellow funds as peers for AIVSX.  Therefore when looking at returns, you get a mixed bag of peer group members for AIVSX.  This would also be the case if you used correlations, another measure of distance, to determine peers.  The author highlights this mix and match approach as part of the limitations involved in many peer group construction methodologies.

We show below, the “collapse” of our seven-dimensional, i.e. seven factor, distance measures.  They are “dropped” or folded down to three dimensions:

Figure 2: Three-Dimensional Peer Group View

Each pink dot represents one of the 247 funds.  As the reader can see, many of the funds fall on a nice line, which indicates forming a peer group will be easy for these funds.  The funds on the right are scattered in both two and three dimensions and represent funds where peer group construction will be hard.

In this article we show an alternative way to construct peer groups when looking across multiple factors simultaneously.  We do this by computing distances between each of the funds.  In our next article, we will look at absolute return funds and see if our distance computations work there as well.