INTRODUCTION

In recent years the competition between mutual fund providers and those entering the ETF sub market has become increasingly competitive [1 ]. Competition can most clearly be seen in the expense battle between large financial institutions such as Vanguard and BlackRock. The underlying reason for reducing the expenses of the funds is to attract investors. The structure of an Exchange Traded Fund is slightly different from your traditional mutual fund because its main objective is to meet the performance of an underlying index. There has been little research into why an investor chooses one ETF over another, especially those that have similar holdings and track like indices. Knowing the characteristics of a fund that are most important to investors can be used in marketing techniques and the development of new funds that highlight favorable characteristics. Having the right combination of key fund characteristics is expected to bring inflow of shares and increase the total net assets of the fund. For a financial institution developing and marketing these financial products, the goal is to increase net assets year over year. If a fund's total net assets increase over time, they are deemed to be profitable investments. For those funds that lack to attract positive share inflows financial institutions backing such products are faced with closing the funds and thus never realizing a return on their investments.

Most academic research has been focused on determining the flows of mutual funds through past fund performance based on returns earned. In the studies reviewed by Nanigian [2 ] fund performance has been explained to have a linear relationship with fund expenses with an increase in expenses having a negative impact on fund performance. Yet, Nanigian [2 ] believes these recent studies in regression analysis are assuming a linear relationship that in actuality is far from linear and more complex. This supports the need for a method of machine learning that can discover relationships that are not linear in nature. Cooper et al performed a study on the effects of recent name changes on the net flows of a fund. In this study Cooper et al found results through the use of regression techniques, indicating funds experienced an abnormal inflow of funds just after a recent name change proving there is more to fund flows than just performance [3 ].

ETFs are developed to meet the performance of an underlying index and not necessarily outperform the market, and thus their attractiveness to investors is not of mere performance of the fund but its broader set of characteristics. Data mining approaches can be of significant use in determining why a particular ETF attracts more investors than others by using machine learning to bring to the forefront linkages and correlations of fund characteristics that have yet to be documented. Therefore, this study is aimed at developing a data mining-based methodology to determine the prevailing factors of an attractive ETF.

The rest of the paper is structured as follows. Section 2 briefly summarizes the recent and relevant literature in the field. Section 3 describes the main steps in our proposed data mining methodology. Section 4 presents the modeling results and discusses the findings. The paper concludes with Section 5 where a summary of the study and its implications are discussed.

SECTION II

LITERATURE REVIEW

Past studies in academia have focused their efforts on fund flows as a result of one or two independent variables. These studies have used linear regression models to explain the relationship between variables and fund flows. When multiple variables were considered in a study they were conducted separately and not together in their analysis suggesting that each is independent of one another with no influence of one variable on the relationship of another. Sirri et al. [4 ] conducted a study over a 30 year period to determine the relationship between fund performance and fund flows. They also looked at the relationship between fund flows and fund's provider as a source of influence on flows. The results indicated that past fund performance has a positive relationship on fund flows. It also indicated fund provider membership in well-advertised campaigns increased fund flows. Fund performance as a determinant in fund flows has been extensively studied by academics and its positive relationship is not only supported in Sirri et al. studies but of many others. Additional research done by Cashman et al. [5 ] has similar results to Sirri et al. The Cashman et al. study [5 ] shows how investors rely on performance as a determining factor for a buy or sell decision. A study conducted by Cha et al reports much of the same results. They concluded in the overall US market, investors tend to put their money into funds based on past security performance [6 ].

Cici [7 ] conducted a study on US equity mutual funds to see if there is a tendency for investors to sell their mutual fund holding in times of positive performance and hold onto the funds during times of negative performance. Cici [7 ] called this the disposition effect in which investors tend to hold onto losses and realize gains. Cici [7 ] compares findings with that of Grinblatt et al. [8 ] using logit regression and concludes the same with past performance being a contributing factor in the flow of shares. Supporting Sirri et al is field work done by Hoffmann et al [9 ]. Their study sheds new light on the psychological factors influencing the prevalence of mutual funds flows and fund performance. The study reveals investors commonly perceive previous positive fund performance will continue into the future and thus they hold on to mutual funds with past strong performance and dump those with negative or weak past performance. This helps to explain why past fund performance has a positive relationship on fund flows.

Gottesman et al. [10 ] explore the relationship between investor behavior and fund flows in up and down markets. Gottesman et al. study [10 ] actively managed mutual funds and the net flow of funds in and out of the funds during economic downturn and economic upturns. The conclusion was investor behavior influences the net flow of funds disproportionately in up versus down markets. In down markets investors are less likely to contribute to inflows even if the fund out performed its peers than if it were an up market. Again this supports the work of Hoffmann et al indicating investor behavior is a key attribute in predicting future fund flows and needs to be considered in the development of a more comprehensive study on fund characteristics.

Daily fund flow volatility and its effect on fund performance were explored by Rakowski [11 ]. Rakowski [11 ] conducted a study on US open-ended mutual funds to see if there is a correlation between daily fund flow volatility and performance [10 ]. Weak performance was due to the increase in trading and expenses incurred as a result of fund volatility. Huang et al also studied fund volatility. Huang et al argued that fund volatility affects fund inflows negatively as investors view fund volatility as a detriment to predicting future fund performance [12 ].

Bollen [13 ] takes a targeted approach to fund flows by segmenting the US mutual fund market and focuses on the socially responsible mutual fund niche. Bollen [13 ] uses a flow performance regression technique to assess the volatility of fund flows for a sample of socially responsible mutual funds. The results indicate that socially responsible funds are more resistant to the fund flow volatility than conventional funds. These results indicate that the type of fund and its meaning to investors has a correlation to overall fund flows. Therefore one can extrapolate there are more attributes to a fund that affect the fund flows than that of just past performance. It brings to the surface the need to do further research on the subject of fund flows and to broaden the variables tested. One of the best ways to determine variables effecting fund flows aside from past performance is through the use of funds that do not exist to outperform the market. ETFs are such funds designed in this way and thus are an excellent source to test.

A slightly different approach to the influences on fund flows is conducted in a study by Kempf et al. [14 ]. The study takes into consideration the family of funds a mutual fund resides in as a predicting factor in fund flows. It also speculates that the position within the fund family also has influence on fund inflows. In order to create a model which tests only the variable of fund family and position Kempf et al. [14 ] attempt to control fund characteristics such as expense ratios, fund size, and age. While it creates a more focused test it does dispel the need to test fund characteristics as determining factors in fund flows. As a result, Kempf et al's research only partially explains fund flows because they build their model purposely to leave out these factors.

Interesting research on the topic of fund characteristics to fund performance was conducted by Fan et al. The study took fund characteristics such as expense ratios, management tenure, and fund age and compared them to the resulting fund performance [15 ]. Popular beliefs would lead investors to believe these three fund characteristics would have an impact on fund performance but the study findings suggested they were irrelevant. Berk et al conducted a study on mutual fund flows and performance in rational markets [16 ]. Of specific concern was the ability of active mutual fund managers to outperform the market. Their test results support the use of actively managed funds by superior fund managers and the tendency of investors to put their money in funds that outperform their peers. Yet, Berk et al. strived to explain the investor behavior as rational versus irrational behavior. This indicates investors are actively researching and making investment choices that affect fund flows based of fund characteristics such as past performance.

How broker incentives affect inflows is a question addressed by Christoffersen et al. [17 ] in a study conducted to determine if the incentives of brokers truly influenced the inflows of affiliated funds. Logic would expect the results of such a test to reveal that incentives do influence fund flows. Christoffersen et al. [17 ] used time series regression to test their hypothesis. They concluded that flows to affiliated brokers are less sensitive to the fees and or commissions of the particular affiliated fund compared to unaffiliated counterparts. These results indicate that broker incentives are an attribute to be considered in predicting future fund flows with brokers receiving above average incentives for a fund which in turn directly influences fund flows.

Barber et al. cross section regression shows investors do make investment decisions based on different types of fund expenses [18 ]. In general investors tend to favor mutual funds with lower commissions and up-front fees but they do so because of how the fees are presented to the investors. Commissions are more readily identifiable by investors and such are more influential in the decision on mutual fund selection creating more inflows to funds with lower commissions than alternatives with higher commissions.

The majority of these studies concluded that past performance is a key contributor in predicting future fund flows as investors tend to trade based off past performance. While this is true for traditional mutual funds ETF's are a different animal and past performance is expected to be less of an influence because the objective of the ETF funds is to not outperform the market. As a result the recent research done on volatility, expenses, fund family, and type of underlying securities of the mutual fund appear to be fund characteristics that influence investor behavior and the ultimate net flow in or out of a fund. Therefore this study is aimed at performing a more comprehensive review of fund characteristics and how they relate to net fund flows of ETFs.

SECTION III

METHODOLOGY