Can a Neural Network Property Portfolio Selection Process Outperform the Property Market?

Journal of Real Estate Portfolio Management, May-Aug 2005 by Ellis, Craig, Wilson, Patrick J

Executive Summary. Evidence of the superior performance of portfolios comprised of 'value' stocks over 'growth' stocks is wide and varied. Despite this burgeoning literature, relatively little is known about the comparative performance of property sector value stocks and the performance of neural network techniques in relation to this market sector. This study addresses both of these issues by applying neural network modeling techniques to the Australian property sector stocks to construct a variety of value portfolios. Risk-adjusted performance measures show that the value portfolios outperform the market by as much as 7.14%.

Introduction

Artificial Neural Networks (ANNs) have a long development history, having been studied since Rosenblatt first applied single layer perceptrons to pattern-classification learning in the late 1950s (Kantardzic, 2003). Despite this long history, it has only been since the advent of faster computers that ANNs have gained wider acceptance. In ANNs, the network is presented with data repeatedly, from which it extracts the key relationships underlying the data. A valuable aspect of neural networks is that they are well suited to deal with unstructured problems, inconsistent information, missing data, and real-time output (Hawley, Johnson and Raina, 1990). There has also been a growing interest in the use of ANNs in finance and economics because of their capacity to imitate nonlinear relationships that are otherwise difficult to identify and also because the use of ANNs require no assumptions about the distribution of the underlying data.

The aim of this research is to ascertain whether it is possible to develop a neural network model capable of building a property investment portfolio that will outperform the market. While the broad objective is to ascertain whether ANNs represent a viable practical tool for portfolio allocation, a more specific objective is to determine whether an investment platform established by a neural network model provides a suitable basis for 'what if type simulations. While it is generally accepted that ANNs are not capable of being used for policy analysis (Brooks and Tsolacos, 2003), there is no reason to assume they cannot be used for 'what if simulations and planning for contingent scenarios generated by changes in the input variables. However, before such a step can be taken it is necessary to answer a basic question: Can a neural network constructed portfolio outperform the market?

While much is already known about the relative performance of value stocks in general, relatively little has been written about property sector value stocks specifically. Likewise, little is known about the performance of neural network techniques in relation to this market sector. Unlike most other industry sectors, property sector stock's primary income is derived from property investments and rental returns. Property trust units are underwritten over the long-term by their corresponding underlying direct property assets, yet over the shortterm property trust units often behave in a similar manner to stocks and very much unlike direct property (Webb, Seiler and Meyer, 1999). If this is the case, then the question naturally arises as to the extent to which units in property trusts have comparable concepts of value and growth to those that exist in the stock market (i.e., value stocks vs. growth stocks), and the extent to which portfolios composed of value units will outperform the property market.

Following from Eakins and Stansell (2003), a select group of fundamental financial ratios will be used to determine the set of 'value' assets to enter the property portfolio. Not surprisingly other research has used one or more such fundamental variables in conventional time series models to predict stock market returns. For instance, Kothari and Shanken (1997) in a study on stock market returns over the period 1926 to 1991 found that both book-tomarket and dividend yield tracked time series variations in expected real stock returns. In the current research, these variables will form the inputs to a neural network model that will have preset limits determined by the literature as indicated later to isolate Value' from 'growth' assets. Value assets are commonly defined as those whose market value is lower than their intrinsic or liquidating value (O'Shaughnessy, 1998:2). The attraction of value assets from an investor's point of view is that the (lower) market value of the asset should rise to meet the (higher) intrinsic value. This being true, portfolios comprised of value assets only, should outperform portfolios comprised of all assets. Value stocks are typically characterized by low market price per share, low cash flow per share, or low book value per share.

Portfolio allocations to value stocks is sometimes referred to as a contrarian investment strategy and a number of studies have indicated that such a strategy has outperformed the strategy of investing in 'growth' stocks (see Fama and French, 1996, 1998; Haugen, 1999; Lakonishok, Shleifer and Vishny, 1994; Levis and Liodakis, 2001; and Yen, Sun and Yan, 2004 as indicative studies). Using stochastic dominance tests, Best, Best and Yoder (2000) confirmed the dominance of value portfolios comprised of high book-to-market stocks over low book-to-market portfolios. These findings all contradict the basic tenets of the efficient markets hypothesis (EMH). Others, however, suggest that the higher returns to value stocks are simply compensation for higher risk (Chen and Zhang, 1998; and Davis, Fama and French, 2000), or that growth stocks actually outperform value stocks in the long run (Beneda, 2002). Given this wealth of evidence, this paper seeks to answer two questions: can a neural network construct a 'value' portfolio, and will this contrarian strategy outperform the market?