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

Forecasting

Donaldson and Kamstra (1996) studied the outcome of combining forecasts of stock market volatility across a range of countries. These authors show that combining forecasts with nonlinear neural networks produces forecasts that routinely dominate forecasts from traditional linear techniques such as moving average variance and GARCH models.

In an interesting experiment, Hill, O'Connor and Remus (1996) compared the outcomes from the original Makridakis forecasting competition (Makridakis et al., 1982) with the outcomes using a neural network on 'level ground' (i.e., the experiment was conducted as if the authors were part of the original competition). The neural network was found to perform significantly better than traditional time series methods when forecasting with monthly and quarterly data, and was comparable with annual data. The authors suggested that the crucial reason for the superior performance was the ability of neural networks to better cope with discontinuities.

In a study of the efficacy of neural networks in predicting returns on stock and bond indices, Desai and Bharati (1998) found that the neural network forecasts were conditionally efficient with respect to linear regression models for large stocks and corporate bonds, but there was no significant difference for small stocks and intermediate-term government bonds. In a study with mixed results, Narain and Narain (2002) compared stock market predictions using a multivariate statistical (MVS) model and a neural network. These authors found that the MVS model was able to predict the S&P 500 and NASDAQ indices just as well as the ANN, while the ANN model was able to predict the DJIA better than the MVS.

Walczak (2001) considered the question of the optimum amount of data to train a network model for financial forecasting. The author examined three common exchange rates: dollar/pound, dollar/mark and dollar/yen. He found that neural networks trained on a larger training set have a worse forecasting performance. Yu (1999) used a three-layer, backpropagation network in a comparison of the forecasting performance of a neural network and a conventional ARIMA model and found the neural network outperformed the ARIMA model in forecasts of the Nikkei Stock Index futures. Similarly, Kanas (2001) used a three-layer network in a forecast comparison with a conventional linear model. There were six neurons in the hidden layer of the neural network while the linear model used lagged percentage change in trading volume and lagged percentage change in dividends to predict stock returns for the Dow Jones and the FTSE. Kanas found that the neural network forecasts were preferable to the linear model in that the network could explain the forecast errors of the linear model, but not vice-versa. This indicated that the inclusion of nonlinear terms in the functional relationship between returns and explanatory factors is important in forecasting.

Research Methodology

The objectives of the research in this study are twofold. First, the performance of neural networks and their ability to identify 'value' stocks is examined from the set of all property sector stocks (individual property trusts) listed on the Australian Stock Exchange (ASX). Once identified, an evaluation of the performance of portfolios comprised of property sector value stocks versus the set of all property sector stocks is conducted. Portfolio performance relative to the market index is measured by the Sharpe ratio (Sharpe, 1966) for risk-adjusted returns, and the Sortino procedure (Sortino and Forsey, 1996; and Sortino, Miller and Messina, 1997) for adjusting returns on a downside risk basis.