potentially exist in the financial time series. This is also one of the reasons why many
machine learning techniques are gaining popularity in modeling financial time series
(Kazem et al., 2013; Matías and Reboredo, 2012; Yu et al., 2009).
Despite its advantages, the ANN also carries some drawbacks, such as overfitting issues
(Tu, 1996). For instance, if the ANN estimation produced a function that is too closely fitted
to the training data set, it could lead to substantially large errors when making predictions
through a new set of observation. Therefore, it is helpful to compare the ANN method
against the regression method so as to determine whether the ANN is outperforming.
There are estimation parameters in the ANN method –it has to be pre-determined before
it can run the neural network estimation, such as the number of hidden neurons, hidden
layers, learning rate and others. There are no fixed rules on how to determine a
configuration of the ANN that could produce the best forecast. Although it has been said
that one single hidden layer is enough to explain the relationship between the variables, this
is on the condition that sufficient numbers of hidden neurons were used (Hornik et al., 1989;
Masters, 1993). Nonetheless, it is still worth examining how a neural network with different
numbers of hidden layers, and hidden neurons perform, given that the sufficient number of
hidden neurons remained uncertain.
This study was performed for the purpose of determining the forecasting performance of
the ANN on the Hong Kong REITs’(HK-REITs) returns. Hong Kong is one of the most
dynamic commercial property markets in Asia (Newell et al., 2010). The HK-REITs carry a
certain attractiveness for investors. For instance, the GZI REITs provide investors with an
exposure to the China property holdings while the Link REITs serve as one of the biggest
retail REITs in Asia (APREA, 2009). As each HK-REIT has different asset holding
compositions in theirportfolio, it would be worthwhile to investigate how the accuracy could
be different across theHK-REITs, and also how the results can be differentacross time since
the portfolio holding the REITs can changeover time. Therefore, it is beneficialto investigate
whether the same ANN method can produce the same forecasting performance across
different HK-REITs when it involved a variety of asset holdings among the HK-REITs.
This study used two common models to predict the HK-REIT returns –the CAPM, and
Fama and French three-factor model. The CAPM was chosen due to its capability to
explain the relationship between the risk, andexpectedreturnswithonlyonefactor.
However, one factor might not be enough to explain all the variations of the excess returns
(Fama and French, 2004). Past studies (Chiang et al., 2005; Peterson and Hsieh, 1997) have
shown that Fama and French three-factor model is better than the CAPM in explaining
the EREITs returns. Thus, Fama and French three-factor model was also employed in
Further to the above, the accuracy of the ANN forecast when used with different
configurations of parameters on the models were also compared with the regression method.
The comparison was made based on all the HK-REITs involved and also based on
individual HK-REITs separately. Each set of comparison was further split into the annual
time series so as to investigate the consistency of the performance across time. Finally, each
result will be compared against the buy and hold strategies so as to determine their returns
performance. This study employed the rolling sample approach in calculating the forecast
because it takes the latest available data in calculating the forecast. This mimics the
practical situation since investors were more likely to use the most recent data from time to
time to predict the nearest future returns of the REITs rather than doing a one-off prediction
to obtain all the forecast. Finally, each result will also be compared against the buy and hold
strategies so as to determine their returns performance. As contribution, the current study
highlights the importance of testing a technique first before applying it, and to seek the
truth of the technique via further testing rather than merely drawing conclusions based on