This paper tries to see the performance of artificial neural network in predicting daily and weekly stock returns in both short and long forecast horizons. This paper also compares performances of neural network with those of linear autoregressive and random walk models in four different forecast horizons. Root mean square and sign prediction are used as two performance measures. From the results, we find that neural network performs better in the long run than in short run in terms of root mean square in the out-ofsample forecasting of daily stock returns. However, it is found that neural network’s performance becomes worse, in terms of correct sign prediction, as the forecast horizon increases. Neural network has superior out-of-sample performance in predicting daily stock returns than linear autoregressive and random walk model under all forecast horizons in terms of both root mean square and correct sign prediction. We do not find a very clear forecast horizon effect on neural network’s out-of-sample performance in terms of root mean square and sign prediction in predicting weekly stock returns. Neural network gives better out-of-sample forecasting of weekly stock returns, in terms of root mean square error, than random walk in longer horizons than shorter horizons. Neural network is also found to give better out-of-sample forecasting of weekly stock returns than linear autoregressive model, in terms of sign prediction, in short forecast horizon than long forecast horizon.

Stock returns are described as nonlinear, heteroscedastic and leptokurtic. Forecasting of stock returns is difficult because of the above mentioned three characteristics. A great deal of effort has been devoted to developing systems for predicting stock returns in the capital markets. Limited success has been achieved. It is believed that the main reason for this is that the structural relationship between an asset price and its determinants changes over time. However, using linear models Chen et al. (1986), Campbell (1987), Fama and French (1988, 1989), Whitelaw (1994) find that stock returns are predictable by publicly available information such as time series data on financial economic variables. Their findings go against the efficient market hypothesis and random walk model, which advocate for the unpredictability of asset returns. More importantly, Hinich and Patterson (1985), Abhyankar et al. (1997) have indicated the presence of structural nonlinear dynamics. Pesaran and Timmermann (1994) find significant nonlinear effects in quarterly and monthly regression of excess stock returns on economic variables and some non-linear terms such as the lagged values of the square of returns.

Artificial Neural Network (ANN), a nonlinear and non-parametric model, has some advantages over other linear and nonlinear models which make it attractive in modelingand forecasting of stock returns. First, neural network has flexible nonlinear function mapping capability which can approximate any continuous measurable function with arbitrarily desired accuracy (Hornik et al., 1989; and Cybenko, 1989), whereas most of the commonly used non-linear time series models do not have this property. Second, being a nonparametric and data-driven model, neural network imposes few prior assumptions on the underlying process from which data are generated.