Forecasting financial time series with artificial neural networks

<div>Financial forecasting has always been one of the biggest challenges</div><div>in theoretical and practical Finance. Because of the speci\x0cc nature</div><div>of most \x0cnancial time series, which tend to be non-Gaussian, non-</div><div>stationary, chaotic and extremely noisy, most traditional models fail</div><div>to accomplish the task of e\x0ecient data \x0ctting and forecasting and,</div><div>therefore, need to be replaced by a more powerful analytical tool.</div><div>Recently, there has been a growing belief that Neural Networks, which</div><div>are used e\x0bectively is pattern recognition and classi\x0ccation, might as</div><div>well be a panacea for at least some problems in \x0cnancial data analysis.</div><div>In this paper we apply feed-forward multilayer Neural Network to</div><div>forecast stock prices of \x0cve largest companies in Russia and evalu-</div><div>ate the possible gains and losses of trading using our Network. We</div><div>\x0cnd that, despite the outstanding ability of the Network to approx-</div><div>imate the data, the standard MLP is unable to generate abnormal</div><div>pro\x0cts to an investor when it is used to forecast one-month-ahead</div><div>stock returns. Moreover, we \x0cnd that increasing the running time of</div><div>the cost-minimization algorithm does not necessarily improve out-of-</div><div>sample performance of the Network.</div>