Financial sequences and the hidden Markov model

Abstract

Our aim is to develop algorithms for learning noisy sequences of symbols taken from discrete, finite alphabets, i.e. predicting the next symbol from the preceding sequence. With this in mind, financial sequences of a five-letter alphabet representing sharp fall, slight fall, no change, slight rise and sharp rise in the stock prices of five technology companies were derived from the daily closing price between January 2008 and May 2011 on the NASDAQ exchange. Accuracy of baseline predictions from a probabilistic analysis was compared with the accuracy of a hidden Markov model. Probabilistic analysis shows clear evidence of non-stationarity of the underlying time series. The hidden Markov model accounts for some but not all non-stationarity. It is argued that an analysis based on contextual probability is expected to outperform the hidden Markov model. © 2012 Springer-Verlag.