The problem of forecasting financial time series has received great attention in the past, from both Econometrics and Pattern Recognition researchers. In this context, most of the efforts were spent to represent and model the volatility of the financial indicators in long time series. In this paper a different problem is faced, the prediction of increases and decreases in short (local) financial trends. This problem, poorly considered by the researchers, needs specific models, able to capture the movement in the short time and the asymmetries between increase and decrease periods. The methodology presented in this paper explicitly considers both aspects, encoding the financial returns in binary values (representing the signs of the returns), which are subsequently modelled using two separate Hidden Markov models, one for increases and one for decreases, respectively. The approach has been tested with different experiments with the Dow Jones index and other shares of the same market of different risk, with encouraging results.