Expert systems for trading signal detection have received considerable attention in recent years. In financial trading systems, investors' main concern is determining the best time to buy or sell a stock. The trading decisions are often influenced by the emotions and feelings of the investors. Therefore, investors and researchers have aimed to develop systematic models to reduce the impact of emotions on trading decisions. Nevertheless, the use of algorithmic systems face another problem called "lack of dynamism". Due to dynamic nature of financial markets, trading robots should quickly learn and adapt as human traders. Recently, a solution for detecting trading signals based on a dynamic threshold selection was proposed. In this study, we extend this approach by adopting several different rule based systems and enhancing it by using the Recurrent Neural Network algorithm. Recurrent Neural Networks learn the connection weights of subsystems with arbitrary sequences of inputs that make them a great fit for time series data. Our model is based on Piecewise Linear Representation and Recurrent Neural Network with the goal of detecting potential excessive movements in noisy stream of time series data. We use an exponential smoothing technique to detect abnormalities. Trading signals are produced using fixed time interval data from Istanbul Stock Exchange. The evaluations indicated that our model produces successful results in trading data. Future work will focus on further improvements and scalability of the model.