Time Series Prediction by Neuro-Complex Fuzzy Systems

Time series are observations recorded sequentially over time. Internet traffic, environmental sensor data, online transaction sequences and stock quotations from financial markets are examples of time series. They are an... [ view full abstract ]

Time series are observations recorded sequentially over time. Internet traffic, environmental sensor data, online transaction sequences and stock quotations from financial markets are examples of time series. They are an important sub-category of data streams in which the data is not only temporally ordered, but the exact time of an observation is also recorded (explicitly or implicitly). Forecasting time series is a high-value machine learning problem. For example, internet traffic prediction aids network management; denial of service attacks and SPAM can also be detected as variations from the forecast. Precipitation forecasting is effective in water resource and drought management. Sales forecasts can be used to optimize inventory management and thus reduce costs. Stock market price forecasts can guide trading strategies covering tens of trillions of dollars’ worth of assets. Complex fuzzy sets (CFS) are an extension of type-1 fuzzy sets in which the membership grades are complex numbers (drawn from the unit disc in the complex plane). Various machine learning algorithms based on a combination of complex fuzzy sets and logic (CFS&L) and neural network have been proposed for time series prediction. The main disadvantage of these systems is slow learning algorithm because of the back propagation neural network implemented in their architecture. In this research, we propose a novel machine learning algorithm based on CFS&L by taking advantage of neural network with randomized learning algorithm in which they update their parameters just in forward pass. The new architecture is suitable for stream data mining.