Applying Machine Learning Methods for Time Series Forecasting

This paper describes a strategy on learning from time series data and on using learned model for forecasting. Time series forecasting, which analyzes and predicts a variable changing over time, has received much attention due to its use for forecasting stock prices, but it can also be used for pattern recognition and data mining. Our method for learning from time series data consists of detecting patterns within the data, describing the detected patterns, clustering the patterns, and creating a model to describe the data. It uses a change-point detection method to partition a time series into segments, each of the segments is then described by an autoregressive model. Then, it partitions all the segments into clusters, each of the clusters is considered as a state for a Markov model. It then creates the transitions between states in the Markov model based on the transitions between segments as the time series progressing. Our method for using the learned model for forecasting consists of indentifying current state, forecasting trends, and adapting to changes. It uses a moving window to monitor real-time data and creates an autoregressive model for the recently observed data, which is then matched to a state of the learned Markov model. Following the transitions of the model, it forecasts future trends. It also continues to monitor real-time data and makes corrections if necessary for adapting to changes. We implemented and successfully tested the methods for an application of load balancing on a parallel computing system.