Time series prediction based on lazy learning

Lazy learning is a kind of novel machine learning methods based on statistical learning theory, which based on memory learning strategy. In the literature, it is generally used for non-linear system identification and function estimation. This paper applies lazy learning to time series prediction. Unlike conventional time series similar analysis, the whole similarity and the individual similarity are discussed. A new similar criterion combined the two similar characters is advanced. Using this criterion and locally weighted learning, one-step-ahead predictors for time series forecasting is achieved. For each single one-step-ahead prediction, the best predictive value will be obtained based on leave-one-out cross validation. In order to show the effectiveness of our method, we present the results obtained on a real-world dataset from the Santa Fe competition and Henon map.