Hyper-Heuristics Using Genetic Programming to Time Series Forecasting

Time series forecasting methods allow companies and researchers to analyze and predict data that change over time, such as stock exchange and climate change. However, because of their complexity and dynamic nature, each type of time series ideally should be modeled using ad-hoc algorithms. To create a more general methodology, we proposed a combination of meta-heuristics, led by Genetic Programming (GP), to enhance the overall prediction ability. GP may not be as popular as the Box & Jenkins methodology for forecasting tasks, but the literature shows appealing outcomes. Swarm intelligence is also a powerful mechanism for searching patterns in large data spaces. Thus, we investigated and proposed a hybrid method using GP together with the Fish School Search (FSS) algorithm, where the latter is used to select optimal parameters for the former. We also used local search techniques for preventing the Genetic Programming to get stuck in local minima, by refining the coefficients on the GP expression. Our proposal was compared to standard autoregressive integrated moving average (ARIMA) model, exponential smoothing (ETS) and standard GP. The proposed method achieved promising results in one-stepahead predictions and was applied to a well-known time series data library.