A Meta-Learner Approach to Multistep-Ahead Time Series Prediction

Abstract

The application of machine learning has become commonplace for problems in modern data science. The democratization of the decision process when choosing a machine learning algorithm has also received considerable attention through the use of meta features and automated machine learning for both classification and regression type problems. However, this is not the case for multistep-ahead time series problems. Time series models generally rely upon the series itself to make future predictions, as opposed to independent features used in regression and classification problems. The structure of a time series is generally described by features such as trend, seasonality, cyclicality, and irregularity. In this research, we demonstrate how time series metrics for these features, in conjunction with an ensemble based regression learner, were used to predict the standardized mean square error of candidate time series prediction models. These experiments used datasets that cover a wide feature space and enable researchers to select the single best performing model or the top N performing models. A robust evaluation was carried out to test the learner's performance on both synthetic and real time series. 

Proposed Dataset

The dataset proposed here gives the results for 20 step ahead predictions for eight Machine Learning/Multi-step ahead prediction strategies for 5,842 time series datasets outlined  here. It was used as the training data for the Meta Learners in this research. The meta features used are columns C to AE. Columns AH outlines the method/strategy used and columns AI to BB (the error) is the outcome variable for each prediction step. The description of the method/strategies is as follows:

Machine Learning methods:

NN:   Neural Network

ARIMA: Autoregressive Integrated Moving Average

SVR: Support Vector Regression

LSTM: Long Short Term Memory

RNN: Recurrent Neural Network

Multistep ahead prediction strategy:

OSAP: One Step ahead strategy

MRFA: Multi Resolution  Forecast Aggregation