TIME-SERIES FORECASTING USING FEATURE BASED HYBRID APPROACH
Date
2019-08-02T17:52:38Z
Authors
Adebimpe, Olashile
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In this research, we explored the unique strengths of both traditional statistical method and machine learning algorithms to propose a hybrid forecasting system based on a decomposition approach, with the objective of improving forecast accuracy across multiple forecasting horizons. Our proposed methodology uses the Seasonal and Trend decomposition using Loess decomposition procedure to break down time-series data into trend-cycle, seasonal and covariance stationary components, where these components produce individual forecasts and these forecasts are aggregated back to whole using an aggregation procedure, with the sole purpose of minimizing errors. Various types of Exponential Smoothing algorithms were employed for the trend-cycle, seasonal components because of its unique weight combination approach while vectors of features were extracted automatically from the nonlinear covariance stationary subseries to create appropriate Machine Learning models.
Description
We carried out our research using NN3 dataset and a large subset of 48,000 real-life monthly time series used in the M4 competition, which is characterized by considerable seasonality, trend and a fair amount of randomness so as to cover a wide range of time series structures. Our result reveals that the combination of decomposition, Exponential smoothing, Machine learning methods, and feature extraction gives less forecasting errors when compared to other combinatory approach and benchmark classical approach.
Keywords
Time series, Forecasting, Forecast combination, Time series features, Hybrid Approach