Abstract
Machine learning based time series forecasting methods are popular and can match the performance of statistical models, in terms of accuracy, scalability, speed, etc. This disclosure presents techniques that incorporate statistical modeling into a neural network framework. The hybrid time series forecasting model described herein is named Seasonality Trend AutoRegressive Residual Yeo-Johnson power transformation Neural Network (STARRY-N). STARRY-N combines the advantages of residual neural network structure (such as N-BEATS) and explainable statistical forecasting models (such as TBATS). The model utilizes a neural network structure with separate stacks for trend, power transformed trend, seasonality, residual correction, and covariate adoption such as holiday effects. STARRY-N has good accuracy and is an explainable forecasting model.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Lichtendahl, Kenneth C. Jr.; Wu, Haoyun; Shen, Weijie; Choi, Chi Po; and Fry, Chris, "A Statistical Decomposition Based Neural Network For Multivariate Time Series Forecasting", Technical Disclosure Commons, (October 28, 2020)
https://www.tdcommons.org/dpubs_series/3707