Premium
Business Forecasting of Double‐trend Time Series: An Improved PLS‐based Time‐varying Weight Combination Approach
Author(s) -
Luo Biao,
Wan Liang,
Li Tieshan,
Liang Liang
Publication year - 2018
Publication title -
canadian journal of administrative sciences / revue canadienne des sciences de l'administration
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.347
H-Index - 48
eISSN - 1936-4490
pISSN - 0825-0383
DOI - 10.1002/cjas.1465
Subject(s) - computer science , series (stratigraphy) , time series , volatility (finance) , partial least squares regression , data mining , econometrics , artificial intelligence , machine learning , mathematics , paleontology , biology
Business forecasting with double‐trend time series (long‐term trends and seasonal volatility) has been challenging due to its complexity. Neither a single time series model nor a fixed‐weight combination approach can fully capture the comprehensive information. We address this issue by proposing an improved partial least squares (PLS) based time‐varying weight combination approach. The proposed method can handle the relations both between the single models involved and between single models and time ordering with time‐varying weights. The test on 20 simulated datasets demonstrates the better and more robust performance of the method. We also apply it to three real datasets. The results show that our approach represents a significant improvement over the existing methods in terms of data fitness and prediction accuracy. Copyright © 2017 ASAC. Published by John Wiley & Sons, Ltd.