Application of linguistic summarization methods in time series forecasting

摘要

The novelty of this research is to use linguistically quantified sentences, the so called linguistic summaries, to improve time series forecasting. The proposed Forecasting with Linguistic Summaries (F-LS) approach combines multiple autoregressive models in line with the Bayesian model averaging. However, instead of defining prior probability distributions, practitioners summarize trends with natural language, and then, the system creates the probability distributions basing on the learning database, summarization methods and selected classification algorithms. F-LS approach is evaluated with experiments on the real life time series from the pharmaceutical market and on the benchmark time series. The results confirm that linguistic summaries are useful and successful for the predictive purposes. Incorporation of the summarization results for forecasting in line with the granular computing approach significantly improves its accuracy in the real-life use case of forecasting pharmaceutical demand, and the forecasting error is around 5 times smaller than for the naive method. Performance of the proposed F-LS on benchmark datasets with automatically discovered linguistic summaries enables to improve the accuracy by 3.5% compared to benchmarks. (C) 2018 Elsevier Inc. All rights reserved.