Paper Title
Forecasting Model Selection in Meta Learning; A New Approach

Abstract
Selecting an appropriate model from among a wide variety of candidates can increase the accuracy of forecasts in business forecasting. Model selection in forecasting literature has mainly been discussed by comparing individual approaches such as penalized likelihood methods and empirical cross-validation with different error measures. However, individual model selection has to implement all the candidate models in the data set to evaluate their performance and select between them. This wrapper procedure is highly time-intensive in the current big data era. On the other hand, filter approaches (such as meta-learning) with feature-based time series representation uncover the hidden patterns in complex time series and use these formodel selection. Although a filter approach can select the appropriate model without implementingall models in the data set, its performance relies strongly on the efficiency of the extracted features from the time series (meta-features). Despite the existence of a plethora of meta-features in the literature,the exploitation of statistical tests (popular tests for model selection) as meta-features is largely neglected. Therefore, this study proposes to use statistical tests as meta-features in forecasting model selection for the first time and to report their efficiency against existing meta-features. Moreover, the efficiency of meta-learning in comparison with individual model selectionis further demonstrated. To validate the proposed approach, the NN3 competition time series has been used, and the results of the comparison with commonmodel selection approaches such as aggregation and penalized likelihood functionhave been established. The results indicate the distinctive efficiency of the statistical tests as a new group in feature-based time series forecasting. Keywords - Forecasting; Model selection; Meta-learning; Meta-features; Statistical test.