Mihaela Simionescu (mihaela.simionescu@faa.unibuc.ro)
Faculty of Business and Administration, University of Bucharest and Institute for Economic Forecasting, Romanian Academy
Bogdan Oancea (bogdan.oancea@faa.unibuc.ro – corresponding author)
Faculty of Business and Administration, University of Bucharest
Oana Simona Hudea (oana.hudea@faa.unibuc.ro)
Faculty of Business and Administration, University of Bucharest
Abstract
Platform economy is an innovative concept that supports the use of digital platforms for business models. In order to take maximum benefit of such platforms, not only the surrounding context is important, but also the ability to precisely determine their destination and to adapt accordingly. With this in mind, we proceeded to the theoretical and applicative analysis of the same, undertaking to underline, based on a comparative analysis of specific econometric versus machine learning methods, how to better forecast short time-series, so as to predict the evolution of the number of participants to such particular platforms. Overall, we identified an important limitation of the Long Short-Term Memory networks, one of the most advanced and effective machine learning techniques for univariate time series forecasting, namely the complexity of computations and the uncertainty regarding the accuracy of results, as compared to the econometric approach, herein mainly represented by SARIMA models. Despite the intensive utilization of machine learning techniques, the current research evidenced the outperformance of the implemented econometric models in some cases. Further research might consider conformal machine learning techniques, to obtain uncertainty quantification too, including a larger number of Long Short-Term Memory networks specific architectures.
Keywords: Platform Economy; Time Series Forecasting; Machine Learning; Long Short-Term Memory (LSTM); SARIMA Models; Teleconference Platforms; Hybrid Forecasting
JEL classification: C53, C45, C32, L86, O33
