Performance evaluation of metaheuristics-tuned recurrent networks with VMD decomposition for Amazon sales prediction

The exchange of goods and services for the sake of production, resale, or other organizational needs is known as business-to-business, or B2B, purchasing. It entails a transactional relationship between buyers from companies and providers. B2B sales often include larger quantities and fewer purchasers than consumer-focused transactions. They also have a special kind of decision-making that involves buying centers and is characterized by complexity. This study investigates if time series forecasting techniques may be used to predict B2B purchases using two distinct types of networks. Performance maximization is achieved by handling parameter selection challenges using a modified metaheuristic optimizer, tailored for this specific task. The methodology is also expanded to use decomposition techniques in order to help attain more accurate predictions by extracting trend data. Two types of forecasting networks are explored. Recurrent neuronal networks (RNNs) and a variation on these networks that utilized reservoir computing echo state networks (ESNs) are both explored on real-world sales data. While netters originally struggle to attain modest results with best models attaining an \(R^2\) $R^2$ score of 0.082790, following decomposition performance is boosted and an \(R^2\) $R^2$ of 0.646967 is attained. These outcomes suggest considerable potential of the proposed methodology in the B2B domain.