Authors

Hybrid neural network model for forecasting dynamics sales indicators on digital marketplaces

In the context of rapid digitalization and the growing importance of online commerce, the task of forecasting sales volumes on marketplaces has become critically important for supporting managerial and strategic decision-making. Despite significant advances in neural network models, their practical application within digital platforms faces a number of limitations, including high demand volatility, data sparsity, the presence of numerous heterogeneous factors with varying dynamics, scalability challenges, as well as high requirements for computational resources and training data volumes. Furthermore, many neural network models operate as “black boxes”, which hinders their use in tasks requiring transparency and justification of forecasts, underscoring the relevance of developing specialized models that combine high predictive accuracy with interpretable results. The aim of this study is to develop and empirically validate a hybrid architecture of neural network model designed to overcome such limitations, taking into account the specific operational characteristics of marketplaces. The proposed model integrates a recurrent encoder for extracting temporal context, modified decoder blocks performing decomposition of the time series into a learnable basis of latent components, and a controlled fusion mechanism enabling adaptive incorporation of contextual information at each decoding level. The applied approach, which forms forecasts as an additive sum of specialized components, each trained to extract certain structural elements, provides an context-aware and structured representation of the time series, enables more accurate capture of long-term trends and periodic fluctuations, and enhances model robustness to noise and data sparsity. Experimental evaluation using data from the Wildberries marketplace demonstrated the model’s superiority in forecasting accuracy over classical and baseline models, confirming its applicability in environments typical of digital trading platforms. Read more...