Time Series Forecasting of Customer Spend Using Inception-Residual Convolutional Neural Networks
IEEE Access · 2026 · 14 · 90376-90385
DOI: 10.1109/access.2026.3700538 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Vanja Mileski und 2 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
This paper addresses next-month customer spend forecasting in retail using a large-scale, multi-year dataset from a major Slovenian retailer. The dataset consists of 247,544 customers and 30 monthly time points, as well as the target month. To solve this regression problem, we construct a per-customer, multivariate monthly time series representing shopping habits, spending, visits, and related behavioral signals. We also enrich this data with static attributes of the customers. To model the shopping dynamics, including short-term shifts, medium-term cycles, and long-term seasonality, we propose a 1D Convolutional Neural Network that combines Inception-style multi-branch modules with Residual connections (IncRes). The Inception branches apply parallel convolutions with different kernel sizes and dilations to capture multi-scale temporal structure, while residual pathways stabilize optimization and support deeper, wider representations. Static features are integrated using a lightweight dense branch and fused later in the architecture, enabling the model to also learn behavior from the demographic context of the customers. The proposed model achieves an RRMSE of 0.679, outperforming the global mean baseline (1.165), a parameter-matched CNN model without Inception-Residual paradigms (0.721), as well as additional competing approaches including LSTM (0.717), GRU (0.714), MLP (0.715), Ridge (0.704), and ElasticNet (0.702). These results show that multi-scale convolutional representations with residual learning improve spend prediction accuracy from raw signals without manual feature engineering, providing a practical approach for individualized retail forecasting at scale, while achieving lower error rates against competing models.
Zitieren
APA
Mileski, V., Vičič, J., & Kavšek, B. (2026). Time Series Forecasting of Customer Spend Using Inception-Residual Convolutional Neural Networks. IEEE Access, 14, 90376-90385. https://doi.org/10.1109/access.2026.3700538
Vancouver
Mileski V, Vičič J, Kavšek B. Time Series Forecasting of Customer Spend Using Inception-Residual Convolutional Neural Networks. IEEE Access. 2026;14:90376-90385. doi:10.1109/access.2026.3700538
BibTeX
@article{Mileski2026,
author = {Mileski, Vanja and Vičič, Jernej and Kavšek, Branko},
title = {Time Series Forecasting of Customer Spend Using Inception-Residual Convolutional Neural Networks},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {90376--90385},
doi = {10.1109/access.2026.3700538}
}
EUVIMED