Adaptive Multimodal Fusion Transformer for Supply Chain Demand Forecasting under Dynamic Market Conditions
International Journal of Business and Social Science Studies · 19.10.2026 · 6 (2) · 319-330
DOI: 10.66795/ijbsss79 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Business and Social Science Studies, online erschienen am 19.10.2026. Verfasst von Lena Cooper und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Demand forecasts in retail supply chains must stay usable when market conditions shift and when input feeds such as prices or event calendars are late or wrong. This paper presents an Adaptive Multimodal Fusion Transformer (AMFT) that combines three information sources: weekly-patched demand history, the price trajectory over the lookback and forecast horizon, and textual event descriptions embedded with pre-trained word vectors together with benefit-payment (SNAP) days. Each exogenous modality enters through an availability-masked, horizon-specific sigmoid gate that adds a residual correction to a demand-driven representation; all pathways are modulated by FiLM layers conditioned on market-regime descriptors, and modality dropout trains the model for missing inputs. We evaluate on 6,100 item–store series from the M5 dataset over seven rolling 28-day windows (December 2015 to June 2016) without retraining, under a known-price-schedule assumption: future prices are taken from the observed M5 price records. AMFT reaches a WRMSSE of 0.809 ± 0.013, statistically indistinguishable from LSTM-S2S, TiDE and a price-aware LightGBM (Diebold–Mariano p ≥ 0.05) and 19.6% below seasonal naive. With prices or events withheld at test time, AMFT degrades by at most 1.9%, whereas an early-fusion Transformer degrades by up to 306%; adding modality dropout to that Transformer restores robustness only at a 5.9% accuracy cost. The analysis further shows that the price–demand relation shifted after the training period and that holiday closures favor tree ensembles, motivating drift-aware price encoders and event-specific losses.
Zitieren
APA
Cooper, L., Powell, S., & Green, S. (2026). Adaptive Multimodal Fusion Transformer for Supply Chain Demand Forecasting under Dynamic Market Conditions. International Journal of Business and Social Science Studies, 6(2), 319-330. https://doi.org/10.66795/ijbsss79
Vancouver
Cooper L, Powell S, Green S. Adaptive Multimodal Fusion Transformer for Supply Chain Demand Forecasting under Dynamic Market Conditions. International Journal of Business and Social Science Studies. 2026;6(2):319-330. doi:10.66795/ijbsss79
BibTeX
@article{Cooper2026,
author = {Cooper, Lena and Powell, Seojun and Green, Sofia},
title = {Adaptive Multimodal Fusion Transformer for Supply Chain Demand Forecasting under Dynamic Market Conditions},
journal = {International Journal of Business and Social Science Studies},
year = {2026},
volume = {6},
number = {2},
pages = {319--330},
doi = {10.66795/ijbsss79}
}
EUVIMED