⭐ Fashion Retailing Sales Forecasting

Proposed Autoregressive Multimodal Transformer modal for fashion recommendation by sales forecasting.
Implemented zero-shot forecasting methodology for improved generalization and addressed limited historical data challenges, achieving 35% improved MAE and WAPE.
Built an end-to-end pipeline for data preprocessing, feature extraction from multimodal and exogenous data
This project is my masters thesis work.
Genuine products sales data are used in this project, collected from 9oz.
Image-based Multi-model Transformer method is proposed.
Additionally NAVER apparel products data also used.
Used Language / Framework: Python (PyTorch).

Graphical illustration of the proposed framework for fashion product sales forecasting. The framework depicts the distribution of fashion products to stores through supply chain management practices. Data D is collected from both the 9oz stores and potential exogenous data sources. Separate data preparation techniques, including scaling and sorting processes, are applied to the collected data across various modalities.Following,positionalencodingisappliedtocapturetherela tive order of elements within a sequence. Subsequently, the framework aggregates the processed feature vectors, and this enriched represen tation is fed into an auto-regressive multimodal transformer module. The transformer utilizes a shared attention mechanism to address the sequential nature of sales data and to facilitate interaction and learning of joint representations across different data modalities. Ultimately, this framework enables accurate sales forecasting, denoted as f(x).

Predicting future sales volumes of fashion industry products is challenging due to rapid market changes and limited historical sales data for recent products. As traditional forecasting methods and machine learning models often fail to address this problem, we propose a novel autoregressive multimodal transformer architecture to anticipate the sales volume of brand-new apparel items bycapturing trends amonginterrelated attributes. In this paper, we utilize authentic data from a fashion company that includes a limited amount of historical time-series sales data and several influencing factors like product image, textual descriptions, and temporal attributes. To mitigate the data inadequacies, we investigate the impact of integrating exogenous knowledgefromane-tailersitefiltered withfashion apparel products. Also, wefoundthatemployingthezero-shotforecasting approach further aids in forecasting with minimal time-series sales data. Our approach achieves the values of 1.546 and 16.42 in terms of MAEandWAPE,respectively, by leveraging exogenous data compared to existing benchmark models. This study demonstrates the potential of our autoregressive multimodal transformer to predict sales volumes with more precision, and it highlights the importance of incorporating the zero-shot forecasting approach in the dynamic fashion industry. (Rajendran & Hong, 2025)

The sample sales profile forecasting approach is illustrated in this Figure. To forecast sales volume, the model searches for similar products in the same or neighboring groups, as shown in Figure (a). The sales volume of the most similar product is then segregated, as depicted in Figure (b). The available data, such as product image, design attributes, temporal attributes, similar product information, and sales volume of similar products, are utilized by the model to predict the sales volume (Figure (c)).

In our study, we consider a forecast horizon of 8 months of data, although the practical horizon necessitates only 6 months of data due to the distinct 6-month duration of fashion periods, such as Spring-Summer (SS) and Fall-Winter (FW), as depicted in the above figure.

The graph displays the Quantitative forecast results with other compared models in terms of MAE and WAPE for the 9oz dataset a, and the Visuelle dataset b. The term GTMTstands for GTM-Transformer, FuMLP denotes FusionMLP, and AMT refers to Autoregressive Multimodal Transformer

Predicting future sales volumes of fashion industry products is challenging due to rapid market changes and limited historical sales data for recent products. As traditional forecasting methods and machine learning models often fail to address this problem, we propose a novel autoregressive multimodal transformer architecture to anticipate the sales volume of brand-new apparel items by capturing trends among interrelated attributes. In this paper, we utilize authentic data from a fashion company that includes a limited amount of historical time-series sales data and several influencing factors like product image, textual descriptions, and temporal attributes. To mitigate the data inadequacies, we investigate the impact of integrating exogenous knowledge from an e-tailer site filtered with fashion apparel products. Also, we found that employing the zero-shot forecasting approach further aids in forecasting with minimal time-series sales data. Our approach achieves the values of 1.546 and 16.42 in terms of MAE and WAPE, respectively, by leveraging exogenous data compared to existing benchmark models. This study demonstrates the potential of our autoregressive multimodal transformer to predict sales volumes with more precision, and it highlights the importance of incorporating the zero-shot forecasting approach in the dynamic fashion industry.

References

2025