Digital Signals and Fashion Trend Forecasting: Distinguishing Micro-Trends from Long-Term Consumer Behaviour

Abstract

Fashion retailing is seeing an increasing tendency to make use of digital attention data as a key indicator of demand. However, note that online visibility is not the same as real purchasing behaviour. Studies to date show some improvements in forecasting based on modern search and social media signals, but there is limited research providing enough evidence as to whether search interest predicts sales, whether the initial popularity of the product continues over time, and how seasonal forecasting models work if they are validated consistently. The present thesis attempts to fill in these gaps through the analysis of two secondary datasets. The first dataset from VISUELLE 2.0 consists of 5,355 products sold in 110 retail stores from 2016 to 2019, and the second dataset is the Fashion Instagram Trends dataset, which has been presented in various studies related to early popularity and its persistence over time. In particular, this research involves examples such as Linear Regression, Random Forest, and XGBoost for comparisons in terms of weekly sales forecasting when using the datasets with lagging Google Trends variables, and SS19 is considered to be the testing season. It should be noted that Google Trends variables did not contribute to improvement in weekly sales forecasting.