Color Trend Forecasting of Fashionable Products with Very Few Historical Data

In time-series forecasting, statistical methods and various newly emerged models, such as artificial neural network (ANN) and grey model (GM), are often used. No matter which forecasting method one would apply, it is always a huge challenge to make a sound forecasting decision under the condition of having very few historical data. Unfortunately, in fashion color trend forecasting, the availability of data is always very limited owing to the short selling season and life of products. This motivates us to examine different forecasting models for their performances in predicting color trend of fashionable product under the condition of having very few data. By employing real sales data from a fashion company, we examine various forecasting models, namely ANN, GM, Markov regime switching, and GM+ANN hybrid models, in the domain of color trend forecasting with a limited amount of historical data. Comparisons are made among these models. Insights on the appropriate choice of forecasting models are generated.

I. Introduction

The fashion industry possesses a long supply chain that includes operations such as design, merchandising, manufacturing, financing, branding, promotion, and final retailing. At the same time, the industry is known to be facing highly variable demand and fashion trend, and the use of information is an important issue [1], [35]. Accurate forecasting is truly crucial. However, forecasting for fashionable products is always a challenging job because fashionable products, generally, have very short life cycles [32], [33]. Moreover, for many fashionable items, decision makers need to conduct forecasting with just one or two historical datasets to make reference to. The current industrial practice in fashion, heavily, relies on human subjective judgment (known as expert advice), and this mechanism, inevitably, would lead to decision bias such as the anchoring bias [28]. The aforementioned industrial practice and observations have motivated us to examine different forecasting models for their performance in predicting color trend of fashionable products under the condition of having very few historical data.