| dc.description.abstract | For fashion products, the demand is very unpredictable and life cycle of products is short. Retailers are required to make decisions in the assortment and quantities of purchases and inventory a long time before the time of sale, when only limited and inaccurate information is available. Decisions are made relying primarily on qualitative data and subjective issues. Errors in demand forecast associated with this model of decision-making can approach 50%. However, demand forecasts can be improved by updating the predictions based on early sales. In this context, the aim of this work is to propose a demand forecast model based on learning with early sales for a footwear retailer supply chain. The model was applied in a Brazilian footwear retailer in the 2013/2014 Summer Collection. The demand forecasting model was proposed in two stages: (i) the first stage that used historical data aggregated by subgroup, considering product sales per store; and (ii) the second stage that used data from early sales to disaggregate the demand forecast into products and colors. To generate long-term forecast the logistic model was used. The Weekly Increment Proportion (WIP), which is the weekly demand forecast per subgroup divided by the number of products sold, was used as input data to decompose the demand forecast into products and colors. In addition to this information, the other entries are the quantity sold in the first week and the quantity of products in stock. Modifying the inventory turnover calculation to include WIP, we have the updated demand forecast. The forecast also considers the substitution demand and broken grade. The results demonstrated that the demand forecast model based on learning with early sales obtained higher results than original demand forecast model. The long-term forecast model was adequate for two of the three product subgroups analyzed. The metrics for measuring the predictive performance of the model used were APE (absolute percentual errors) and the MAPE* (adjusted mean absolute percentage error). Two forecast horizons were considered, six and eight weeks. The model performance according to the metric APE forecasting six weeks was 55,199 for the model and 207,511 for the original model prediction. Forecasting eight weeks it was 51,232 for the model and 93.212 for the original model prediction. According to the metric MAPE* forecasting six weeks, the model presented a result of 87.598 and the original model presented 239.777. And forecasting eight weeks the result was 88.454 for the model and 167.515 for the original model prediction. As the model was applied to only one case, it cannot be considered validated. The same results are not expected in different cases. | en |
