Forecasting swimsuit sales for the next month to assist in inventory management for Heatwave

Project Details


Fall 2018


Jheng Kai-Ru, Adam Yu, Silvia Yang, Zoly Chang






Heatwave is a swimsuit e-commerce seller on the They designed and manufactured their swimsuits and sell them on the Tmall platform. In this project, our team collaborated with Heatwave to work on the forecasting job of predicting swimsuit sales for the next month to assist in inventory management.

The goal of Heatwave is to estimate the future inventory and decide the number of products to put into manufacture for the next month. Moreover, they would also like to adjust their marketing strategies through forecasts. For instance, they would put promotions on the products if the forecasts are lower than their expectation.

At the beginning of the project, the data pre-processing took us lots of time and we made a big effort on integrating two sources of data. The data quality was low and inaccurate. Eventually, we got only 19 months of sales data in total.

We then focused on the top 5 products and generated forecasts using both monthly and daily data. The seasonal naive model was chosen as a benchmark to compare with other models. We found that seasonal naive actually worked best for the monthly data. On the other hand, the ARIMA and ETS model performed well in the daily data.

In the implementation phase, we ran into several difficulties. First, Heatwave lost the majority of their historical data during the platform transition, causing the short length of data. Secondly, the low quality and inconsistency of the data made it hard to pre-process and integrate, which cause some losses of data. Thirdly, we were unable to forecast newly released products due to the short data length. Lastly, the characteristic of the short product cycle in the swimsuit industry makes the dataset relatively short in nature, the forecast will work only if the data can be processed and delivered immediately.

In summary, we suggest Heatwave do the followings:

  1. Save the data periodically to retain data autonomy (ownership) and data quality.
  2. We found similar patterns between new and old products. The company can use the sales of old products as part of the ingredient when forecasting new products with a similar pattern.
  3. Modify models as the datasets become longer. The company can have more precise insights into the product life cycle. Eventually, reaching the goal of lean production.
  4. Make decisions not only with the forecast value but also the manager’s domain knowledge.

Application Area: