Retail Stock Prediction



The goal was to deliver repeatable stock predictions for a retail clothing group from Southeast Asia. We had to make predictions of the data divided into two categories: products with and without a sales history, a few months in advance.

Our work

To predict the sales of products with a history, we combined two approaches: an individual model for each store item and a top-down approach, according to which we summed up categories, built a time series model for each, and then split these global values into items according to a distribution. The prediction in the second category, meaning products without sales history, was more challenging. We applied the Convolutional Neural Network (CNN) for feature extraction to find similarities between the new and the historical products; then, we combined them to obtain a history for learning.

The outcome

  • An improved stock prediction which increased sales and minimized the number of unsold items.
  • The algorithm predicts the sales of products with and without historical data, based on the visual similarity derived from the CNN.
Technology used
  • Python
  • TensorFlow
  • Convolution Neural Networks
  • R