In this analysis we combine the features, order input history and the non-numeric variables, such as region and class, in a deep learning method called 'embedding', to make highly accurate near-term forecasts. The accuracy of this method on this client's data is 5% MAPE over all dealers forecasted individually.
This technique includes the ability to use the time interval since a promotion, product introduction or any other market activity that was either started or ended in a prior time period. No other technique can efficiently include categorical variables such as region and class along with the time dependent interval features as well as order input history. The accuracy of this forecasting method arises from the inclusion of all variables: categorical, interval, ordinal and continuous. It provides the accuracy that this client needs to rely on these forecasts.
Jeremy Howard's Fast.ai disclosed this deep learning solution for structured data. Embedding is typically applied in language processing but is adapted to structured data here.
Accurate forecasts using the neural network embedding method for each of this client's dealers are displayed in this chart. An accuracy of under 5% averaged across all dealers is achieved.
This data table is interactive and displays the deep forecast for each dealer with the blue line. The deep learning trend forecast is red and the statistical time series forecast is green. By selecting a dealer from the drop down box a new dealer is displayed.
Accurately forecasting each dealers near-term performance is the final challenge this client presented. Accurate forecasts by partner is the last client challenge. The dealers can now be reliably sorted for expected performance, allowing this client to more effectively focus on each dealer's individual strengths and weaknesses when applying its resources.