In 2016 JPGAL enlisted the help of deep learning specialist Matthew Oakes PhD to try to learn from the data hidden within Epicor ERP.

Championed for image recognition and autonomous driving, could deep learning provide a framework to better understand business and more importantly help Epicor’s customers.

From the outset, the challenges were not the limitations of deep learning, but how should Epicor’s data be prepared to allow deep learning to do its magic.
Several prototypes were built around sales forecasting and prediction. Over 1 million anonymised records (20 years of Sales data) were kindly provided by @Johnathan McCoy of @wd40 company for research purposes, and away we went. Public holidays, weather records and promotional activities were all presented to the algorithm. All of which were included in its learning process, creating biases and weighting the information accordingly.

The results were far from perfect, but the lessons learnt from this exercise brought about a change in mindset at JPGAL. For deep analysis, transaction changes must be recorded rather than overwritten (change logs). Business rules and processes should be known and recorded.

Although we have not pursued deep learning further this research lead to the partnership with QPR Process Analyser to help tackle the business process and data issues that hampered deep learning. In the long term we hope to revisit this project, with a deep learning ready dataset.