When Old School Methods Trump Fancy Data Science
A customer in the assembly sector contacted us to evaluate the logistics forecasting and replenishment algorithms. The objective was to look for improvements via data science techniques, mainly regression in all sorts of flavours, time series and –to make sure we didn’t miss out on the latest fads- artificial neural nets.
We were up against years of experience in implicit knowledge, heuristics and judgment calls that delivered reasonable results in … a massive Excel sheet.
In came our SAS, Clementine, KNIME and Python specialists. We mimicked the practices of the client as much as possible, compared these with knowledge discovery facilitated by the tools, only to find out that we could only marginally improve on the forecasting results.
Though the new approaches had a slightly better mean square error (MSE) and the confusion matrix showed less false negatives and false positives, the cost of keeping the system alive compared to the status quo proved to be prohibitive.
Nevertheless, the client became aware he needed better documentation of his team’s tacit knowledge.