Solving your stock challenges with machine learning
Stock-out or overstock – can you relate? Modern buyers are spoiled. They want everything, right now and if possible, delivered to their home addresses. On the other side of this we can find quite complex operations and supply chains which make it all possible, and which allow you to receive something from Amazon in under 2 hours if you are a member of Amazon Prime (with some restrictions which are less important to this story).
These big pushes from major companies are also forcing small-medium companies into a position where they also have to offer huge numbers of items, and where they must be able to provide great service levels and maintain complex supply chains for this to happen. Regardless of the reason, the situation is that most consumer-related companies today are dealing with many SKUs (stock keeping units) and providing great service levels on all of them, and these items are probably coming to their warehouses from all over the world, which is quite a challenging and data intensive job. The real impact on a business happens when replenishment of such items leads to either stockouts or overstocks of obsolete items that can’t be easily sold or disposed of.
In our experience, most of these problems and complexities can be approached and solved with forecasting and optimisation techniques. Modern sales forecasting and stock optimisation solutions like the ones BE-terna are developing solve this challenge in three steps.
The three-step process explained
(1) Estimating demand and market behaviour
A brief summary of the process is as follows – The first step is to identify the type of item for which we are trying to recommend an order. The identification procedure analyses the behaviour of the sales data (sales value, volatility, demand, sales frequency, etc.) and classifies it into one of the type groups. For example, an item can either be constant, new, seasonal, a slow-mover, a fast-mover, promotional, etc. The next step is estimating the demand for the item already classified in a specific type group. This is tackled using machine-learning techniques. We train multiple machine-learning models using data from different sources, and dynamically self-evaluate each of them. In our experience, external data like weather or traffic helps estimate demand in future periods much more precisely ( depending on the industry, of course). The best-performing model is then used to forecast the demand for the next period (in months and up to a year). The definition of the next period depends heavily on your reordering point and other restrictions that we’ll discuss later.
(2) Optimising stock – the Holy Grail of purchasing
Once demand for the next period is estimated, the second step is to maximise stock availability without gaining in overstocks, or in other words, decreasing the stock level without gaining in stock-outs. Doing this means defining reordering points in terms of frequency and ordering quantities. Moreover, quite frequently there are many additional restrictions, again depending on the business. Since this problem has multiple, contradictory objectives, we tackle it using a multi-objective optimisation technique. In short, the quantity and time of the next order means balancing the costs of having too much stock in your warehouse on one side, and between transportation and all transportation-related costs on the other side. The complexity comes from introducing all kinds of restrictions which exist due to the nature of the company’s business or industry. To list just a few, these would include - minimal order quantity, minimal order amount, minimal safety stock etc.
This is more or less the Holy Grail for purchase departments because one possible consequence is stock-out, meaning not only lost sales but probably also the loss of customers, at least for some time, because the competition is just a click away. Also, some distribution companies have penalties if a retailer’s order is not fulfilled, meaning if they experience stock-out. On the other side, overstocking, which might seem like less of a problem from a sales-perspective, has a negative impact on inventory turnover ratio, it ties up working capital (cash flow), and also raises overheads (storage rent, property tax, insurance, write-offs, labour etc.).
(3) Bring the solution into action
The third, and perhaps the most important component for success, is user experience. This entails integration of the solution into people’s everyday business through user-friendly interfaces. Usually this means two things. The first is to provide seamless integration into ERP systems or wherever purchase orders are being placed, and the second is to provide some simple self-service analytics which can be used by employers for interpretability (preferably with a previously used visual interface). This might not look like a step related to the business impact of the project, but it is indeed critical because it is focused on helping people gain trust in and understanding of the estimations provided by the system. If we don’t manage to get this step working, it will be quite hard to realise the potential returns on investment that projects like this can deliver, as the users may not accept or use the solution.
As an example, one distribution company which we developed and implemented stock optimisation and forecasting for managed to get the number of stock-outs to zero in a given period, while managing to save roughly 25-65% in cash-flow depending on the item group. Of course, the potential savings and benefits depends on a company’s current situation, but in our experience those are quite significant for most SMB or enterprise companies which have warehouses at the heart of their operations. Translate this to your company and you can quickly realise that your savings could range from a few hundred thousand to millions of euros / dollars / pounds.