Using Demand Forecasting to Optimise Stock

Warehouse Stock Manufacturing


Leading manufacturer

Tech stack

Google Cloud


Demand forecasting


AI + Machine Learning

We helped our client to optimise stock profiles at each of their branches across the UK. With the aid of a new Machine Learning-driven capability, our client is now able to make decisions on whether to push products to or remove products from its branches, informed by an expectation of its future demand and sales. With this in place, our client has been able to reduce its logistics costs, increase sales volumes, improve order fill rates and reduce stockout instances.

Our impact

  • Developed a single, scalable solution that predicts the optimal stock profile across all products & branches
  • Led to a 31% increase in volume sales and a 64% increase in order fill rate
  • Enabled an 86% reduction in orders that incur logistics costs due to being sourced from another location


The challenge

When a branch runs out of stock for an order, a phenomenon known as “stockout” in retail and supply chain, this typically results in longer wait times for customers, additional operational costs from ad-hoc shipping of the product requested and in some cases, non-conversion of the order; a disruption of overall stock distribution. 

Before working with Datatonic, our client used a business rule to determine whether a particular product should be stocked at a branch in order to combat stockouts. Per business rule, decisions on future stock were dictated by sales and demand in the past 12 months; the branch stock profiles were effectively catering to a market that existed one year before, instead of the month ahead.

Also, tens of thousands of products share one single threshold on sale and demand volume, regardless of whether it is an oil filter which cars need renewing regularly or a more long-lasting product such as a steering wheel.

In addition, the business rule only specifies whether a product should be added to a branch but does not recommend pulling back “cold” stock, resulting in wastage due to branches holding unnecessary products. Our client wanted to utilise Machine Learning’s prediction and generalisation capabilities to recommend for each branch, which products in its full catalogue should be stocked because they are more likely to be looked up and/or purchased in the next 30 days.

Due to its extensive branch network and vast product catalogue, the solution was required to be highly scalable.


Our solution

Following a Proof of Concept with Datatonic, our client has implemented Machine Learning capability to predict which of their tens of thousands of products customers are most likely to make an inquiry for and purchase at each branch, transforming stock decisions from a single business rule to a data-driven predictive solution which recommends stock for each specific product at each specific branch. Overall, Datatonic delivered an innovative strategic solution which combines 4 Machine Learning models to produce an actionable recommendation: whether a product should be stocked at a branch or not. The aims of the solution proposed by Datatonic are threefold:

  • Aim 1: Identifying possible waste → remove these products from current stock
  • Aim 2: Identifying possible opportunity → add these currently not stocked products
  • Aim 3: Identifying PNOs with no benefit of stocking at branch → remove them 

Definitions of possible waste and possible opportunity were powered by 2 propensity models, forecasting demand and sales likelihood in the next 30 days respectively for every branch-product level. The intuition behind these definitions being products that are very likely to be looked up and sold at a branch but are not currently stocked are possible opportunities; possible wastage refers to currently stocked products which are forecasted to be unlikely to be sold or even looked up. 

As for “benefit of stocking at branch” in aim 3, it refers to the additional look-up to sale conversions brought by having a product in stock. In order to predict conversion rates for when a product is stocked versus not, Datatonic devised two additional machine learning models. To test its effectiveness, our client’s Business Analysis team rolled out a pilot in selected regions in the UK and have yielded compelling results of a 31% increase in sales volumes, a 64% increase in order fill rate, and an 86% reduction in VOR orders (order during branch stockout).