Improving Demand Forecast Accuracy with AI at a Construction Materials Manufacturer

How better forecasting cut inventory 17% and stockouts 52% at the same time

The challenge

A global leader in the design, manufacturing, and distribution of construction materials had two problems at the same time. Their warehouses were full. But their customer couldn't get what they needed.

Their planning process was out of sync with actual customer demand. They keep making and stocking products that became obsolete or expired, and the products the customers did want kept running short.

As a result, working capital piled up in the form of stock no one wanted. They needed to free up working capital by forecasting demand more accurately and aligning their supply to what customers actually wanted to order.

The work

We started by mapping demand by SKU, location, and time. Then we looked at the variables that might realistically have an impact on demand (historical pricing, weather, promotions, distribution dynamics, etc.)

Rather than a single silver-bullet model, we used a variety of methods to figure out which variables drove demand.

  • Moving averages and exponential smoothing where demand was stable.
  • ARIMA for time-series structure.
  • Hierarchical models to link category trends down to individual product behavior.
  • Econometric models to separate the effects of weather from pricing from promotion.
  • Segmentation grouped products by category and seasonality.

The end result was a set of forecasts for each SKU, based on how each actually sells.

The last phase put the forecasts to work. We gave the client an optimized operating model, which helped them reorganize inventory around the model outputs. We also helped change their planning processed based on what the machine learning algorithms and models recommended.

The impact

Metric Result
Inventory levels-17%
Stockouts-52%
Obsolete / expired product expenses-46%

Typically inventory reductions result in stockouts, and stockout reductions cost you inventory. This client was able to get the best of both worlds, Capital came out of idle stock, waste from expired goods fell by almost half, and planning long-term got more accurate across their SKUs.

Why it worked

We realized there was not going to be a single model that could get them a consistent result. We leaned on classical statistics when it was sufficient, machine learning to isolate variables, and used segmentation to keep the models honest at the SKU level. More importantly, the planning process finally ran on accurate forecasting, which could create durable results.