Demand Forecasting

In the building supplies industry, where demand for certain products can be both seasonal and sporadic, accurate forecasting is critical to managing inventory and meeting customer needs. Before my involvement, forecasting efforts were highly manual and inconsistent, often relying on broad assumptions or overly simplistic models that failed to account for the diverse demand patterns across product categories and regions.

To address this, I led the development of a robust and scalable forecasting pipeline tailored to the complexities of the business. The solution needed to support weekly demand forecasting for over 1,000 product-location pairs, many of which exhibited low-volume or intermittent demand. Recognizing that a one-size-fits-all approach wouldn’t suffice, I implemented a model selection strategy that dynamically chose between multiple algorithms including Croston’s TSB (for low volumes items), Moriai, and Amazon’s Chronos model, based on each SKU’s historical demand characteristics.

Built in Python, the pipeline ingested multiple years of invoiced quantity data and applied a preprocessing layer to handle gaps, zeroes, and anomalies. To ensure the solution was operationally viable at scale, I engineered a parallel processing framework using the concurrent.futures module, cutting runtime from over 12 hours to under 4. This allowed forecasts to be refreshed on a weekly cadence, keeping pace with business needs.

Forecast results were structured for both traceability and usability and output files were archived by version and loaded into Power BI, where stakeholders could view forecasts alongside actuals, filter by region or category, and assess performance using KPIs like MAPE and forecast bias. Over time, the accuracy of forecasts improved substantially, with MAPE dropping from 38% to 10% due to better model selection and the incorporation of demand seasonality.

The most impactful outcome, however, wasn’t just the statistical performance—it was the cultural shift. By embedding the forecasting outputs directly into the Power BI dashboards teams already relied on, and by framing the models as tools to complement rather than replace human judgment, adoption grew organically. The business began moving away from reactive, guess-based planning and toward a data-driven forecasting rhythm that informed purchasing, staffing, and customer service decisions.