Forecasting and predictive models

Forecasting predicts how a measure changes over time: demand per product, sales per branch, the number of admissions or staffing levels per shift. The result is a forecast for each period, ideally with a range.

Predictive analytics is broader: it also includes predictions for individual cases, such as whether a patient will be admitted at St. Antonius. Both are part of our AI applications. If you’re interested in machine maintenance, read about predictive maintenance.

Historical data by product, location and period, ideally covering several seasons. You also need anything that explains demand and is known in advance: promotions, prices, opening hours, Christmas, Easter and school holidays. Record when something was out of stock, too, because demand was higher than sales at those times.

You usually don’t need personal data to forecast demand. If you do make predictions per customer, bear in mind that, according to the Autoriteit Persoonsgegevens, a hashed customer number is usually still personal data, so the GDPR applies.

The benchmark is a naïve prediction, such as the same period last year. Next come statistical time-series models such as exponential smoothing and (S)ARIMA(X), followed, where worthwhile, by machine learning such as gradient boosting with LightGBM, or deep learning such as DeepAR and LSTM. For items with many periods of zero demand, there’s Croston’s method.

In a proof of concept for Rijkswaterstaat and the NGinfra partners, we showed that we could simulate the impact of a bridge opening on traffic congestion, using an LSTM model and ships’ AIS data, among other things.

With a backtest: you have the model make a prediction from a point in the past and compare it with what happened afterwards, repeating the process with a moving start point. That way, the model never uses data from the period it’s predicting.

Look beyond MAPE: it breaks down in periods with no demand. MASE compares the forecast with a naïve prediction. Forecast Value Added shows whether each step, including a planner’s adjustment, adds anything. Read more in our blog about why your forecast model is right but your inventory isn’t.

Sales wants growth, finance looks back and the supply chain team favours certainty. A data-driven forecast gives them a shared starting point; planners add knowledge that isn’t in the data and record why they made changes. Read about how this works in practice in our blog on demand forecasting.

If you forecast by product and region, those figures must add up to the total. This is called hierarchical forecasting: the levels are aligned so that procurement and finance work with the same figures.

Not with a time-series model alone, because there’s no series. Instead, you predict demand using comparable products: how did earlier launches perform in the same category, price range or shop? Use several comparable products, not just one.

As sales data comes in, the model takes over step by step. Record which comparisons you chose so you can see afterwards which were the best predictors.

On Databricks, AutoML forecasting trains statistical models and DeepAR, using a country’s holiday calendar as an additional input; the best model goes into Unity Catalog and the predictions into a Delta table. In Microsoft Fabric, you can use Prophet, SARIMAX and LightGBM, among others, in notebooks or with AutoML, and display the results in Power BI.

After launch, we monitor the gap between forecasts and actual results and retrain when drift occurs. We handle this through MLOps and management and continuous improvement.

When demand is stable and the same period last year already gives a close estimate, when there’s hardly any historical data, or when you can’t make adjustments because the lead time is longer than the forecast horizon. The same applies if your ERP or planning software has a forecast module that’s good enough.

For orders by shop, we have Prediqt, which provides sales forecasts and ordering recommendations. If the data itself is the problem, start with the data foundation. If you mainly want to see what’s happening, a dashboard will do.

Want to compare providers first? See our overview of the best forecasting consultancies.