Forecasting and predictive models
We build forecasts for demand, sales or hospital admissions that feed into planning in time, as they do at the St. Antonius emergency department.
At the St. Antonius emergency department, our algorithm predicts whether a patient needs to be admitted
As part of the hospital’s AI team, we developed an algorithm to help doctors decide whether a patient can go home or needs to be admitted. Departments are notified in time and can start preparing earlier. “Professional collaboration with dedicated specialists” Marleen Vreeburg, emergency department doctor - St. Antonius
Why is our forecast always off?
The planner takes last year’s sales from the ERP system, adds an uplift and adjusts the figures in Excel for promotions. A week when a product sells out is then counted as low demand. A predictive model forecasts demand for each product and location, with a range of possible values.
A forecast planners can verify and adjust
Demand, not just sales
When a product sells out, demand is higher than sales. We correct for that first; otherwise, the model learns from empty shelves.
Better than repeating last year
Every model must beat a simple benchmark, such as the same week last year. We test it on periods the model has not seen.
Room for the planner
If the planner knows about a promotion that is not in the data, they adjust the forecast and record why. We assess that adjustment afterwards.
For parts and inventory too
At Athlon, we use machine learning to forecast demand for maintenance and repair parts, supported by a data pipeline in Azure.
First the decision and the horizon, then the model
-
Decision and level of detail
Which decision depends on the forecast, how far ahead does it need to look, and is it needed by product, location, week or day?
-
Benchmark and backtesting
A simple model provides the benchmark. We test every improvement against periods the model has not seen.
-
In planning and under management
The forecast goes into ERP, planning or Power BI. We monitor deviations and retrain the model when patterns change.
Can we base our planning on a predictive model?
A good fit if
You plan inventory, purchasing, staffing or capacity based on estimates, have historical data by product or location, and inaccurate estimates have a noticeable cost. Or your planners spend a lot of time in Excel.
Not a good fit if
Demand is stable and the same period last year already provides a close estimate, or you have very little historical data. In that case, a simple method in your planning system is enough.
What determines the scope
The number of series (products multiplied by locations), which additional sources matter, such as promotions or prices, how often the forecast is refreshed, the connection to ERP or planning, and who manages it.
Will planners trust the model?
Only if they can see where a number comes from. That’s why we show the range, track their adjustments and compare the model and the planner over the same periods.
Time series model, machine learning or the module in your ERP?
Statistical time series model
Choose this if you see a clear trend and a regular seasonal pattern for each product. Exponential smoothing or ARIMA is quick, explainable and often a strong starting point.
Machine learning
Choose this if you forecast many series, such as parts for the Athlon fleet, or if promotions and prices matter. Our advice: use it only if it outperforms a simple benchmark.
Module in your planning software
Choose this if your ERP or planning software already produces forecasts and your series show no unusual patterns. Setting it up properly is often enough.
Predictive models for A&E, the workshop and road traffic
Data Science Lab has worked with data and AI since 2016. Based in Amsterdam, we are a Databricks and Microsoft partner. We have built predictive models for St. Antonius, Athlon and Rijkswaterstaat, and can manage them afterwards if you wish.

First, we talk about the decision that depends on the forecast
Stef Stoelwinder
Business DevelopmentStef is your first point of contact. In our first conversation, whether at your offices, in Amsterdam or online, we’ll discuss what you want to predict and what historical data is available.
E-mail info@datasciencelab.nl
Phone +31 (0)20 3636 724
Which prediction would help our planning most?
Describe what you want to predict and how you currently plan. You’ll get a reply by email, or we can arrange an initial conversation.
Message sent
Thanks for your message. We’ll be in touch as soon as possible.
Frequently asked questions
Forecasting predicts how a quantity changes over time: demand per product, sales per branch, the number of hospital admissions or occupancy 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 looking at machine maintenance, read about predictive maintenance.
Historical data by product, location and period, preferably 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 sold out, too, because demand was then higher than sales.
You usually don’t need personal data to forecast demand. If you do calculate forecasts per customer, bear in mind that, according to the Dutch Data Protection Authority, a hashed customer number is usually still personal data. In that case, the GDPR applies.
The benchmark is a naïve prediction, such as using the same period last year. Next come statistical time-series models such as exponential smoothing and (S)ARIMA(X), and, where worthwhile, machine learning such as gradient boosting with LightGBM, or deep learning such as DeepAR and LSTM. For items with many periods of no demand, there’s Croston’s method.
For Rijkswaterstaat and the NGinfra partners, we showed in a proof of concept that we can 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 next, repeating the process with a moving start date. 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 performance with a naïve prediction. And 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 supply chain 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 how this works in practice in our blog about 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 reconciled so that purchasing and finance work with the same figures.
Not with a time-series model alone, because there’s no series. Instead, you forecast based on comparable products: how did previous launches in the same category, price range or store perform? Use several comparable products, not just one.
As sales data comes in, the model gradually takes over. Record which comparisons you chose so you can see afterwards which predicted demand most accurately.
On Databricks, AutoML forecasting trains statistical models and DeepAR, with 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 difference between the forecast and what actually happens, and retrain when drift occurs. We handle this through MLOps and management and continuous improvement.
What else can you predict or plan with data?
- AI applicationsAll our applications, from predictive models to generative AI and business intelligence.
- Optimise planning with AIHow to turn a forecast into an actionable plan for people and capacity.
- Data sciencePredictive analytics, machine learning, computer vision and text analysis: which technique suits which question?