Predictive Analytics & Forecasting

Last year plus five percent is a surprisingly common planning method. That works as long as little changes. Add seasonal effects, promotions, price changes or volatile markets, and the deviations grow. On one side there is too much stock tied up in the warehouse, on the other delivery capability and staff are stretched to the limit.

We build forecasts for demand, sales, utilization, cancellations and payment defaults. Behind these sit two kinds of questions. One is about developments over time: how many units will sell next week, how high will utilization be next quarter? That is forecasting in the narrow sense. The other is about probabilities for individual cases: how likely is this customer to cancel, which invoice will probably be paid late? The result is a value between zero and one that lets you rank the cases. Planning in finance is covered under FP&A & Finance Analytics, and general model development under Machine Learning Development.

What exactly is being predicted?

Sales in units or revenue in euros, per item or per product group, per day or per month, for the next four weeks or the next forty? That depends on the decision attached to it. Ordering goods with a long lead time calls for a different horizon and level than shift planning.

The yardstick is a baseline, a simple benchmark forecast: last year’s value, the same weekday of the previous week, a moving average. A model is only worth the effort once it clearly beats that baseline. It is measured through backtesting. The method forecasts past periods without knowing their values, and the hits are compared with the actual figures. Testing several periods matters so that a single outlier does not distort the picture. The error measure should fit the decision, for instance the error in units or in euros.

Which method wins is decided by backtesting

Classical time series methods such as exponential smoothing, ARIMA or Theta model the trend and seasonality of a single series. They are fast, easy to follow and often hard to beat, especially with regular patterns. Machine learning, such as gradient boosting, learns across many series and takes many drivers into account. That pays off when items, stores and promotions are looked at together. For cancellations and payment defaults, classification models are used accordingly.

The third group is pretrained time series models like TimesFM from Google or Chronos from Amazon. They produce forecasts without being trained on the individual series, and newer versions can also take drivers into account. They do well in public benchmarks, although the figures mostly come from the developers themselves. They can be interesting with many series that have short histories. Whether they beat the baseline and classical methods on your own data only backtesting can show. That includes a look at the license, because for some model versions the weights are released for non-commercial use only.

Which group wins cannot be said in general. An ensemble of several methods is often the most reliable. Sporadic demand, for spare parts say, has methods of its own, because many days without sales distort the usual approaches.

Drivers, hierarchies and uncertainty

Drivers include season and weekday, holidays, school breaks, prices, promotions, competition and weather. What matters is that a driver is known for the forecast period. Holidays and planned prices are. For weather there is only a forecast, and the model should be trained with that too. Training on actual weather gives you a forecast that does worse in use than in testing. Special cases such as lockdowns or supply shortages distort the history and should be flagged. And if you expect yearly patterns, you need several years of data.

Most companies need forecasts at several levels: per item, product group, region and for the company as a whole. Calculated separately, the numbers do not add up. Hierarchical forecasting reconciles them, bottom-up, top-down or with methods like MinT that take all levels into account together. That way the item forecasts sum to the regional forecast.

Every forecast contains uncertainty. Prediction intervals show it as a range in which the actual figure falls with a stated probability. That answers concrete questions: how much safety stock is needed, how many people have to be scheduled for the unfavorable case? A single point value hides this. For predictions about individual cases, the probabilities have to be sound: of the cases rated at 20 percent, roughly one in five should actually happen.

Why forecasts end up unused

A forecast only works once it reaches the place where planning happens. So the models run automatically, and the results are available as a table in the data warehouse or through an interface. BI tools show actuals, forecast and range in the same report, planning platforms take the value as a suggestion, and the ERP can use it for requirements planning and dispatching. Planners can accept the suggestion or adjust it. We record both, so that it can be checked later whether the manual correction improved the forecast. If values are overwritten by hand and nobody notes why, nothing can be learned from it.

Monitoring starts after go-live. Once the actual values are in, the deviation is evaluated continuously. If it gets worse, because of a new product range, changed prices or different customer behavior, the model is adjusted or retrained. A forecast that is never compared with reality loses quality unnoticed. How such an operation is set up technically is covered under MLOps.