Demand planning and forecasting

Demand planning in IBP turns history and market knowledge into a single number per product, location and period that the business agrees to plan against: the consensus demand. The steps are familiar from APO DP, but the tools differ: forecast models are configured in the Manage Forecast Models app, run as application jobs, and reviewed in the Excel add-in or the web planning apps.

Forecast models

A forecast model in IBP is a sequence of steps: preprocessing (for example outlier correction or substitution of missing values), one or more forecasting algorithms, and postprocessing such as error calculation. You define the input key figure (cleansed history), the output key figure (statistical forecast), the periodicity and the history and forecast horizons. SAP offers classic time-series algorithms (moving average, exponential smoothing variants, Croston for intermittent demand), regression-based and machine learning algorithms such as gradient boosting of decision trees, and an automated selection that picks the best fit by error measure. The exact list changes by release; check the algorithm list in your release's help.

Step by step: creating Contoso's fleet model

1. Open Manage Forecast Models and create a model for planning area CMCPA01, periodicity month.
2. General settings: 36 months of history, 18 months of forecast horizon.
3. Preprocessing: outlier correction on the cleansed history key figure, using a method that replaces flagged values with the expected value.
4. Forecasting steps: input key figure cleansed history, output key figure statistical forecast; algorithms triple exponential smoothing and automated selection.
5. Postprocessing: calculate the ex-post forecast and MAPE, weighted MAPE and bias for the last 12 months.
6. Assign the model to fleet products with a forecast model assignment, then schedule it as an application job each month after the history load.

Measuring forecast quality

An ex-post forecast is the forecast the model would have produced for past periods. Comparing it with actual history gives error measures. Three matter most to a reviewer. MAPE (mean absolute percentage error) averages the percentage misses. Weighted MAPE weights them by volume so a small product with a big miss does not dominate. Bias tells you whether the model systematically over- or under-forecasts.

In this example MAPE is (5.3 + 4.9 + 10.0) / 3 = 6.7%. The forecast summed to 11,950 against actuals of 12,400, so the model is biased low by 450 pieces over the quarter. A test of the forecast job does not prove the forecast is good; it proves the job calculates these measures correctly on known data. Prepare a small, controlled data set and check the results against a spreadsheet.

Demand sensing

Demand sensing adjusts the near-term forecast using short-term signals such as open sales orders and recent shipments. SAP's current algorithm uses gradient boosting. SAP's learning material lists constraints worth testing: it works at daily or weekly granularity, needs enough history (the default threshold is 30 weeks; streams with less receive default outputs), runs only on the baseline version, and does not perform well on intermittent demand. For Contoso, demand sensing makes sense for aftermarket parts with steady weekly orders, not for fleet tenders that arrive as lumpy contracts.

From statistics to consensus

The statistical forecast is a starting point, not the plan. In the demand review, sales adds known deals (a Great Lakes Fleet Services fleet renewal), marketing adds promotions and phase-outs, and key account managers adjust for customers who share their own forecasts. Each input is a separate stored key figure, so the review can see who changed what. Consensus demand is then calculated from them, usually with a rule such as "statistical forecast plus sales and marketing adjustments, unless a planner override exists". Sunita closes the review by locking the consensus for the next three months, so late changes go through the next cycle.