An effective demand planning model starts with sound business design decisions. Historical data quality, calculation granularity, product segmentation and the level of automation all determine forecast reliability. In this article, we explore the key features of a demand planning model in Anaplan, along with the architecture and configuration choices that make it easier for teams to adopt.

1. Classification, product linking and historical data cleansing

The first critical element is the quality of the sales signal. Unprocessed historical data introduces bias into every downstream calculation. This is one of the most underestimated areas of investment during the early stages of a project. We configure product linking features to maintain historical continuity across product code changes, product substitutions and cannibalisation. This approach is essential when product portfolios evolve regularly.

We model plausibility ranges using configurable dynamic thresholds. These identify outliers and trigger automatic corrections. Promotional effects are captured and separated from the baseline sales signal.

We also incorporate the reconstruction of lost sales during stockouts. Unfulfilled sales are estimated to avoid underestimating actual demand. The result is a linked, cleansed historical dataset that can be used directly for statistical forecasting.

In our experience, product lifecycle management and ABC/XYZ segmentation are fundamental pillars of the model. ABC measures each product’s contribution in terms of volume and revenue. XYZ assesses demand volatility using the coefficient of variation in sales. These segments inform data cleansing rules and help prioritise planner alerts. They also guide sales signal modelling, including the choice of statistical algorithms, the use of Machine Learning and the need for manual review. Without well-calibrated segmentation, maintaining the model quickly becomes unmanageable as the product portfolio grows.

2. Statistical forecasting, Machine Learning and Best Fit in Anaplan

The value of a forecasting engine depends primarily on how appropriately it is configured. It is not simply a matter of using more complex models. Statistical models cover most forecasting requirements, typically around 80% of SKUs depending on the portfolio. They account for seasonality and trends through moving averages, exponential smoothing and model families such as ETS and ARIMA.

At OneHive, our demand planning consultants configure Anaplan Forecaster according to demand profiles, product segments and the performance indicators being monitored. They adapt the parameters to the specific characteristics of each portfolio to improve forecast accuracy. A standard implementation of the forecasting engine is therefore not enough to guarantee reliable results.

Machine Learning is used for segments where it can deliver measurable improvements. Its application is not limited to erratic demand patterns. It can also incorporate explanatory variables such as promotions, cannibalisation, calendar effects and specific product characteristics. Depending on the context, a dedicated model can deliver a 3 to 5 percentage point improvement in Forecast Accuracy, subject to confirmation against the results of the relevant project.

We configure the Best Fit functionality at node level, according to client requirements. Anaplan runs several model families in parallel and automatically selects the best-performing candidate. The selection metric can be configured using indicators such as MAPE, WAPE, RMSE or bias. RMSE is our default starting point.

We also use ABC/XYZ segmentation to guide model selection. An A/X product with stable demand should not be treated in the same way as a C/Z product with volatile demand. Demand planning teams retain the ability to manually override the selected model when business circumstances require it.

We continuously measure forecast quality through backtesting. These analyses are based on validation periods defined in advance. Forecast Accuracy dashboards identify deviations and help teams refine model parameters over time.

3. Promotions, events and product lifecycle management

Promotional planning is an area of business design that is often underestimated. Before any configuration begins, the promotional mechanics must be clearly defined: offer type, discount depth, duration, channel, customer and the category of products concerned. A promotion on a fast-moving SKU does not have the same effect as a clearance campaign or a new product launch. When poorly modelled, these activities distort the statistical baseline and forecasts for subsequent campaigns.

At OneHive, we build dedicated modules to separate promotional uplift from the baseline sales signal. Bulk integration of promotional activities relies on configurable rules. The model proposes forecast volumes based on the type of promotion, channel, customer and product category. Execution tracking then makes it possible to compare actual sales against forecasts and measure the performance of each event.

We also feed promotional data back into Machine Learning models. This approach captures halo effects, cannibalisation and promotional uplift. It also accounts for changes in demand before and after a promotional event. Forecast accuracy improves as the promotional history becomes richer.

We manage the product lifecycle directly within the model. Product launches can be configured manually or automatically. Two approaches are possible: applying a volume to a launch curve or using a comparable product to reconstruct historical demand and calculate a baseline forecast. Cannibalisation mechanisms can be configured between SKUs. End-of-life products are managed using dedicated decline curves to progressively reduce planned volumes.

4. Forecast adjustments and scenario planning

The granularity at which the sales signal is managed is one of the most important design decisions. Working at too detailed a level multiplies manual activities without adding value. Working at too aggregated a level hides useful demand signals. We configure top-down and bottom-up approaches according to business needs: some adjustments are made at a detailed level, while others are performed at an aggregated level. Combining both approaches within the same model is essential for an efficient review process.

We configure multiple scenarios, including baseline, optimistic and pessimistic. Each scenario incorporates business assumptions relating to pricing, marketing budgets and supply constraints. Differences between scenarios are displayed in absolute or relative terms, with contributions broken down by segment. The analysis metric is configurable (kg, units, cases, revenue or euros), allowing teams to quickly assess the impact of their assumptions on demand and service levels.

5. Demand review and alignment with S&OP

We structure a shared validation process that connects local teams with a centralised view. Bottom-up forecasts submitted by countries or business units are consolidated and compared with the top-down perspective derived from global objectives. Differences are made visible and provide the basis for consensus decisions during Demand Review meetings. For our clients, this structure significantly reduces the time required to prepare these meetings. On some projects, the time needed to complete the S&OP cycle has been reduced by more than 30%.

Consolidation covers multiple planning horizons, from short-term operational requirements to long-term budgeting. It incorporates the financial perspective, including revenue forecasting and margins, and supports the development of a consistent overall plan, from the Demand Plan to the Supply Plan and ultimately the Financial Plan. Gaps between commercial ambitions and operational realities are identified, measured and translated into decisions and actions.

Conclusion

An effective demand planning application is not built around an algorithm alone. It relies on three foundations: reliable, cleansed historical data; forecasting models configured for each demand profile; and structured management of commercial events and the product lifecycle.

Anaplan provides the technical framework needed to bring these components together in a consistent, auditable application connected to the S&OP process. The value delivered then depends on the quality of the business design, the precision of the configuration and how effectively the process is embedded in teams’ day-to-day practices. A high-performing model is therefore about more than its features. It must also be understood, used and maintained by the teams responsible for managing demand.

In practice, we consistently observe a progressive improvement in forecast accuracy over the 6 to 12 months following implementation. The most successful projects achieve Forecast Accuracy levels above 90% at S&OP granularity, with a bias below 5%. This progress does not simply reflect an immediate improvement in statistical performance. It also demonstrates the gradual adoption of the model by teams and the cumulative benefits of historical data cleansing mechanisms.

Are you building or improving your demand planning model in Anaplan? Get in touch with our experts.