Demand planners are often faced with the task of converting large amounts of data into an accurate forecast. This becomes more difficult the shorter the planning horizon. The planner has to react quickly and often under intense pressure to understand a vast array of data and to identify trends and errors.

You might have implemented the Demand Planning (DP) module of SAP's Advanced Planner and Optimizer (APO) application to help you with this task. DP offers “consensus-based,” Internet-enabled forecasting tools such as Collaborative Demand Planning and Promotional Planning, and interactive and statistical forecasting functionalities. These tools allow customers, demand planners, marketing and sales teams, and corporate decision-makers to interact and build an accurate demand plan in real time.

Information overload is still a potential problem, but APO provides help here in the form of alerts. Alerts enable management by exception, and they have many guises in the APO modules. As an exception-based management tool within DP, an alert is an effective way of providing the input used to modify both historical sales data and future forecasts. Whether the aim is to smooth and negate the impact of outliers, correct the impact of seasonal trends, account for the effect of promotions, or include the influence of product life-cycle changes, alerts provide early warning signals that some action is necessary.

I'll explain the different types of alerts and show you how you can use them as a forecasting aid. First, let me show you how alerts work within APO.

Alerts and the APO Demand Plan

Let's review the building blocks of an APO demand plan:

1. Define an InfoSource and map an APO InfoCube to the DataSource.

2. Define a planning area. Raw data is retrieved and stored in an InfoCube or live cache and moved in and out of live cache during the planning process. The planning area is defined using an InfoCube, and it uses updates from SAP Business Information Warehouse (BW) InfoCubes, legacy, or R/3 systems. Updated data is written to live cache and the InfoCubes (Figure 1).

Figure 1

APO DP components relevant to generating an alert

3. Define units of measure, currency, and currency conversion parameters for the planning area.

4. Assign key figures and characteristics to the planning area.

5. Determine aggregation and disaggregation factors to be applied to the key figures and characteristics.

6. Create a planning book with a data view and select relevant key figures and characteristics as well as statistical forecasting methods to be used. Remember, each planning book can have multiple data views.

7. Make promotions and product life-cycle assignments to the data view.

8. Use the Macro Builder to assign macros and alerts to the data view. Macros and alerts can be accessed directly in the data view. Alerts can be created and amended depending on the requirement.

9. Assign a user profile to the planning book and data view.

10. Assign an alert profile to the planning area.

Alerts come into play during step 8, assigning macros and alerts to the data view. I'll describe this process later, but first a little more background would be useful.

Create Multiple Views of the Same Data

Within the APO planning book, an administrator can create any number of data views specific to a user's requirements. This focused view of the data allows the user to identify trends and errors in either historical or forecast data. Even more important, the ability to switch between statistical and interactive forecasting modes allows different analyses of the same data.

APO's statistical forecasting tool uses statistical error analysis to determine forecast accuracy. A series of previous forecasts for a particular period is stored, and each deviation from this series is compared to the actual cell values for the same period. The deviation can then be projected into the future. Switching between different forecasting modes allows the user to see the expected deviation, placing the forecast in context and enabling “what if” planning. For example, it enables your company to evaluate future scenarios for maximizing production throughput or sales revenue, minimizing inventory levels, and helping to provide positive cash flow.

Here, alerts are the core method of illuminating and solving forecast errors. You improve the accuracy of your statistical forecast by monitoring predefined tolerance thresholds for the standard errors. With each iteration of the forecast, the model is adjusted where any of these thresholds are exceeded.

APO DP uses two types of alerts: dynamic alerts for interactive forecast planning and database alerts for batch processing of larger data volumes. Dynamic alerts reflect the current plan and are viewed as the data is being planned; the alert is not stored in any long-term database. Dynamic alerts are also macro-dependent. They can be automatically executed when entering or leaving the demand plan or manually triggered from within the forecast plan. The dynamic alert is used to evaluate the current plan or results, and these values are stored in live cache. This interactive alert type is not suitable for dealing with large volumes of data because of performance issues that occur when dealing with large arrays of data.

Database alerts are used to show the forecast as it was during the planning run. When dealing with large data volumes, it is best to perform the analysis as a background job using SAP's batch-processing technology to evaluate the data. The results of the planning run show the situation as it was at the time of the run. In other words, with database alerts you see a snapshot of the plan during runtime. You have the option of creating customer-specific dynamic or database alert types to be used with DP macros.

The Alert Monitor provides a reporting tool with which planners can monitor the state of any APO plan to see if its conditions have been violated. Database alerts are generated in the planning run, and alerts are created when a threshold is exceeded. You can define multiple alerts and assign them to a plan to report on inconsistencies most relevant to the needs of a user.

Using the Alert Monitor, the planner can view alerts and interactively change data. Forecast errors, sales trends, seasonal factors, promotions, and “phase-in/phase-out” products can be isolated and identified, allowing the planner to take appropriate action to validate and correct distortions as he finalizes the demand plan.

You can view alerts in a number of ways. Interactive or collaborative DP alerts can be accessed directly in the interactive mode. Follow the menu path Planning> Interactive Demand Planning or Interactive Collaborative Planning.

You can access alerts through the Supply Chain Cockpit (SCC) control panel (transaction /SAPAPO/SCC01), where you can also see application-specific alerts as well as alerts for individual products, locations, and resources (Figure 2). From the SCC, you can resolve problems through direct access to the corresponding transactions via context menus. You can also resolve problems through the Alert Monitors for the respective applications.

Alerts can also be accessed via the Alert Monitor window as standalone components, which you can access using menu path Supply Chain Monitoring>Alert Monitor or transaction /SAPAPO/ AMON1. Alerts are assigned to the data view created from an APO DP planning book.

Clicking on the Display/Change button at the top of the Change Profile screen (Figure 3) toggles between display and change modes depending on your authorization level. Select Forecast after clicking on Display/Change to view statistical alerts for a planning area, planning book, and data view. For example, at the bottom of Figure 3 you see:

Planning area: ZDP_PA_DDC01
Planning book: ZPB_DDC_01
Selection description: DDC 7040

The selection description DDC 7040 represents a predefined selection variant used in the Forecast Planning screen. This variant is a subset of the data view CO_01. Users are most likely interested in a specific product line, customer, or sales organization, so the alert would be applied to this subset of data. If necessary, a user may also define a variant for all values in the data view.

Figure 4 shows a typical result of an alert. To analyze an alert, right-click on the appropriate row. Then select the Supply & Demand Planning option. This puts you in the forecast planning book/data view (Figure 5), and from here you manually take necessary corrective actions.

Figure 6 shows a macro used to trigger a dynamic alert in interactive DP. In this example, if the value in the Preliminary Final Forecast cell is greater than the value in the Target Stock Level cell for the same period, then it sends a status alert that both the Preliminary Final Forecast and Target Stock Level are under-covering the demand based on the values in Preliminary Final Forecast.

Figure 7 shows an example of a database alert that has been updated in the Alert Monitor. You can compare dynamic and database alerts in Supply and Demand Planning interactive planning.

Figure 2

Alert view from the Supply Chain Cockpit

Figure 3

Select Display/Change (1) and then Forecast (2) to view alerts (3) for Planning area, Planning book, and the appropriate subset of the data view (4).

Figure 4

A typical result of an alert

Figure 5

The forecast planning book

Figure 6

A dynamic alert macro

Figure 7

Example of a database alert

Statistical Forecasting Alerts

Univariant and multiple linear regression (MLR) forecasting use alerts to determine forecast error. The composite forecast method combines forecasts from alternative forecasting methods including judgmental and mathematical models. Each forecast is based on the same historical data, and a weight is applied to the significance of each forecast.

Univariant forecasting. A forecast profile uses a diagnostic group to determine upper threshold limits, and it is assigned to a planning area. During interactive or background planning, you are able to use the statistical error analysis to improve the accuracy of your forecasts by monitoring these predefined tolerance thresholds for the standard errors. An alert is issued whenever a threshold is exceeded. The forecast is then manually adjusted when any of these thresholds are exceeded. Alerts indicate which cell entry requires a correction. Making iterative changes allows you to adapt the plan to remove forecast errors.

The accuracy of the forecast is measured using any one of six forecast error measurements and tested against a normal distribution (Figure 8). The basic technique involves storing a series of forecasts for a particular period and comparing each deviation of this series to the actual values for the same period or previous period. Whenever a deviation occurs, an alert is triggered. The six forecast error measurements available in SAP's APO are mean absolute deviation (MAD), mean absolute percentage error (MAPE), error total (ETl), mean percent error (MPE), mean square error (MSE), and root of the mean square error (RMSE).

MLR forecasting. Unlike univariant forecasting, an upper and lower limit are set in a diagnostic group, which is assigned to the MLR forecast profile. This profile is assigned to the planning book and planning data views. The measure of fit defines the upper and lower limits of the threshold, and an alert is issued if either of these thresholds is breached (Figure 9). You use alerts to monitor the MLR measures of fit in two situations. The first is to run MLR interactively or in the background, and the results are calculated automatically. The process is to run the forecast, check the alerts, and change the model until you are satisfied that you have optimized the model. MLR allows you to compare one dependent variable to many independent variables. The second situation is when you have already created the MLR model and are running MLR with mass processing. The aim here is to see no alerts. This happens if the model you built was a good one. The models used include R square, adjusted R square, Durbin-in, Durbin Watson, and T-test.

Figure 8

Univariant statistical forecast alerts

Figure 9

MLR statistical forecast alerts

Alerts are assigned a specific priority, and in statistical forecasting it is necessary to define separate diagnosis groups in the univariant and MLR forecasting profiles. Diagnosis groups contain the threshold values that cause an alert to be triggered. For univariant forecasting, these are always maximum values when determining the accuracy of the forecast using forecast errors. Forecast errors are determined by storing a series of forecasts for a particular period and comparing each deviation of this series to the actual values for the same period. Whenever a deviation occurs, an alert is triggered.

For MLR, the diagnosis groups define the upper or lower limits, depending on the nature of the measure of fit. When using MLR to compare two models, it is important to always use the same dependent variable.

When combined with APO DP's interactive and statistical forecasting tools, alerts provide your organization's planner with a clearer vision of disruptive factors that are difficult to discern when using traditional forecasting tools. The improved accuracy of the forecast and clearer vision of the past and future will enable your organization to plan with confidence.