By John Yuva, Editor, SAPinsider
Robust is a difficult term to define. It’s something that is useful for the organization, something that makes sense. A robust demand forecasting process must be able to handle variances in demand and enable control of your supply chain — with the intent of avoiding situations where you have 1,000 days of inventory or are missing 50% of production orders that should have been delivered to the client. In both scenarios, you can lose a client in the process.
With a robust system, you can adjust quickly to your forecasting needs and how your market is changing. And to view recommendations to adjust. This is where SAP Integrated Business Planning (IBP) for demand can help. SAPinsider sat down with Dr. Marco Sisfontes-Monge, Managing Partner of Arellius Enterprises, to discuss demand planning and forecasting when utilizing SAP IBP.
Q: What types of capabilities does SAP IBP have in the area of demand planning and forecasting?
A: In forecasting, not only SAP IBP in the supply chain, but many other models running in SAP Business Planning and Consolidation in combination with HANA, the important part is that you define a clear process. To give you an idea of how to implement it, when you start off a supply chain project linked to forecasting, you need to define the process.
Where is the forecasting going to happen — where is the physical location the algorithms are going to run in IBP or in another system sending data from IBP to that system, and then back to IBP for analysis? Not all companies use IBP algorithms. Some might use other systems such as Python.
In this case, algorithms are running in a separate system when completed, the data is sent back to IBP reviewed and adjusted and then sent back into another system unless the algorithms will run inside IBP. So, the process is very important.
The second part is linked to the supply chain. Your supply chain is already very complex. Keep it simple. Don’t complicate your supply chain even more with the implementation of software. Keep it simple because any mistakes will be amplified many times if your complexity increases. For example, we have a client that created a complete retail model for supply chain, and then has $500 million differences in a single month for actuals versus SAP APO and SAP ECC for the complete region of Europe. The complexity of their supply chain model has complicated their implementation.
Focus is also required on the supply chain side. When you’re going to do forecasting, you can apply the same models for all product lines. When performing forecasting in your organization, SAP IBP can classify a product based on either data volume or how uncertain it is the item to manage. Basically, forecasting is the way we’re going to manage uncertainty in the market. Thus, we want to be more accurate. So, first, we need to focus based on how much volume you have for a particular product and how certain or stable is demand for that product is.
After that, you work on algorithms applicable to all those segments. And then you evaluate how accurate the algorithms are. You then execute it. You monitor to see if the algorithm is constantly changing or needs a new algorithm. You may also need to adjust the timeframe that’s being fed into the algorithms to learn and apply it to the type of data that you’re handling because this will determine its accuracy. For example, a seasonal model will have different variations than a trend or more constant demand product that requires just smoothing algorithms and without significant changes in its average and median quantity values over time.
So more or less, that’s what a robust system should do is continuously monitor to ensure your forecast is accurate, that it meets your needs, and how closely it affects your supply chain.
Q: Each product then has its own demand dynamics? Treating them the same way is not always applicable?
A: A key example is a supermarket. As a supermarket, you cannot only sell profitable products since this retail environment depends on what is called category management, where there are multiple product families that drive each other’s demand. And you need non-profitable products in order to drive demand on more profitable products. For example, the most profitable area of a supermarket, which you can see based on the square footage dedicated to it and the variety of products is the produce department, where the margins can be up to 80% depending on the perceived value by the customer. Check the difference in price of organic versus non-organic vegetables as an example of customer perception. Some people say when we’re doing forecasting, we’re only going to focus on profitable products. You can’t take this approach as a supermarket because you must sell all products. And you make your profit based on the mixture of product categories. At the end of the day, your shopping bag has a mixture of different products. The supermarkets are making 60% overall because it’s selling profitable and unprofitable products. If you buy milk and you buy eggs, you lose money in eggs, but you make money in milk. So, it’s the same logic amplified.
Q: What are some of the forecasting tools and algorithms available within SAP IBP?
A: Specifically for SAP IBP, it depends on your process. IBP has very powerful algorithms. Also, we have SAP HANA with its PAL library that has many algorithms and SAP BW also has an algorithm that can be used. However, there’s an important item we need to consider. It depends on what type of planning you’re going to do. SAP IBP has two types. One is like demand planning and the other is what’s called demand sensing. The difference is how long in the forecasting you’re going to monitor. Demand planning is more high level with a time horizon of 52 weeks or a year or longer — which is also very challenging.
And demand sensing covers a six-week to the eight-week horizon. So, if you’re a supermarket, it would make more sense to have demand sensing algorithms running because of demand volatility for weekends and seasonality. Six to eight weeks would be better than a long-term demand planning strategy.
With that in mind, this will determine how much noise or variability you have in your models. Let’s say a month-level algorithm will have less variability than daily or weekly levels. It will have more noise. So, the idea is that you’re in control of uncertainty. For most algorithms, when we are developing custom models using statistics like moving averages, single exponential smoothing, and trend models for linear double exponential smoothing key factors need to be used to evaluate its accuracy, that must be reviewed and adjusted over time
So, the good thing is that SAP has a good capability to choose the best algorithm that works for you based on the data that you’re feeding it. Those are the type of algorithms available. But again, I do recommend that people do not apply all of them. Instead, focus on the best fit options available using the parameters that SAP provides.
Q: What are some of the best practices in forecasting, for example, choosing the right KPIs?
A: SAP provides a lot of tools to do forecasting. For SAP IBP specifically, the most important part is to read your models.
Thus, first and foremost, you need to know your data well as it relates to your product. Is it an ABC or XYZ product? If it’s a high-volume product, then you’d use the XYZ classification. An AX (where A means a HIGH quantity volume and X means LOW variability or very constant demand) product is one that is high volume and easier to forecast and predict its demand.
On the opposite side of the equation, you have a C product, where it’s a low volume but high fluctuation (CZ). These are the products that in one month it may be zero units and another month 2,000 units, and then five more months with zero units. In that type of classification, your product demand data is important to know what algorithm to use. Or perhaps, no algorithm at all is required for this type of product since its volume is low and maybe its profitability is similarly low, so a manually input forecast could be the best choice.
Once you know what algorithm to use, use the parameters available for your model. For example, you want to use six weeks of data to prepare for the Christmas season to forecast a specific product. You suspect that the forecast demand will increase compared to the previous year, and you want to optimize how much product inventory is required.
Thus, you need a way to measure how accurate the actual data is and how your planned data behaves. To do that, there are plenty of options to measure what’s called forecast accuracy. So, we can have what’s called mean absolute percentage error, which is very common. It’s very important to know what kind of forecast accuracy you need to use to guarantee that the model is doing what is expected. The lower the value of the indicator used to measure forecast accuracy (variability), the better the accuracy. The most common indicators used in industry are MAPE (Mean Absolute Percentage Error) the lower the percentage, the more accurate. However, it is not good with low-volume products or with products such as CX with many zeros or no demand during several periods because it makes the indicator infinite.
As an alternative to MAPE, we can move to other indicators such as MASE (Mean Absolute Scaled Error). When compared to MAPE, it handles zeros or null values much better during the period for intermittent demand scenarios it is independent of the scale, and the best solution when the MASE value is <= 1. And in the case of mean squared errors, it’s a value close to one. However, when MASE is higher than 1, and with multiple Zeros during the periods we need to keep searching for alternatives such as Error Total, and MSE (Mean Square Error) as further alternatives.
Q: As we come out of the pandemic, how are forecasting trends evolving?
A: A lot of companies are trying to do demand sensing in certain places of the world that they’re sourcing. Because lockdowns continue in different regions, forecasting short-term demand is extremely difficult. A clear example of that could be Australia right now that has a lockdown in different cities almost every other week. So, managing supply is very challenging in geographies that are still highly susceptible to lockdowns.
Watch the video here.
What Does This Mean for SAPinsiders
- Create forecasting models with plenty of data periods as close to reality as possible that reflect the real behavior of your real-life demands.
- Do use forecast accuracy measures to evaluate your forecasting models constantly and dO use ABC/XYZ segmentation as guidance to the type of algorithms you might need to use for the different types of products.
- No forecast accuracy measures over 100% or even 3200% variability used by some of our clients, even though I told them many times, demonstrate the forecasting algorithms are not accurate. Adjust the parameters in the model to find the next best solution, but certainly, variability of actual versus the forecast of over 40% might be an indicator your model parameters must be adjusted because your history has adjusted. Also, you are probably using incorrect algorithms.
- Do not build forecasting algorithms or anything in your production environment directly. It is quite common now in the SAP IBP world for entire teams to make developments in a single environment without having programs properly tested, validated, approved, and the security roles did.
- Overfitting is a constant risk that means having a model so accurate that you trust your algorithms blindly without tracking the real data, and how forecast accuracy has changed over time. For example, in supermarket chain category management, there can be products that are cyclical, but for a brief period they become AX (high volume and low uncertainty), and the rest of the year they are CZ (low volume high uncertainty), such as turkey in during Thanksgiving. On top of that, these products will drive other products. If you have turkey, you will buy potatoes, gravy, beer, wine, and more. And if you don’t have turkey, you don’t sell anything of the rest of the products and the consumer will buy it all somewhere else. Think about this next time you scan your Preferred shopper card in your supermarket, that’s what they are tracking the demand of products based on customer purchases.
MEET THE EXPERTS
Dr. Marco Sisfontes-Monge is a managing partner of Arellius Enterprises. For more than 15 years he has supported SAP implementers, direct clients, and other customers in Europe, North America, Latin America, Asia, and Africa in the industries of automotive, insurance, pharmaceutical and healthcare, logistics, software, utilities, chemical, oil and gas, exploration and natural resources, discrete and process manufacturing, retail, and financial services. His background includes project management and performance measurement, product- and activity-based costing, design optimization, discrete and process simulation, system dynamics, artificial intelligence and machine learning, and advanced statistical methods. He also has finance specializations from the London Business School and Saïd Business School from University of Oxford in England. You may contact him via email at firstname.lastname@example.org.