Demand Planners Rulebook
What should demand planners do and not do? First, we emphasize critical practices for demand planning (viewing forecasting as an information game, monitoring forecast value added, and leveraging a reliable automated forecast engine). Then, we discuss what demand planners should do (gather insights, refine data, and enrich forecasts) and not do (trim outliers, tweak forecast parameters, or include supply constraints in demand forecasts). Finally, I advocate for a shift from product-centric to insight-driven methodologies.
As demand forecasts drive purchasing, production, delivery, cash flow, and financial planning, supply chains heavily invest in tools and planners to create the best forecasts possible. Unfortunately, 15 years of academic research show that, on average, only 50% of human-made adjustments result in better forecasts. This begs the question: How can we improve this? This article will answer this question by defining demand planners’ role within the overall demand planning process.
Acknowledgments
Tim Young, Malek Djoudi, Guilhem Delorme, Robin Watkins, Sophia Reyes, Navid Tavoli, Prince Boadu, Enzo Perez, Piet Vandenbroucke, Abhishek Dolbaruah
Three Key Concepts for Demand Planning Excellence
Before discussing demand planners’ rulebook, we need to examine three critical practices that will allow them to unlock their work: 1. see forecasting as an information game, 2. use a bulletproof automated forecast engine, and 3. track forecast value added (FVA).
Forecasting is an Information Game
To predict the future, you need (a lot of) insights.
Supply chain planners need to see forecasting as an information game. The more you know about the past, the present, and the future, the better you’ll predict — or even shape — the future.
The central insight you need to predict future demand is accurate historical unconstrained. Unfortunately, most supply chains struggle to capture unconstrained orders and, instead, track and forecast supply-constrained sales, shipments, or invoiced amounts. When forecasting supply-constrained sales rather than unconstrained demand, a vicious cycle ensues, where recent shortages breed pessimistic forecasts, consequently reducing supply orders. In short, supply-constraining demand plans create self-fulfilling prophecies of out-of-stocks. I call this the vicious supply-sales circle (I discuss it thoroughly in my latest book). As a result, we strongly recommended to use unconstrained demand to generate future forecasts.
Depending on your industry, in addition to historical unconstrained demand, you must track various business drivers such as launch activities, prices, promotions, shortages, client incentives, sell-out and stock-in-trade, orders booked, etc. Again, more (accurate) information means better forecasts. I like to call this an insight-driven approach to forecasting.
This approach is similar to Marketing Mix Modeling, a data-driven framework for assessing the impact of marketing campaigns.
You Need a Bulletproof Automated Forecast Engine
Supply chains need to forecast demand for tens or hundreds of thousands of combinations of products x locations at scale. Relying on an army of demand planners to manually review and update every forecast is impossible and likely to be counter-productive (remember that academic research showed that 50% of the enrichments don’t add value).
Instead, supply chain decision-makers need a forecast engine to process as many insights as possible to generate accurate forecasts automatically. Imagine your forecasting engine capturing several business drivers (such as promotions, prices, or point-of-sale data) as inputs to make its predictions. The ability to leverage various insights will make your model bulletproof. By bulletproof, we mean an engine able to cope with most of your business (as well as seasonality and trends) without requiring human inputs or enrichment. For example, if you’re in a promotion-driven supply chain, you don’t want to rely on a forecast engine that isn’t aware of promotions (and therefore can’t forecast their impacts).
Moreover, you need a fully automated model that doesn’t require monthly or weekly fine-tuning. You want your forecast engine to automatically deliver forecasts for all your products: old and new, stable and erratic, seasonal and non-seasonal.
Machine learning models will be the backbone of your bulletproof automated forecast engine. They can successfully process various business inputs to generate accurate forecasts. And they can forecast all types of products, including brand-new and erratic ones.
On the other hand, statistical models will struggle to cope with business drivers. Even on datasets without business drivers, machine learning will deliver more accurate predictions than statistical ones (see this case study for a detailed example). Providing more insights into a machine learning model will only widen this gap (as shown in this second case study).
You Need to Track Forecast Value Added
Tracking forecast value added (FVA) is the cornerstone of demand planning excellence. In short, FVA means tracking the accuracy added value of every single step (or even individual) in your forecasting process. Without this measure, you will be blind to leading and improving your process. As a supply chain leader, implementing FVA should be your main priority. Learn more about FVA in this article, my book, or this webinar (with two supply chain leaders presenting how they implemented it).
What Should Planners Do?
Now that we understand the philosophy (insight-driven) and the main tools of our demand planning process (a machine learning bulletproof automated forecast engine and FVA), we can define what planners should or should not do. Their main tasks should be: 1. Clean data, and 2. Gather extra insights and enrich forecasts (when appropriate).
Clean Data
Planners are data stewards. They should ensure that all data used in the demand planning process (and fed to the forecast model) is correct. There are two main types of data: master and transactional.
- Master data. Planners must ensure that all master data (such as product family, brands, and transitions) are correctly recorded. Many forecasting engines (especially when relying on machine learning) will use these features to generate better forecasts.
- Transactional data. Planners must ensure that all orders are correctly registered, especially their initially requested delivery date and the code for rejection of canceled orders (if any).
This should include all data fed to the forecasting engine, including promotional activities, pricing, and client sell-out data.
Collect Extra Insights and Enrich Forecasts
If you know something, do something.
Planners are investigators. They should gather insights beyond what existing systems already feed into the forecast engine and use them. For example, if the forecasting model can’t consider clients’ sell-out and inventory-level data, planners should use this (at times unstructured) data to enrich forecasts.
If planners realize that the same insights are often used to enrich forecasts, they should discuss with the data science team how to upgrade the model to automatically include these inputs.
What if planners can beat the forecast engine but without specific insights?
If humans can beat a model at forecasting time series just by looking at graphs, it means that the forecasting engine isn’t bulletproof. (Data) scientists should improve it.
What Should Planners Not Do?
Clean Outliers
Many planners spend time spotting and trimming down outliers (by trimming down, we mean bringing back historical demand amounts into reasonable ranges). Outlier detection is usually so crucial for planning leaders that they often include it in their planning software RFQ.
At SupChains, we think data cleaning is tremendously important. But we usually don’t spend time detecting and trimming outliers. In short, there are the golden rules we follow,
- Never trim historical sales data based on deviation to the mean.
- Clean your transactional data as explained why and how in this webinar and this article. For example, highlight erroneous transactions by looking at their price/unit. Once data cleaning is done correctly, you are usually left with zero outliers in your sales time series.
- Censor exceptional periods such as covid.
- As an ultimate last resort, trim historical data based on historical forecast errors.
Planners shouldn’t be left to judge what is or is not an outlier because,
- Variability and Biases. Different planners will assess what is or is not an outlier differently. There is inherent variability (and bias) in asking humans to determine numbers.
- Data Hacking. Allowing planners to change historical values opens the door to data hacking.
- Poor Corrections. Many planners and software vendors see outlier detection and cleaning as a trimming exercise: planners (or models) trim extreme values in sales time series by reducing them to a reasonable range. Unfortunately, this is a dubious way to correct sales patterns: how do you know the correct value?
Models will always be faster and, by design, more predictable and less biased than humans at spotting outliers.
Don’t ask humans to do the work of a computer.
Tweak Models
If you can manually tweak and improve a forecasting model (for example, by changing its inner parameters), it means the model isn’t bulletproof. It’s time to improve it.
Account for Supply
As discussed earlier, demand forecasting is about predicting unconstrained demand, not supply-constrained sales. Forecasting sales will lead to a vicious shortage circle, where any shortage will likely lead to a perpetual out-of-stock situation.
If you want to forecast actual future sales (or revenues), start from an unconstrained demand forecast. Then, let your supply planning engine compute expected future production, deliveries, and shortages. This will mechanically give you a supply-constrained sales forecast that you can easily translate into revenues. (Or even into a cash forecast.)
Consequently, planners shouldn’t consider supply and shortages when forecasting demand.
How to Review Forecasts?
Don’t Use Basic ABC XYZ Segmentation
I strongly advise against prioritizing forecast reviews using basic ABC XYZ Segmentation. The usual ABC XYZ classification (as portrayed on the right) uses historical demand volumes and variability to segment products, which is a poor and outdated technique.
Historical volumes and demand variability shouldn’t determine priority:
- Historical volumes aren’t always correlated with future volumes (think seasonality, trends, promotions, or strategical shifts).
- High historical volumes don’t guarantee good forecasting accuracy or high margins.
- Demand variability is a poor indicator of forecast ability (I dedicate a section to this in my book).
Use a Better ABC Classification
Instead, as explained in my book Demand Forecasting Best Practices, if you want to use ABC XYZ to prioritize your work, you should define it based on future (forecasted) revenues and historical forecast errors (expressed using MAE% and bias — in any case, never use MAPE).
Focus on Insights, not Products.
I advise planners to avoid product-driven reviews and instead focus on insights.
- Product-Driven Mindset
Most planning software and planners enrich forecasts by reviewing products. To do so, they will use various prioritization methods (like ABC segmentations) where the core idea is to focus on the biggest (or worst) products. But it’s not because a product is stable, erratic, high-volume, or low-volume that humans are likely to add more value than a model.
Remember, if you can beat your model without specific insights, it means your model needs to be improved.
- Insight-Driven Mindset
Instead, we advise to focus on insights: What do you know about future business drivers that your model isn’t aware of? Then, review and enrich your products based on these insights.
In practice, planners should start their review process with insight-driven activities, such as, for example,
- Contacting their clients
- Review new products (and their impact on existing products)
- Discussing promotions and advertisement campaigns with marketing
- Discussing current clients and contracts with sales
Planners need to look for insights, thinking and acting like journalists. Then, based on this information, they can enrich their forecasts based on their assumptions (we can call this insight-driven or assumption-based forecasting). In short, if you know something, do something.
Key Takeaways: A Rulebook for Demand Planners
What are the key practices to implement to unlock demand planners?
Supply chain leaders need to see demand forecasting as an information game, implement a fully automated machine learning forecast engine to do the heavy lifting, and finally track the overall process’s forecast value added (FVA).
What are demand planners’ primary responsibilities?
Demand planners need to: 1. Ensure that input data (both master and transactional data) is cleaned and consistent; 2. Collect insight from markets, clients, sales, and marketing teams; and 3. Enrich forecasts using these insights.
What should demand planners not do?
Demand planners should avoid spending time 1. fine-tuning statistical models (an automated engine should do this instead); 2. focusing on supply or shortages (they should forecast unconstrained demand, not supply-constrained sales); and 3. detecting and cleaning outliers manually (instead, focus on creating a robust data cleaning automated logic).
Should planners use ABC Segmentation to focus their work?
No. The usual ABC segmentation (based on historical volumes and COV) will not point to any meaningful segment. Instead, segment your products based on expected (value-weighted) volumes and historical forecast errors. Better yet, enrich your forecast based on insights.
What is the difference between Product-Driven and Insight-Driven reviews?
Product-driven reviews mean planners first review forecasts with the biggest (or most important) product. This technique is unlikely to add much value as planners don’t have specific insights for these products (compared to the forecast engine).
Insight-driven reviews (or assumption-based forecasting) means that planners will look for specific information regarding future demand (discussions with clients, marketing, sales, …) and then update the forecasts based on these insights.
