Analytics-driven forecasting means more than measuring trend and seasonality. It includes all categories of methods (e.g. exponential smoothing, dynamic regression, ARIMA, ARIMA(X), unobserved component models, and more), including artificial intelligence, but not necessarily deep learning algorithms. That said, deep learning algorithms like neural networks can also be used for demand forecasting, depending on the situation.
When it comes to demand forecasting and planning these days, many people believe demand-driven and market-driven are interchangeable. There are some basic similarities - you need to forecast true demand, and use more advanced analytics to sense and shape future demand. We all know that the true demand signal is POS (point-of-sale) channel data, and/or syndicated scanner data (Nielsen Company and Information Resources, Inc.), right? We also know that shipments are the supply signal, and sales orders are the replenishment signal. So, if you're sensing sales orders, you're not really sensing true demand, but rather replenishment. So, what signal are you forecasting?
Artificial intelligence and machine learning
There’s also been a lot of hype and discussion regarding artificial intelligence (AI) and machine learning (ML), and rightfully so given that we've solved data collection and storage challenges, as well as processing and scalability challenges. Even with all the buzz about AI and ML, not many are using either for demand forecasting - or if they are, it's on a one-off basis, not on a large-scale across the entire business product hierarchy.
In fact, I’m finding that the majority of companies who have a formal demand forecasting and planning process are still using 1990’s forecasting processes, and applying 1980's mathematical methods (e.g., moving average, and/or non-seasonal exponential smoothing). In fact, many companies feel that AI and ML is the new demand forecasting “easy button,” which I remind them only works in those Staples commercials.
You still need data scientists to monitor and tweak models using analytics-driven methods to make corrections if something dramatically changes. And demand planners are needed to own and manage the demand planning process. Close to 80 percent of a demand planner's time is spent managing data and information - not a productive use of their time. As a result, demand planners do very little real analytical forecasting.
So, why are companies so excited about using AI and ML for demand planning?
Maybe they think with all the digital disruptions, AI and ML can replace demand planners and bypass the need for data scientists, creating real-time demand execution. I’m not sure how these same companies plan to move from a 1990’s demand planning culture using 1980's mathematical methods to AI and ML overnight. Almost all of the companies we talk to are currently using moving average and/or non-seasonal exponential smoothing methods supported by Excel. They continue to cleanse the supply/replenishment history (sales orders or shipments) into baseline and promoted. We all know this is a bad practice (see "Stop cleansing your historical shipment data!")
You can’t measure all the demand patterns if your mathematical methods can only measure trend and seasonality, and then hope that collaborative “gut feeling” judgment can explain away all the unexplained variance. It requires analytics-driven forecasting and domain knowledge, not intuitive judgment.
The answer to all your supply chain challenges
Companies have been led to believe that sales & operations planning/integrated business planning (S&OP/IBP) is the “holy grail,” which will solve all their supply chain challenges. It certainly has a viable purpose, but the process is only as good as the forecast driving it. It’s the old "garbage-in-garbage-out" analogy. If you don’t believe me, then why are most companies who have implemented S&OP\IBP looking for more advanced analytical demand forecasting and planning solutions?
We've also been hearing that the forecasts being produced by legacy demand management solutions are "good enough" for companies using simple moving average\exponential smoothing methods. Even in the face of aggregate level forecasts, accuracy averages between 50 - 65 percent, and the lower mix accuracy is between 35-45 percent. So, what does “good enough” mean? Can someone tell me, please?
Companies need to move to a more analytics-oriented culture if they want to better understand their customers and more accurately predict future demand. It will take time to transition from limited analytics to a broader use because that's an enterprise effort requiring an analytics-driven corporate culture, as well as new analytics skills, horizontal processes and large scale technology.
"Good enough" forecasts driving your S&OP\IBP process is not the answer, and machine learning is not the “easy button.” You need to start with analytics-driven forecasting using large scale automatic hierarchical forecasting technology supported by data scientists who have the domain knowledge, and the advanced analytical skills to monitor, track, and tweak models as the market and consumer preferences change.
For more information, check out my book: Next Generation Demand Management: People, Process, Analytics and Technology.
4 Comments
In some forecasting applications where the business users have used only Excel based projections like CAGR and project these growths as their "forecast", if they don't have time-series background they typically fail to see the value of increased accuracy. They also fail to see the value of analytic based forecasting and not even sure how to measure the ROI since they are not currently measuring it the cost of poor or inaccurate forecasts today. How would you approach such an issue?
Hi Randy,
Thank you for the great questions.
You touch on many of the reasons why analytics-driven forecasting is not well accepted in many companies. The first is the fact that most demand planners are not really planning demand. They are actually planning supply or replenishment. As I mentioned in the article, true demand is POS/syndicated scanner data, particularly, if you are in the retail, CPG/FMCG, auto, electronics, pharmaceutics, or industries where downstream channel data is available for companies' products. Also, a large majority of demand planners do not have strong analytical skills, as they are pretty much managing data and information, and as such, Excel is their preferred technology.
When I worked at Coca-Cola prior to working at SAS, my team provided quarterly "10 year momentum estimates" for the entire Coca-Cola product portfolio. We used time series methods (ES, ARIMA, ARIMA(X), MLR, and other methods). Those momentum estimates were use by marketing for planning purposes, as well as other departments. We would also use those projects to calculate CAGR's.
There is an old saying, "what gets measured gets fixed". So, in order to see the real value of analytics-driven forecasting you have to measure it against the companies current process. It has been found in study after study that analytics-driven forecasts outperform intuitive judgment. At SAS we have developed ROI calculators that demonstrate a direct correlation between improved forecast accuracy and increased CSL's, the reduction of inventory costs, waste, and working capital. Not to mention the increase in revenue and profitability.
What you describe is a classic financial and operations planning view of demand forecasting and planning. Demand planners who report into operations planning are too far upstream removed from the consumer. As a result, they tend to be more operations focused looking only at reducing costs associated with inventory, and increasing order fill rates, and CSL's. I'm not saying that those metrics are not important. Sales/Marketing use POS/syndicated scanner data, not sales orders/shipments. Plus, they are concerned with revenue growth and profit. Not only are sales/marketing using different data, but they are measured against different, sometimes conflicting, performance metrics. In many cases, sales/marketing doesn't show up for the S&OP meetings. In those situations, those companies are doing only OP, not S&OP. The key to the success of S&OP is shared horizontal performance metrics between sales/marketing and operations planning, not vertical performance metrics. The only performance metrics they share is CSL.
I would recommend moving demand planners downstream closer to the consumer in sales/marketing. They don't necessarily need to report directly to sales/marketing. Actually, demand planning should be independent of sales/marketing, operations planning, and finance reporting into an independent department that reports directly to the CEO/President. Plus, sales/marketing should be accountable for the unconstrained demand forecast. Integrate POS/syndicated scanner data into the process using consumption based modeling driven by analytics-driven forecasting--use the MTCA (multi-Tiered Causal Analytics) process to link consumption to supply (sales orders or shipments). Use "What If" analysis to sense and shape demand using those factors that influence demand (i.e., prices, sales promotions, in-store merchandising and others) to close the gaps between financial objects and current market conditions, not intuitive judgmental overrides. Implement a set of common (horizontal) performance metrics for the S&OP process. Finally, introduce FVA (Forecast Value Added) to measure the touch points in the process to demonstrate that analytics-driven forecasting outperforms intuitive judgmental overrides.
For more information, read "Next Generation Demand Management: People, Process, Analytics and Technology".
Hi Charlie,
You brought up some very valid points in the context of demand management. In retail specifically, I have seen that two teams ( Marketing and Supply chain) work in isolation and I sometimes find this challenging to convince them on bringing the their aspects together. They also fail to see the value in bringing the two worlds together and less prone to make anay change in the process. How would you approach this issue.
Hi Shweta,
Thank you for the comments.
There as several things your company can do to bring together marketing and supply chain together in the demand planning process.
First, implement a S&OP (Sales & Operations Planing) process. This process requires inputs from sales, marketing, and operations planning to come to consensus regarding the demand plan. Some companies are now transitioning to an IBP (integrated Business Planning) process. IBP adds additional dimensions regarding financial inputs, and marketing aspects. However, in order for S&OP/IBP to be successful sales, marketing, and operations planning must share common horizontal performance metrics across all the different organizations. Also, it also requires more accurate forecasts using more advanced analytics to effective model all the marketing inputs, and run 'What If" simulations to not only sense the demand signal, but also shape the demand signal.
Second, marketing's primary role is to increase revenue and profit, while operations planning's role is to reduce costs. So, that puts them at odds with one another. Also, most demand planners report into operations planning to far upstream removed from the customer/consumer, and they are actually creating a replenishment/supply plan as they are forecasting sales orders or shipments. Marketing is downstream closer to the customer/consumer and in many cases are using POS/syndicated scanner data. Not only are they at odds with one another, but they are not using the same data. It's like Marketing is on Mars and Operation Planning is on Venus. So, you need to embed the demand planners downstream in marketing, and train them with more analytics skills, or hire data scientists. Then, conduct consumption based modeling using the MTCA (Muti-Tiered Causal Analysis) process linking consumption (POS/syndicated scanner data) to the sales orders/shipments data as a leading indicator. This will require a corporate culture change, new horizontal processes, investment in demand planner's analytics skills--or, hire data scientists, as well as new large scale analytics technology. In addition, this will require ongoing change management to only gain adoption, but for it to be sustainable.
Third, make marketing accountable for their forecast inputs. In other words, build in forecasting accuracy into their MBO's. This will also require change management.
These are only a few approaches. I am sure there are others. For example, implement FVA (Forecast Value Added).
I hope this helps answer your question.
Charlie