Some thoughts on the difference between ordinary demand planning and AI/ML-enabled demand understanding.
Full Article Below -
Here is a question: Isn’t it a limited view that production is based on demand curve/price elasticity as a key or definitive decider of demand and supply?
Within that formula, we have a view that is: If we change the price, say down, the desirability of the product goes up and, therefore, we have to produce more supply.
Are the Old Models Obsolete?
Innovation is moving along rapidly in supply chain these days. Users and solution providers are learning together how to apply more advanced analytics and data. But so many organizations are still relying on past philosophies of demand and supply. Too often, price elasticity curves (and the accompanying promotions)1 drive forecasts for production.
We all know price is not the only factor that influences demand. However, in the past, with fixed concepts of data, this was all many forecasters could analyze and calculate. “Oh, we do a lot better than that now,” you are thinking. Even if your organization has more nuanced forecasting systems, most companies have basically ignored many important inputs in their forecasting systems.
Marketing input—Marketing people often chafed at price elasticity as the method, instead, presenting other views and market analyses of demographics and other methods for defining customer grouping, as well as the customer’s desire for/need for a certain product or service. In fashion, “prestige” or “budget shopper” laces marketers’ language. Since they include income data in their spreadsheets, these various views allow them to conjecture pricing and demand for certain groupings of customers. Problem. It’s all in the spreadsheet! Or, for ultra-modern users, it’s from social data gathered with natural language processing from a separate platform and, therefore, it is hard to synthesize/include these inputs systematically. And even if marketing does have a system, the credibility of the data is often not validated.2 It becomes an artifact of the recent past as production gets rolling.
Product design/engineering input—Product people also chafed at the demand/supply elasticity curve as it did not take into consideration product attributes and features and their utilitarian and sustaining value. Importantly, price elasticity does not take into consideration the customer’s ability to actually evaluate the increased value they receive and, thus, their justification of an increased price. After all, how do you measure the reasoning process of each consumer? (Turns out there are several ways, but that is for another article). Product designers often supported their argument with competitive analysis based on feature/function. Again, this data does not fit neatly into the forecasting system your company may have purchased as recently as two or three years ago.
Logistics input—Then we have the issue of availability. Currently, consumers in more “developed” economic sectors have access to lots of home shopping/delivery options to get deliveries, even within an hour. Of course, they pay for it in a variety of ways, either through delivery payments, minimum order quantities, or some other factor. In other words, free shipping and, possibly, increased ticket price accompany the availability/convenience. Delivery today plays a big role in demand. And one way or the other, consumers are paying. We must find a way to fit this data into the forecasting system.
Historically, we could look at some of these factors and attempt to glean whether our strategy and investments in customer segments or product development were effective or not. We say “attempt,” since there are so many other factors that affect demand. Fact is, in the very recent past there was no consistent way to validate that the results could actually be tied to the causals. Many organizations today still don’t practice building smarter models based on these and other demand-factor inputs that they may even know about through experience.
Widening the Horizons in Demand
So, let’s dive a little deeper into this whole discussion about widening our ability to understand demand and supply. Simply, products and pricing strategies can be looked at in two fundamental ways, assuming the necessity of the product: commodity,highly available from multiple sources (often cheap sources of production); or limited supply, often single source (where often production cost is higher, with expensive material, but not always).
Commodities—A popular commodity like soap, we know, is a necessity. Soap will be bought. But it is also a highly competitive market with several equal or almost equal offerings. As well, there is abundant supply. Thus, buyers often feel little pressure, in aggregate, to buy without some incentive at a given time. Therefore, the amount of soap and other abundantly available consumer goods you sell, to a great extent, can be gauged on price. This is why brand companies and their channels expend so much effort on promotions.
However, promotional management comes at a huge price. To wit:
Packaging and set up costs
Loss of sales at run up to and post promotion
COORDINATION WITH CHANNELS
Excess inventory if promotion does not meet goals
Those are just a few costs associated with this questionable practice. (Of course, promotions are a great way to introduce the market to new products or to bleed off excess or sunsetting inventory.)
Commodities products can, at times, have so much availability that, in essence, they are dumped in the market. Yet we can still spend a fortune on the associated information, relationship management, tweaking and tuning, and logistics costs and coordination.
Often people say, “Well we had a pretty stable forecast.” But if you dig more deeply, the whole system, technology, and the supply chain processes were crafted to stabilize things, rather than fine tune sensitivity to customers. Today, we can change that.
An oft quoted statement of Michael Dell was, “If we can’t forecast the products, make the products more forecastable.”
In many industries that means reducing the number of choices.
But remember, often not stated, but part of the ecommerce explosion is that along with the convenience of getting things home delivered is the explosion of product and packaging choices offered to the customer.
This means we must widen the horizon of how we forecast.
Single source/limited supply—These are products for which there is only one source, or limited sources, of supply. Generally, sellers in these markets can claim higher prices for their products. These conditions may be a permanent factor of supply markets, or temporary. For example, some products, such as lifesaving medications, will be bought at any price. On the other hand, the condition could be temporary, such as in our recent past when we went through a significant supply shortage of toilet paper. Though traditionally this would be a commodity product, toilet paper, hand sanitizers and disinfectants, and so on, recently moved into the category of short supply, and prices moved up accordingly. No matter how much money we threw at the supply chain, we could not produce more in a time interval. No matter how much money the consumer had, they could not buy a roll of toilet paper at the grocery store.
Fixed forecasts are out. Flexibility in forecasts is in. And in reality, that kind of always was the truth. Products, and the calculating approach used to plan them, may not be fixed values with a fixed relationship between demand/supply and price.
Whether a regular part of the landscape, the normal process within the cycle of product introduction and fulfillment, or emergency market issues, grappling with these issues requires a widened horizon in looking at, understanding, and, therefore, forecasting and managing demand and supply. As we have stated many times, the impact of weather, a social trend, or some kind of national emergency can be understood and expressed as part of an analytical model—i.e., as fields, parameters, values, and equations—to capture, calculate, and communicate said impact and its resultant values.
Enter the Non-Fixed Parameter
So, let’s get back to toilet paper. What changed here?
If we think about traditional forecasting systems, the forecasting parameters are fixed—in the toilet paper example, we use history + some safety stock value and demand flows by channel/customer/location: 2- or 1-roll packs go the convenience stores; 12-roll packs to small form grocery chains; 12-, 24-, 36-roll sizes to grocery stores; and 48- and 72-roll packs to the clubs (BJ’s, Costco, Sam’s); pallets with these cases/packs to the wholesaler.
Within safety stock, there is a value that either a planner or the system set based on history. In this example, my field has a fixed parameter/a fixed forecast method—reorder point, weekly forecast/order from channel, maybe a little MAPE thrown in. This is the way we did things. Now, however, history, channel, and safety may not mean much with a range of factors influencing how we plan even the simplest things. History can tell us a lot, but not maybe in the fixed way it did before.
Grocers and distributors know,either by history or intuitively, the dynamics that previously impacted actual sales or shortages. For example, severe weather (a hurricane is expected on October 21, 2020) will create a run on food, water, and toilet paper before the storm. Hence, the lower limit on safety may no longer apply, since demand will spike and we don’t want to run out. Or an outbreak of flu creates a run on the health and pharma aisle for cold and flu remedies, as well as on foods like juices, soups, and so on.3
How would this flexibility in parameters, methods, and data work? If I have an ability to flex or change my parameter for toilet paper for each storm, say based on category of the storm, I can select a time period that contains the last major storm of this magnitude, say, August 28, 2019. I can select with the sales history date-range and leverage that forecast. I can also change the actual way I utilize those specific values, for example:
Limiting the range of the value. This can be automated or manual. For example, a pallet or trailer size affects an upper or lower limit (an automatic input). Manually, our confidence in, say, sky-rocketing demand, or our doubts about the channel’s forecast may cause us to fix an upper (or lower) range.
Smoothing. Again, this can be automated or manual based on historical values or input from customers or suppliers (we can only handle X factor of receiving each week or we can only produce Y each week).
Best fit algorithms that reflect the statistical characteristic of the product or variant.4
In essence, most of the fields, the ranges, and the actual forecast method used can change for each product at each location at any time. In theory, we could continuously tune our parameters every time we forecast. In practice, we might not want to do that since our mind can’t grasp all these dynamics, and too many changes to tasks and processes challenge the auditability of what we are doing. Certainly, when we have major changes, we will want to tune or change our parameters.5 And we would want this to occur automatically. Whether the systems are autonomous or just send alerts to the user, we do need the system kind of running on some level of autopilot since this is just too much for a human, or even a team of humans, to figure out.
Conclusions—Dynamic, Expanded, Continuous, and Autonomous
We know that the idea of using history, alone, in forecasting the future was somewhat questionable, at times.6 While that approach may not be obsolete, it is limiting, since now we can look at an expanded view of our markets, customers, channels, suppliers, carriers, and so on, and see the many demand-impacting factors as well as the dynamic interrelationships between them that we never saw before. And if we are customers of a large supply-chain provider, we can now gain the benefits of an aggregated view of history and multiple in-process logistics flows to monitor and react to changes. We can see that by the SKU for each time increment, allowing constant tuning and optimizing of many of the factors where the money gets spent and made—price, order quantities, supply volumes, inventory investments, safety stock, on and on.
We now have the ability to capture all that data, have a machine digest and continuously learn from it, and produce some insights. Using formulas, the system can pick the right forecast method based on all of the factors.
Why would we want to do this?
We now have the ability to dynamically change plans, safety stock, production values, warehouse stores, and so on based on the variety of demand impacts and their interrelationship at a specific timeline and location, giving us much better perceptions and actions.
Based on this, we can improve sales and profits by producing at the right time, reducing logistics costs or excesses; and at the right price, optimizing prices across the lifecycle of a product and with more dynamic pricing schemes based on market characteristics.
This is increasing sales, increasing profit. A real win. But we do have to make a lot of changes to get to this point. And that means upgrading the data and technology. If the recent past has taught us anything, those who had a more robust and smarter supply-chain operating model or had the smarts to react early (read A Crisis is a Terrible Thing to Waste, in this issue) are doing
OK. And those that didn’t, well, we can hope there is still time.
1 Consumer-oriented supply chains from automotive, electronics, apparel, food, health and beauty (thus, the chemical industry) and many other industries at some level promote products through price, thus, using price elasticity within the forecast model. -- Return to article text above2 Today there are methods to build consensus using machine learning. This topic will be further explored in an upcoming report on AI/ML use cases in supply chain. -- Return to article text above 3 As another example, currently, there are floating shortages of flu and other vaccines. They may compete against a COVID-19 vaccine if one becomes available soon, by clogging up manufacturing and logistics chains. -- Return to article text above 4 There is a lot more to these parameter issues and we encourage users to discuss them with their solution provider. ChainLink will discuss the nitty gritty of methods in upcoming publications. -- Return to article text above 5 There will be a lot more on this topic in upcoming reports from ChainLink in Q4. -- Return to article text above 6 In the case of toilet paper, probably the total consumption didn’t change, but where it was needed—i.e., at home vs. office or public/commercial facilities—changed. It was the same with food consumption. A lot of the change was to home rather than commercial use, but pack size/volume orders also changed with the channel: 5-pound bags of flour for the home vs. 25- or 50-pound bags for the restaurant. If you consider this factor, it is easy to see that possible supply chain breakdown could occur, not in milling, but in packaging. -- Return to article text above
To view other articles from this issue of the brief, click here.