"Do you want to improve your demand forecasting capabilities?
I have developed a new way to use machine learning to support supply chains. Here's how it works:
I have developed a metamodel capable of dealing with virtually any supply chain demand data set. The metamodel will work with your data to create the best possible ML model."
This is the beginning of a post on Linkedin published on October 12, 2021. The promise is basically "I will improve your forecast with artificial intelligence (AI)".
And I ask, what do we really gain if this were possible? I'm not even questioning whether an "automated learning" algorithm is capable of improving forecast accuracy. My question is even more fundamental. Suppose if it does improve the forecast, what have we gained?
Determinants of inventory size
Inventory is necessary only if customers do not have the patience to wait for the product from the moment they express their need, so the objective of having inventory is to satisfy immediate sales. And we also know that the demand for a particular product at a point of sale has a high variability.
If the objective is to satisfy sales and demand has wide fluctuations, then the inventory we require must be sufficient to satisfy the maximum demand before the next replenishment, which means that most of the time we must hold inventories in excess of the actual demand at the time.
The next replenishment will depend on the time we decide to allow between replenishments, and also on how long it takes for the product to arrive after we order it. Both times together make up the replenishment time.
The longer the replenishment time, the bigger the inventory needed.
Consequences of a large inventory
The larger the inventory, the more space is required to store it, and the more tied-up capital we have. Since space and capital are both limited resources, the larger the inventory, the less variety we can offer customers, which reduces sales.
In addition, if the replenishment time is long, then we will also have higher risks associated with the inventory: risk of scrap and risk of obsolescence. In addition, a lower overall ROI of the operation.
Impact of the forecast on the consequences of the inventory
If our forecast is way off, we will have lost sales due to stock out, and we will have excess inventory accumulation.
If we improve the forecast, we will reduce the stock out, and also the accumulation of excess. However, the main factor that determines the size of the inventory is the replenishment time, and a lower forecast error does not reduce this factor at all, so the required inventory remains high.
I stop here for a minute, because I can already hear the counter argument "if the forecast is accurate, I need what is fair and necessary". I agree. But remember that demand is highly variable: some SKUs will have high sales in the period and others less so, and they alternate. For a particular period, the inventory required is the combination of high and low demand multiplied by a long time.
Why did I deduce that the time is long? Very easy, how many selling days does one need to forecast and how often does one forecast a product line? Once a week, once a month, every other month? It's not every day for all SKUs, that I think is pretty safe to assume.
Therefore, if I am using forecasts, it is very certain that the replenishment time is long. And even worse, if in addition to forecasts, I use MIN/MAX, the time to replenishment is also variable, so you should also forecast how long it will be until the next replenishment. And I am still accepting that we can improve the forecast.
From this reasoning, I deduce that a more accurate forecast, without changing anything else, has not reduced much space or capital tied up. Suppose that the improved forecast eliminates the stock out, and sales increase. The capital tied up will not be reduced very significantly. Remember that before there were stock outs, which means less inventory. Now there is more inventory of those products and less of the others, but the total effect is that inventory is still proportional to replenishment time, so it cannot be reduced very much.
What determines the profitability of a company?
Maybe I should have started here. A company is a system and its profitability depends on how much margin we can generate with its scarcest resources. Another way to look at it is how much money can be generated for every penny spent on trading.
In TOC - Theory of Constraints, Dr. Goldratt defined only three ways to measure money in a company: throughput, inventory, operating expense.
Throughput is the speed of generating money through sales. Inventory is the amount of money trapped in the system and can be converted into throughput. Operating expense is the money spent to convert inventory into throughput
I know that inventory defined in this way can be confusing. Let's use investment instead, and leave the term inventory for the units of products stored.
In mathematical optimization (linear and nonlinear programming), an objective function of the system is defined. In the case of a company, it is the profit. And that function would give infinity if it were not for the fact that the company's resources are limited, so we say that the optimum is determined by the active constraints.
In systems we already know that most of the resources must have slack (see refutation to line balancing) so there are only a few active constraints.
Thus, a company's profitability is determined by what its active constraints are and how they are used.
Space and capital are two constraints that are used more or less depending on the size of the required inventory. If the inventory needed is larger, these two constraints are being used more, even to the point of exhaustion. In that case we are forced to accept a level of stock out because we cannot increase inventory.
Relationship of forecast and profitability
As we have already seen, using forecasts is associated with a long replenishment time, so the aforementioned constraints, space and capital, will be used almost to the maximum. Now it is time to answer, how does better forecasting improve profitability?
A better forecast will make us use those space and capital constraints better, but it will not make us use them less. That is, if they are active constraints, we will be able to "move the needle a little" with more sales by reducing stock outs, but not much more, because the replenishment time has not changed and we will continue to use a lot of those constraints.
What if we reduce the replenishment time?
By reducing replenishment time we immediately alleviate the need for inventory to meet maximum demand. In other words, we can have fewer units of inventory and at the same time that inventory is proportionally larger than the previous one. This reduces stock out and reduces the use of space and capital constraints simultaneously.
By reducing the use of constraints, we can now exploit those constraints in a better way by expanding variety, for example, achieving much higher sales.
In a typical retail environment, outlets can restock every day from their distribution center. And I don't think I'm wrong when I say that outlets receive merchandise every day. What happens is that the fundamental change is that we replenish all SKUs every day.
How does a better forecast make a difference now if we are no longer coping with space or capital constraints? I don't think it makes any difference. And for such a short time, the best forecast is to repeat the immediate past: replenish today what was consumed yesterday.
But there may be changes in demand for each SKU, and inventory levels may not be adequate over time. For that we need to detect in which direction the demand is moving, but we do not require an exact number of units to be replenished. At TOC we have a simple mechanism we call Dynamic Buffer Management, which can be automated and adjusts the investment according to the actual demand. This is the origin of what has been called "Demand Driven".
One feature of this system is that it requires effort only to collect the daily data, which is already collected anyway. And no time is spent processing it, because it is done by a computer (although it is good that there is always human supervision).
When is it appropriate to forecast demand?
The capacity decision is a strategic decision. Normally capacity cannot easily be varied by significant amounts. Doubling or halving are moves that cannot be made frequently and require requirements planning. It is for this type of decision that S&OP (Sales and Operations Planning) is required.
At the capacity level there is a lot of statistical aggregation. It's easy to deduce that. If a company makes 3000 different SKUs, it will hardly have hundreds of production lines. A very large factory has less than ten lines, so the demand for each line has a lot of statistical aggregation. That also makes it possible to deduce that the aggregate demand forecast for each line has a much smaller error than the sales forecast for each SKU.
In such circumstances it is advisable to make demand forecasts to plan capacity expansions. The difficulty in this topic is how people do not understand the exponential function, but that is a topic for another article.
On the other hand, the complexity of today's supply chains also lends itself to AI applications to study capacity utilization at different nodes, such as factories, means of transport, ports, containers, etc. In that field, it is very impressive what Throughput Inc. has achieved with its ELI application (I think in honor of Eli Goldratt).
If I am offered a system to improve demand forecasting for replenishment at the points of sale, I already know that it is a system that operates with long replenishment times for each SKU, so I cannot expect a great improvement in profitability. Yes, there will be an improvement, but not a big one.
On the other hand, without any demand forecasting system, but with a dynamic buffer adjustment system, with short replenishment times, the improvement in profitability will be the same or better than that of the other system, but with less effort and less capital, and in addition to the release of constraints to generate even more margin.