One-number plans are bad plans. A one-number forecast, for example, looks across the range of demand that might emerge in a period and decides to create a forecast that is mid-way between the most and the least demand that is apt to emerge. A probabilistic approach to planning would be better, but our supply chain tools have not supported this approach well.
Matt Hoffman, vice president of product and industry solutions at John Galt Solutions, uses the analogy of a trip to an airport to explain why a probabilistic approach is better than one-number planning. John Galt Solutions is an innovative supplier of supply chain planning solutions.
“Probabilistic planning is a lot like trying to figure out what time you need to get up and what time you need to leave for a 5:00 AM flight. So the night before, when you’re setting your alarm, you don’t know what that traffic’s going to be like.” You need to consider the probability of congestion. “But not just the traffic, but also the chances of an accident, the impact of weather, and other things as well.” The decision one makes impacts how much sleep one gets and the likelihood of arriving at the airport on time.
This is similar to the problems companies face in supply chain planning, where they have to deal with variable lead times, inventory levels, demand, and information lags.
Mr. Hoffman asks us to imagine we are planning a trip to a metropolitan airport. Google Maps will show various routes with estimated route times if you look at the app the night before the flight. The longest route, which routes a driver around the city center, takes 45 minutes. The fastest route, which takes a driver right through the heart of the city, only takes 25 minutes.
The fastest route is the most likely to have severe traffic congestion, but for a 5 a.m. flight, that probably won’t be a problem. However, a traffic accident on this two-lane road could lead to severe delays. The slowest route is on an eight-lane highway. Even if there is an accident, it has a very low likelihood of impacting the drive.
Further, is the person planning to drive their own vehicle to the airport? Do they need to add time for dropping off a rented car? If so, how much time should they add? Or are they taking an UBER? What are the chances the UBER will arrive on time?
Also, how long will it take to clear the security check? Does a person have TSA PreCheck or do they need to stand in the longer line? And at some airports, a Fast Pass really is not any faster. What is the situation at this airport?
What an individual plans to do when they arrive at their destination is another factor. Do they need to get right to a conference, or is the schedule looser? Further, how many flights are there per day? Would it be possible to get another flight in two hours if they miss the first?
Finally, perhaps the most crucial factor is just how risk-averse a person is. I have relatives who are nervous if they are not sitting in front of their gate two hours early.
In short, most people intuitively use probabilistic planning all the time without even realizing it.
John Galt’s New Solution for Probabilistic Planning
John Galt Solutions will introduce a new end-to-end planning solution based on a probabilistic framework in the third quarter. “What we’re looking to do,” Mr. Hoffman explains, “is move beyond a rigid structure of planning with a single number to accounting for a full range of complexity and uncertainty when developing a plan.”
John Galt is using a Markov Decision Process to quantify the value of supply chain processes and power a more comprehensive model of real-world supply chain probabilities in their Atlas Planning Platform. For example, in a trip to the airport, there is a chain of events that determines whether a person will arrive on time. Each event is a “state.” I slept in late, or I didn’t. The Uber arrived on time, or it didn’t. There is a probability attached to each state. What actually happens in one state impacts the next state. That is a “transition.” Based on the states and the transitions, a person can change the actions they take. Those actions have different costs or rewards associated with them based on the potential value of making (or not making) an action at different points in the future.
The Markov Decision Process is effective at estimating value in an environment where long-term costs and benefits are assessed in the context of uncertainty at each step in the decision process. To create the plan, the planning engine runs thousands of simulations at different stages of the process, with different probabilities attached at each step. John Galt Solutions is using “Q-learning” to drive the simulations. Q-learning is a machine learning-based approach that generates a range of relevant simulations.
This is not a static plan. As the likelihood of different factors changes—such as through an adverse event—Atlas re-optimizes based on this “new view” of risks.
The analogy between planning a flight and creating a supply plan is good, except that a supply plan is far more complex. The supply plan has many more variables to consider. There are lead times. However, it should not just be an average lead time; companies should have lead times for every raw material they buy. Every different product has different demand variability associated with it. The throughput time on a factory line can also vary depending on the age of the equipment, what is being made, and even the order in which goods are being made. Product yield is variable.
Traditional planning is based on a single number forecast. The supply plan is designed to meet the one-number forecast. However, the demand-supply match is done with a limited consideration of risk, uncertainty, and vulnerability.
A probabilistic plan still has an initial plan, a one-number plan, but it is based on a stronger understanding of the range of supply chain outcomes and how those supply chain outcomes affect core business goals. This kind of planning understands whether it hurts more to miss high or to miss low. And it understands just how much a miss hurts across the range of possible outcomes.
In short, probabilistic planning better aligns with a company’s tolerance for risk, goals, and strategies. Despite a more rigorous approach, a probabilistic plan will also be wrong. However, a probabilistic plan allows a company to pivot faster when new information emerges. So, rather than running away from uncertainty, probabilistic planning embraces it and rigorously plans around diverse sets of variability across a plethora of essential factors.
