Managing truckload procurement variability under uncertainty
There are a few things that truckload trucking and Wall St. have in common — managing resources (trucks and money), responding to markets (demand for loads and drivers, or competing for investors), and dealing with uncertainty.
On Wall St., uncertainty is endemic to the nature of the field. You might say that finance as an industry would not exist without uncertainty. Not surprisingly, Wall St. has developed a variety of strategies for managing uncertainty, which can all be grouped under a single term: hedging.
A recent Medium post from the FreightLab in MIT’s Center for Transportation and Logistics highlights the costs associated with truckload procurement in the presence of predictable variability (hour-of-day, day-of-week, seasons, holidays) and uncertainty (which represents random deviations from expectations), and tries to offer a solution by separating predictable, well-balanced traffic lanes from those that fall elsewhere on the spectrum of balance and stability.
In this article, we are going to argue that what is needed is accurate models that make it possible to design strategies that work in the real world. In the process,
This article will make the following points:
- The MIT approach ignores the fundamental interactions of truckload networks, which will make it impossible to separate “nice lanes” from others that are more variable and less balanced. I believe that the MIT methodology is guided by the deterministic models that were first developed in the 1980s and 1990s, which create unrealistic expectations for shippers when carriers submit bids since the bids are derived from models that ignore variability and uncertainty that represent a fundamental fact of life in freight transportation.
- I will introduce a new class of tools developed over 40 years at Princeton University at CASTLE Lab, and recently licensed to Optimal Dynamics, which are based on a new unified framework for optimization under uncertainty. The tools, marketed as CORE.ai, span strategic planning (using a highly detailed and carefully calibrated simulator), tactical planning (which plans activities 1–2 weeks out, performing load acceptance and planning drivers while recognizing the uncertainty in forecasts), and real-time operational dispatch.
- The tactical planning tool in CORE.ai will actually reduce variability and uncertainty by planning a week or more into the future, identifying the loads that will fill in gaps given the current status of drivers and loads. It captures the uncertainty by simulating 20 different scenarios, and computing the probability of covering each load and getting drivers home.
- I point out the power of using the full network of a carrier, in addition to the different forms of substitution, to absorb the variability and uncertainty that is endemic to the truckload industry.
- The strategic simulator can be used to compute load coverage curves that give the probability of covering different numbers of loads in a lane on a given day. Carriers can influence these curves through incentives, where higher incentives provide higher service, but at a higher cost.
The CORE.ai toolbox will provide realistic estimates, not available from current deterministic tools, of how often a carrier can meet the expectations in a bid, and how often the shipper will have to climb the ladder of higher bids. The level of coverage is controlled by incentives, which not only allows the carrier to strike a balance between cost and service, the shipper can also see the true cost of higher service.
We believe that both carriers and shippers can perform better planning and set more realistic expectations by using more accurate models that capture the dynamics of truckload networks.
The uncertainty challenge
In the trucking industry (and throughout business, science and engineering), people endlessly struggle with uncertainty, which refers to deviations from what was forecast or expected. The most common response that people have when faced with uncertainty is to ask the question: how do we get rid of it?
The approach most often quoted to eliminate uncertainty in business settings is to improve forecasting. There is today the widespread belief that this can be achieved using “AI” which has produced a rush to use the latest fad in the form of neural networks (especially deep neural networks, since bigger must be better).
The problem is that neural networks have so many parameters they tend to overfit noisy data (see my post Managing problem structure and noise in machine learning) which limits their predictive power. This means that throwing big machine learning models at complex problems such as forecasting the number of loads (an activity that is inherently uncertain) may actually add uncertainty due to the problems of overfitting.
One (of many) areas where truckload motor carriers have to deal with uncertainty is in the process of bidding on freight. Shippers will provide files of aggregate freight volumes, typically drawn from the previous year’s history which is their best estimate of how many loads will move in each lane next year. Needless to say, this provides minimal guidance about how many loads might move in a particular week, and even less guidance about the loads moving on a particular day.
Using this rough guidance, carriers are supposed to prepare bids that capture their costs for moving freight in a lane. These costs depend on the entire network of flows, since drivers in one lane feed capacity to other lanes. These interactions depend on how many loads are moving to each destination and when, none of which are known with any precision when the carriers have to prepare their bids. Note that this did not stop the development of what became known as “combinatorial auctions” which are little more than deterministic models of future freight flows.
Now pity the shipper. The carrier will bid to move a certain number of loads in a lane, but again, this is done on an aggregate level. So the carrier may agree to move on average 20 loads per week in a lane, but if they don’t have enough drivers to serve these loads on a given day, well, … they don’t, forcing the shipper to turn to more expensive carriers.
In a recent Medium post “In Search of Alternatives to Truckload’s Fragile Freight Contracts,” Chris Caplice of the FreightLab in MIT’s Center for Transportation and Logistics notes:
To an outsider, it might seem odd that contractual agreements that are binding in price, but not in volume or capacity, are dominant in the truckload transportation industry. These contracts are used because it is the only way that shippers and carriers have been able to handle the volatility and variability of trucking demand and supply. But they do add costs and reduce the efficiency of freight networks.
The post goes on to describe a process from the 1990s known as “combinatorial auctions” which are basically deterministic optimization models that choose loads for a carrier’s bid by optimizing flows (deterministically, and statically) over the network. Perhaps it is not surprising that there are problems with implementing this strategy. In fact, the blog is quite honest in noting:
The transactions are completed when the freight is tendered and delivered by the designated carriers — but not always. All too often, the procurement process breaks down at this second moment of truth when an actual load is tendered to a designated carrier.
How do you fix this? A well-worn path preferred by business professionals is to get rid of uncertainty, such as through better forecasting (a goal mentioned by CTL director Yossi Sheffi in previous interviews). This is a pipe dream in truckload trucking since the number of loads moving in a particular lane on a particular day is influenced by an entire spectrum of factors, including production capacities, shipping policies, short-term variations in market demand, longer-term market trends, weather, …
The availability of drivers in a lane is governed in part by the number of loads moving drivers into a location (which means subject to all the same sources of variability), compounded by driver management issues (of which there are many), equipment problems, traffic delays, and dwell times at pickup and delivery points.
I claim that truckload trucking is like Wall St. — uncertainty is a fundamental property of markets, whether it is a truckload market or a financial market. Wall St. has developed a set of tools to work with uncertainty; truckload markets (carriers and shippers) have to do the same.
The MIT solution
The Medium post by Chris Caplice describes their solution strategy:
Is there a better way? MIT CTL’s FreightLab believes so and is developing alternative contract formats that solidify full truckload agreements and give shippers and carriers more certainty about future freight movements.
In fact, the MIT proposal does nothing to reduce uncertainty. Instead, it just produces a strategy for classifying lanes into three broad categories based on the degree to which they are balanced and stable:
A more nuanced approach is to classify lanes according to where they are positioned between two extremes: those associated with consistent, balanced freight flows, and those where freight patterns are inconsistent and unbalanced. The type of lane is matched to the type of truck transportation best suited to carry the freight. Stable lanes are supported by dedicated fleets (trucks committed to respective lanes). Lanes in the middle of the spectrum are serviced by contract relationships (long-term contractual agreements with carriers). The spot market is reserved for unstable lanes.
The problem is that it is virtually impossible to divide lanes in the way that MIT envisions. Truckload carriers are networks, as illustrated in the figure below showing the flows for a major carrier. In our detailed simulations, plots of driver tours have never shown drivers simply moving back and forth between two points (even when one point is a major source such as a warehouse or manufacturing facility).
Lanes are rarely balanced on average, and even those that are roughly balanced may be highly imbalanced on any particular day. A truck moving in a so-called “balanced lane” might easily be dispatched after delivering a load to another load in an “unbalanced lane” just because there is a load available at the right time. The load in the “unbalanced lane” may offer the driver better miles, or it may be a better choice for getting him home.
If we classify lanes based on balance and consistency, a carrier would lose the powerful portfolio effect of choosing among loads out of a region. In the figure below, we show probability distributions of the number of loads moving out of Indianapolis. These distributions are not independent. If there are more loads to Cleveland one day, there are likely to be fewer loads to other locations. Being able to choose loads across outbound lanes (taking into consideration other issues important to the driver) is comparable to the financial analyst combining stocks in a single portfolio to reduce risk. However, if the lane to Cleveland is balanced and consistent, while the lane to Columbus is not, the MIT system would not capture the ability of drivers to choose between the two on any given day.
Real truckload operations take interchangeability to a much higher level, since the real world is not actually divided into regions and lanes (these are just constructs for reporting). If we only assigned drivers to loads in his region, we might find ourselves with the situation in the figure below on the left. Here we have four drivers and four loads, where each driver has been assigned to the closest load. This solution leaves one driver unassigned in the east, and an uncovered load in the midwest.
This is not how real carriers operate. The company will know that more loads originate in the midwest and terminate in the east, so dispatchers will tend to assign drivers to loads toward the west, a process that one carrier called “checkerboarding.” This is exactly the problem solved by older load matching systems developed in the 1980s, but it is a strategy used by carriers even without an optimization model.
In a nutshell, while all carriers will report flows based on lanes (from region to region, as well as out of, or into, regions), they know that drivers and loads interact across the country. This is because freight is imbalanced almost everywhere, starting with the most obvious pattern of loads originating in the midwest and terminating near the population centers on the coasts. However, there are many local exceptions, some of which are highly seasonal (such as the periods when agricultural products flow out of Florida).
The MIT solution is to identify pieces of the network where the deterministic, static models seem to be a good approximation, and then allocate dedicated fleets to these “stable lanes.” What this ignores is that drivers moving in these “balanced” lanes still need to move over other lanes. Imagine that the lane from Chicago to Atlanta is one of our “balanced” lanes (see the figure below). However, a driver pulling a load from Chicago to Atlanta may be assigned to a load from Charlotte to Detroit (after dropping off his load in Atlanta) because a driver terminating in Jacksonville (FL) would be stranded if he was not assigned to the return load from Atlanta back to Chicago.
The “problem” that has been identified by industry is that the bids based on static, deterministic models produce rates that do not reflect the variability and uncertainty of real world flows. When these variations happen (which occurs daily), spot shortages of drivers will arise forcing shippers up the bid ladder until they find a carrier with capacity, which means they are paying more than expected. This is a natural byproduct of variability, so the “problem” is not that it happens, but that carriers are expecting to be offered the loads contained in the bid package, and shippers are expecting the carriers to move the freight that is offered to them. Neither actually happens. The problem is really a matter of expectations due to simplistic models.
Next, I will introduce a new toolbox based on an entirely new set of modeling and algorithmic strategies designed explicitly to handle the complex dynamics of freight transportation. The fundamental idea is to use realistic models and modern algorithms that accurately capture the problem, rather than older tools that can only solve highly simplified representations. The goal is primarily to use a more realistic model of truckload operations so we capture the natural variability and uncertainty in the industry. However, we can use the tools to actually reduce the uncertainty through better planning and load acceptance.
CORE.ai — Modern tools for modeling variability and uncertainty in truckload trucking
The fleet management tools of the 1980s and 1990s for truckload trucking, which span load matching to combinatorial auctions, use simple, deterministic models that are unable to plan into an uncertain future.
CORE.ai, developed over a 40 year period at Princeton University in CASTLE Laboratory, is a modern set of tools that can plan into the future, even if it is an uncertain future, just as financial planners will plan returns on investments into your retirement without knowing what the market is going to do over the next 20 years. Unlike the older tools based on deterministic optimization, CORE.ai builds on the fields of stochastic optimization, but draws on a new, unified framework for sequential decisions under uncertainty supported by an extensive body of research in the most prestigious journals in the field.
CORE.ai does more than just model variability and uncertainty in truckload trucking; it actually reduces uncertainty by planning into the future. CORE.ai has a tactical planning module for load acceptance that plans up to two weeks into the future, without assuming that we know what is going to happen. It identifies the best loads given the current status of drivers, the current set of booked loads (which are not guaranteed), and the current forecasts of loads from different sources.
CORE.ai works with probabilistic forecasts (rather than the more traditional point forecasts) and draws samples of what might happen, representing any source of uncertainty which might include:
o What loads will be called in, over time, by your A-shippers, your remaining shippers, and your brokerage?
o What loads will be cancelled after they are accepted?
o Which drivers will accept their assignments?
o How long will it take to move the load from origin to destination (which may involve congestion and weather delays)?
o Will there be equipment delays?
o How long will it take to load and unload the trailer at origin and destination?
There are two fundamental strategies used by CORE.ai to handle uncertainty:
o Tactical planning to make decisions now about which loads to commit to moving in the future, and how to manage drivers to balance transportation costs, service to shippers, and meeting driver needs.
o Creating flexibility through substitution.
We tune and evaluate new strategies using the fleet simulator. Since this is a fundamental tool for strategic planning, I will briefly introduce this first.
The fleet simulator
The most powerful planning tool in the CORE.ai toolbox is the fleet simulator, which is a highly detailed, carefully calibrated model which captures the arrival of loads and the dispatching of drivers over time. The simulator models each driver (and load), capturing hours of service, equipment type, and driver domicile. The figure below illustrates the process of optimizing the assignment of drivers to loads, then stepping forward in time (during which we learn about all the sources of uncertainty listed above).
The fleet simulator can be used to estimate the profitability of loads, identify the most economical driver domiciles, and evaluate the fleet size. It can also be used to tune an array of parameters that control the tradeoff between operating costs and revenues, on-time service, and getting drivers home.
We also use this same logic to project driver dispatching in the future, which is the basis of our tactical planning system.
The tactical planning system
The tactical planning system projects loads being called in and driver dispatch over a horizon of a week or two. Load acceptance is probably the most important decision in tactical planning, but not the only one. Carriers have more flexibility working with shippers setting pickup times at the time that the order is booked. In addition, we can do a better job getting drivers home and meeting revenue goals.
The first question in everyone’s mind is: how can we plan a week or two into the future when there is so much uncertainty? We use Monte Carlo simulation (the same tool used by financial planners) to create 20 samples of what might happen over the next two weeks, as illustrated in the figure below. We then simulate the full real-time dispatch system for each of the 20 samples (using the same logic as the fleet simulator), where we model each driver the same way if we were performing real-time dispatching. If we can cover a load with, say, 90 percent probability based on our sample, then we are confident we can cover it in the real-world, given the incentives we would provide once a load is accepted.
The tactical planning system is one of the most powerful tools in our modeling arsenal. Anticipating events farther in the future, even in the presence of uncertainty, provides far greater flexibility for carriers and shippers to work together. The most important decision is whether to accept a load or not, especially when you are given access to a set of loads (such as through a carrier’s brokerage division) where you can pick and choose the loads that best fit the network.
As of this writing, we have tested this idea on over 20 different carriers, and the results are little short of dramatic. We have run careful simulations that allow us to compare the estimated probability of covering a load with the probability that we actually cover the load. These simulations have shown that the actual probability of covering a load is consistently higher than the estimated probability. The benefits have often been in the tens of millions of dollars in increased profits, which could also be used to make more competitive bids.
Creating flexibility through substitution
It is important to realize how we might be covering a load despite all the different sources of uncertainty. Financial engineers on Wall St. will often talk about hedging to handle uncertainty. In freight transportation, the key word is substitution, which comes in several forms:
o Spatial substitution — We illustrated this above when we showed how drivers could be optimized around the country so that surpluses (say, on the east coast) could be used to cover deficits (such as in the midwest).
o Temporal substitution — It is important to pick up and deliver loads within specified time windows, but when we are planning in the future, before we have committed to a load, we have additional flexibility to work with the shipper. While we do not know exactly which driver will handle the load, we can determine the times that are the most likely to work with the carrier.
o Shipper prioritization — Each carrier can identify its most important shippers. Carriers should aspire to the highest level of coverage and on-time performance for their best shippers.
o Driver prioritization — Just as carriers know who their best shippers are, they also know their longest serving, most reliable drivers. We want to make sure that these drivers get the loads that meet their needs.
Better planning through better modeling
Shippers complain when the carrier that submitted the lowest bid cannot handle freight requests, forcing shippers to use carriers that submitted higher bids. The problem, as they see it, is that carriers need to do a better job meeting the commitments they made in their bid. What shippers do not realize is that those bids were based on a simplified model of freight flows where loads terminating in one region create drivers that can be used to move loads out of that region.
The simple reality is that there is always going to be variability in the number of loads, and the supply of drivers. Just as Wall St. has learned to live with the uncertainty of financial markets, shippers and truckload carriers need to live with the reality that there is always going to be variability and uncertainty in truckload markets.
The problem as I see it is the deterministic (and static) models of freight flows which have been used for bidding in the past create a simplistic view of the world, creating unrealistic expectations for shippers. The bids being submitted are based on an idealized, unrealistic model of the real world.
A better approach is to use more realistic models. The CORE.ai system has what is today the state-of-the-art fleet simulator for modeling drivers and loads for truckload carriers. The model captures drivers and loads at a high level of detail (including full hours-of-service, driver domiciles, and equipment types), as well as an accurate model of the booking process with probabilistic forecasting. The model estimates the cost of moving a load over a network while capturing the variability (hour-of-day, day-of-week, seasons, holidays) and uncertainty (including accurate load booking profiles) in the booking process.
Instead of just providing a bid, this model can provide estimates of the probability of covering 1, 2, 3, …, loads in a lane, called the load coverage curve, illustrated in the figure below. This curve captures the inherent uncertainty of dispatch operations, but creates measurable performance goals for carriers and shippers based on an accurate model of operations.
In addition, the CORE.ai fleet simulator makes it possible to shift the load coverage curve through an incentive parameter that encourages higher load coverage for a particular shipper. However, a carrier should be able to specify, for example, 90 percent coverage when they move up to 10 loads on a given day, and perhaps 70 percent coverage for the next 5 loads. It will be up to the carrier to tune the incentives in the model to achieve these probabilities. Of course, increasing the coverage probability will increase costs, forcing an increase in the bids by the carrier. Most important, the model will set realistic expectations based on an accurate model of real dispatch operations.
The CORE.ai fleet simulator can be used by shippers to help them understand how higher load coverage probabilities affect costs. This would be an industry first, since shippers are accustomed to making demands on carriers without expecting to pay for it. We used the simulator to estimate the cost of a request by a major shipper to narrow the time windows. The results, reported in
H. P. Simao, J. Day, A. P. George, T. Gifford, J. Nienow, W. B. Powell, “Approximate Dynamic Programming Captures Fleet Operations for Schneider National”, Interfaces, Vol. 40, №5, pp. 342–352, 2010.
found that it would cost $1.9 million per year to meet these tighter limits, which made the shipper drop the request. Without the model, the shipper might have been able to strong-arm the carrier into accepting the tighter requirements. What shippers do not always understand is that higher service comes at higher cost. The question is: how much higher? That can only be done with a model.
Closing remarks
It is time that we finally outgrow the deterministic models (and deterministic thinking) of the 1980s and 90s. Realistic models capture the variability and uncertainty of truckload markets, including both the availability of loads and the supply of drivers.
Better tactical planning will allow carriers to anticipate problems before they happen. They can accept loads that will fill gaps in the network and help get drivers home. It will also help carriers set realistic expectations for serving shippers, giving the shippers better visibility into the future.
Finally, accurate strategic planning models will allow shippers to appreciate how the dynamics of the truckload market, and their own behaviors, affect the performance of carriers. Ultimately, the two sides need to work together, both in the design of their networks (strategic planning) and the operational implementation, where the most important activity is tactical planning a week into the future.
