How Machine Learning Improves The Efficiency Of Freight Operations

Just a few years ago, the state of the art in freight brokerage technology was software-assisted pricing and matching. A simple tool might take a base rate for a certain origin-destination pair and add a seasonality modifier based on historical data, then give a human broker a ranked list of the carriers most likely to service the freight. 

Then the phone calls would start, and often the broker would find out that the market had turned and that carriers weren't available or weren't willing to move the freight at the price originally quoted to the customer. If the broker was unwilling to take a loss on the load, it was given back to the shipper or came back with a higher price.

Today, digital freight brokerages are significantly more advanced. Pricing is based on near-time volume, capacity indicators, and carrier quality rather than just historical data, and matching is normally completed without any human intervention at all. But automation now goes far beyond those tasks. For example, Convoy uses machine learning to assess its carriers' service quality, reduce crash risk, and combine multiple loads into round-trips, seeking out optimal routes from countless possibilities.

FreightWaves spoke with Ziad Ismail, Chief Product Officer at Convoy, to get a better sense of the current state of digital freight brokerage technology and how machine learning helps Convoy adapt to the complex dynamics of the freight market. 

Ismail said that in computer science, engineers write a program with explicit rules for handling various situations. But sometimes a system is so complex that it's impossible to plan or even understand all the different scenarios. The freight market exemplifies this type of complex system.

That's where machine learning comes into play. With machine learning, data scientists create a model that takes massive amounts of data and then figures out the rules on its own. The more data the model sees, the smarter it gets.

For example, instead of trying to account for every possible influence on trucking spot rates, a model would look at training data that shows how rates have historically moved given a variety of market conditions. This data provides a starting point for a machine learning model to generate accurate spot rates based on future changes to volume and capacity as well as differences in freight characteristics, pickup and drop-off locations, seasonality, time of day, and hundreds of other variables. 

Machine learning was crucial to Convoy's performance during the COVID-19 pandemic's disruptions to freight markets, which saw large, unpredictable swings in volumes across many different markets simultaneously. 

"The models performed well even when we were seeing unprecedented demand shocks," Ismail said. "There was some additional volatility, but our routing, pricing, and availability models still worked as intended, and we were able to flex our capacity up by more than 50% in March to meet the spikes in demand."

Ismail explained that Convoy views the broader machine learning opportunity through the lens of the shipment lifecycle, where the processes of tendering freight, matching loads to trucks, and hauling can all be made more efficient.

For example, when a shipper tenders a load to Convoy, a machine learning model assigns the load a ‘supply availability score.' This score determines how easy it will be for Convoy to find a truck and service the freight.

"The model doesn't just look at historical capacity trends, but also a range of real-time variables that are constantly in flux," Ismail explained. The supply availability score is based on things like tender lead time, capacity in the market, the density of Convoy's carrier network, the required truck type, whether a lane is a headhaul or backhaul, and other factors unique to the shipment. 

By crunching data as soon as the freight is tendered, Convoy immediately knows how easy or hard it will be to service the load. That means that when Convoy accepts a tender, it has a high degree of confidence that it can find a carrier to complete the job successfully. As a result, the company gives back far less freight to its customers than traditional brokers.

Machine learning is also used to price each load and match it to a carrier. Ismail said it's not just about finding the cheapest truck, but combining price with quality in what Convoy calls a ‘scored auction.' In a scored auction, Convoy is willing to pay more to high-quality carriers that are unlikely to cancel, likely to be on time, and whose records include fewer safety incidents and cargo claims. Every time a carrier works with Convoy, more data on its quality and performance is collected, and this data is fed into the machine learning model. Over time, rewarding higher quality carriers is a self-reinforcing cycle, enabling Convoy to offer more reliable service to its customers.

As each load is matched, Convoy's platform also looks ahead to the location and time of the drop-off. If another shipment is available in the vicinity near the scheduled delivery time, machine learning is used to perform a process called ‘batching.' Here, Convoy can offer carriers a backhaul to boost asset utilization, reduce wasted time, and maximize carriers' revenue. On the shipper side, batches help reduce carbon emissions caused by empty miles. 

Batching loads, Ismail said, is something that computers are naturally better at than humans because of the difficulty of sorting through billions of possible combinations.

"If you're running a marketplace with even 1,000 shipments available on a daily basis, and you're looking at combinations of three shipments at a time, 1,000 cubed is 1 billion permutations. Even at a minimal scale, there is no way a human can do it."

Once a load is accepted and the driver is en route, Convoy uses machine learning to predict shipment delays and, in some cases, proactively fix the problem. For example, if an inbound truck is delayed due to a mechanical issue or traffic, Convoy can often identify that the pickup time is at risk and reassign the load to another truck that can meet the appointment time.

"The goal is to identify problems as early as possible and then automatically course-correct before it puts the shipment at risk. This is important for both shippers and carriers. We are able to give shippers a more reliable service and we can often find alternative loads for the delayed carrier."

Through the use of machine learning, Convoy is trying to find the ‘efficient frontier' of the transportation industry, Ismail said, using a metaphor from finance. The ‘efficient frontier' is a curve that describes the optimal risk-reward ratio of a portfolio of securities – the line represents the best possible level of expected return given a certain risk. In transportation, the ‘efficient frontier' would be a curve that describes the highest possible carrier revenue given a certain level of shipper spend. 

The goal is to reduce shippers' cost of purchased transportation while simultaneously increasing carriers' revenue per week.

"Freight has historically been a zero-sum game," Ismail said. "Either shippers save money, or carriers earn more. The opportunity is to flip that model on its head by driving inefficiency out of the system. The more that we can match the best carriers to each load, provide reliable service, and reduce empty miles, the more we can help shippers drive down costs while also helping carriers earn more. Machine learning is at the heart of that."

This article is published jointly with our partners at Convoy. To view more Future of Freight content, click here.