Highland Hills Comprehensive Plan Public Meeting #1
Lightning Talk - Transport: Overcoming Deadhead in the Trucking Industry Using Efficient Load Matching
1. Decarbonizing Transport: LightNing Talks
Overcoming Deadhead in the Trucking Industry Using
Efficient Load Matching
Kilian Heilmann, Lyft Inc.
2. Deadhead in the trucking industry
● Deadhead = miles a truck drives empty without a load: wasteful
● Deadhead has negative externalities through congestion and air pollution
● Why do trucks drive empty?
○ Natural deadhead due to geographic and temporal dispersion of economic activity:
there is no load
○ Deadhead due to information frictions creating matching inefficiency: truckers can’t find
a load
● Question: How can technology and marketplace design solve the matching problem?
● Today, I will evaluate the effect of a centralized matching technology developed by a digital
freight broker on deadhead in the trucking industry
3. Background: Truck transportation market
● Very fragmented market with...
○ …millions of shipments (truck loads)
○ …half a million trucking companies
○ …over 17,000 registered brokers
● Uber Freight is a tech-enabled freight broker
connecting shippers to truckers using a mobile app
● The Uber Freight platform processes hundreds of
thousands of truck loads
→ Potential for centralization of the matching
process
4. Load bundling
● Uber Freight introduced feature in
September 2019
● Bundling algorithm creates low-deadhead
combinations of two truck loads to form
round-trips (called bundles), and allows
truckers to book them together
● Advantage: No need for truckers to
browse through thousands of available
loads to create efficient routes themselves
5. Measuring deadhead from the data
● I don’t measure deadhead directly, but we can
infer it from the data
● I only observed loaded miles
● Focus on users that primarily use the Uber
Freight platform and measure deadhead as
distances that are likely being driven empty
● Average deadhead in the Uber Freight data =
18%
● How to evaluate the feature in the presence of
network effects?
6. Evaluating load bundling: Counterfactual
● Basic idea: Compare bundles to round-trips that
were booked without the feature
● Since the algorithm only picks low-deadhead
trips, treatment group is positively selected
● Instead, I go for a before-after comparison and
compare deadhead of drivers that use bundling to
their previous deadhead on the same routes
● Results: Average deadhead decreases by
17 miles or 22.6% per round-trip (even when
controlling for drivers and location fixed effects)
7. Environmental consequences
● Trucking sector in the US has a massive environmental impact (436.5m tons of GHG in
2017) and the potential to decrease greenhouse gas emissions is huge
● We can use emissions factors to translate avoided miles into avoided greenhouse gas
emissions (Assumption: emissions are proportional to ton-mileage, 161.8 g/ton-mile)
● Back-of-the-envelope calculation:
○ Bundling reduces deadhead by about 17 miles
○ A typical trip is around 415 miles long, of which 76 are deadhead
○ Driving empty emits about half the emissions of driving with an average load
○ 17 * EF_empty / ((415-76) EF_full + 76 EF_empty) = 0.022 metric tons per trip
● Bundling reduces per-trip emissions by .02 metric tons or 2.2%
● If the whole industry were to reduce emissions by 2.2%, this would translate to about
9.6 million metric tons per year
I want to talk about an important problem in the trucking industry, namely deadhead.
Deadhead is simply the amount of empty miles driven by a truck.
These miles are wasteful, because they don’t serve any purpose, but they create negative externalities through congestion and air pollution
Now why do trucks drive empty? There are two distinct reasons:
There is spatial and temporal dispersion of economic activity. A trucker might just have delivered a shipment, and now wants to go home
Then there is a very different reason caused by information frictions that create matching inefficiencies. Maybe there is actually a load, but the trucker can’t find them
The question I want to ask today is: “How can…
And “I will evaluate this effect”…
After “can’t find a load”: This creates a matching problem
I want to talk about an important problem in the trucking industry, namely deadhead.
Deadhead is simply the amount of empty miles driven by a truck.
These miles are wasteful, because they don’t serve any purpose, but they create negative externalities through congestion and air pollution
Now why do trucks drive empty? There are two distinct reasons:
There is spatial and temporal dispersion of economic activity. A trucker might just have delivered a shipment, and now wants to go home
Then there is a very different reason caused by information frictions that create matching inefficiencies. Maybe there is actually a load, but the trucker can’t find them
The question I want to ask today is: “How can…
And “I will evaluate this effect”…
After “can’t find a load”: This creates a matching problem
Let me first talk a bit about how the truck transportation market works in the United States:
-It is a highly fragmented market with hundreds millions of shipments every year.
-There are about half a million trucking companies with millions of truck drivers competing for these shipments
-And these million of actors are connected by over 17,000 registered freight brokers
Here I will use data from Uber Freight. Uber Freight is literally Uber for Freight. It offers a mobile app that allows truckers to book loads right on the app.
“It processes”
And the technology has “potential for centralization…”
This is exactly what Uber Freight built.
Uber Freight introduced a feature called load bundling.
Here, an “algorithm creates low-deadhead combinations …”
This has the advantaged that…” “. Taking away the information friction
Here is an example of a bundle: The first load goes from Houston to Dallas, then drives a little bit empty, and goes back to Houston. Overall, out of 563 miles, only 30 are driven empty
So, how do we measure deadhead? Actually, we don’t, but we can infer it from the data
Now, how do I evaluate the impact of the feature?
First, we need to think about what is the counterfactual?
I will spare you the exact calculations
9.6 million metric tons per year in the US