Outsmarting Algorithms (and Firms): Anticipatory Behavior in the Labor Market

Earlier this week, the LA Times had an interesting article on the discriminatory nature of the gig economy’s algorithmic pricing model.

Meaning different workers receive different offers with different “wages” simultaneously:

“Avedian shared an experiment he ran in which two Uber-driving brothers in Chicago sat side by side with their apps open. They recorded in real time which rates they were offered for the same ride — and one brother was consistently offered more for every trip. The brother who kept getting higher offers drove a Tesla and had a history of accepting fewer rides, while his brother had a rental hybrid sedan and a higher ride acceptance rate. This suggests that Uber’s algorithm is offering higher rates to the user with the nicer car and who has historically been more picky, in order to entice him onto the road — and lower ones to the driver who was statistically more likely to accept a ride for less pay.”

The crux of the claim:

“The notion that people should be paid the same wages for doing the same work is one of the most fundamental assumptions about a fair labor market. And yet, according to new research from Veena Dubal, a law professor at UC Hastings, on-demand app and tech companies have been undermining this crucial compact in ways that stand to influence the future of work in deeply concerning ways.”

Part of what makes the gig economy model scalable is the “algorithmic boss” it’s based on.

From the workers’ point of view, the gig economy offers flexibility and the option to either supplement income on their regular jobs or to earn money when in between jobs. 

But at the same time, there’s a dehumanizing aspect to gig work, like the fact that workers aren’t entitled to restroom breaks and are solely responsible for their taxes and expenses. Or the fact that there’s more of an adversarial relationship between gig workers and the people that gig workers serve, as exemplified by the declining tips:

“Some restaurants have ended their delivery options. And customers, conditioned during the pandemic to prefer ‘contactless’ deliveries that drivers say now feel dehumanizing, seem less inclined to generously tip someone with whom they’ve barely interacted.”

And, of course, these platforms can decide to deactivate gig workers at any time (more on that in a future post). 

(Un)Fair Wages

So, we’re already looking at a complex working situation. But the article touches on the very essence of this economy: At any point in time, an algorithm is making an offer to a driver to pick a customer or do any other task (pick from a local supermarket or drop off food at a customer’s house) in what the algorithm (in its finite wisdom) thinks will be the lowest amount needed to ensure the ride happens. 

The article points out an essential aspect of any type of work: fairness —the idea that everyone should earn the same wage. 

However, even outside the gig economy, the idea of fairness varies. In most workplaces pay is somewhat related to seniority. And while some may argue that seniority demonstrates an accumulation of knowledge, I don’t think this is always true, and I’m not sure that seniority-based pay is always transparent and fair either. 

There are many ways to address the disparity between how much various workers are being paid, but since I’m not a legal expert, I will refrain from discussing that aspect, as well as the morals surrounding it. 

But an interesting aspect is the fact that most gig workers are able to work for more than one firm. If a company discriminates against them or doesn’t offer them competitive wages or fair treatment, they can switch to another. In most cities, drivers have the option to work for Uber, Lyft, Amazon, DoorDash, Grubhub, Instacart, and several other firms that are vying to access gig workers in peak times (but are unwilling to employ them otherwise). 

Sounds simple? Well, it’s not. 

This type of “hopping” requires workers to anticipate where they should be physically and in which options to look into. It also makes me consider my most recent work on this topic. 

My long-time readers know that I often feature research from people in Operations Research and Operations Management —mainly other people’s research. But once in a while, I share my own work, and this is one of those times. 

Workers’ Anticipatory Behavior

Our just completed research paper, Measuring Strategic Behavior by Gig Economy Workers: Multihoming and Repositioning, is a joint work with Daniel Chen, a Wharton doctoral student, and Ken Moon, a colleague at Wharton.

The paper aims to understand how gig economy workers make decisions by being strategic. More specifically, this paper focuses on two strategic decisions that gig economy drivers must continuously make to optimize their earnings: multi-homing and repositioning. Multi-homing refers to workers accepting jobs from more than one platform, e.g., driving for both Uber and Lyft on the same day. Repositioning involves workers choosing a new physical location to become eligible for more lucrative jobs, e.g., an Uber driver intentionally driving to the airport for their next pickup. 

However, these decisions can be complex due to their anticipatory nature. Both multi-homing and repositioning involve trading off a worker's immediate earnings for future earnings that rely on the expectations of future demand.

For example, when a driver chooses an Uber trip over an available Grubhub delivery, they earn a different amount for that specific trip and end up in a different location, which might impact future job opportunities. Similarly, when a driver repositions, they forgo the immediate earnings from accepting a trip at their current location in favor of the potentially higher earnings at another location. As a result, gig workers must have a forward-looking approach and consider the potential outcome of their decisions when using these strategies.

How often do they do that?

Both activities are common: in our data (primarily from NYC), 56% of drivers work with multiple platforms, and all of them reposition by one or more miles after some trips. Between individual trips, workers switch platforms 9% of the time and reposition by at least one mile, 34% of the time.

But before we move on to discuss the method and the results of our research, it’s important to note that, one way or another, we all engage in anticipatory behavior in our professional life. I’m writing this as the semester here at Wharton is nearing it’s end,f and students come to me with questions about which job offer, among multiple options, to accept. Some students have to choose between a simple product manager position in a big tech firm and the opportunity to join a smaller firm but in a more senior role. Given the advent of AI and ChatGPT-like products, many question the long-term viability of some of these career paths. “What if I take a job that will pay well for the next 2-3 years and then be eliminated?” For those thinking about joining a specific industry (e.g., fintech), how can they know that they’re not joining a sector that will be overregulated in the next few years? All complex questions require tradeoffs and making some assumptions and predictions. 

The gig economy provides an excellent testbed for studying our ability to exercise such behavior because to a large extent, drivers can use past experiences to predict the future. In New York, not all days or times of day are the same, but an experienced driver will know the city’s most popular times and destinations, which are also more likely to generate ride requests. In most professions, people have a sense for these things, but in the gig economy, this knowledge is more precise. 

The Method

We assume there’s a cost for multihoming, which, while not truly an expense, is used to model the difficulty of juggling multiple platforms simultaneously. Think of it as some type of cognitive cost. We also assume that  there’s a cost of repositioning additonal to the cost of merely driving. This is used to model that repositioning may be viewed by drivers as a task harder than merely driving between two points in time. In the event that these costs are not actual, the model will return zero for both. 

To estimate these costs, we have created and tested a model that considers workers' different preferences when working on multiple platforms and changing their position or location. At its core, the model predicts the earnings, destinations, and time spent on a trip that workers can expect when starting a new task on a specific platform and at a particular place and time (this is where the ample data we have is being used)

As they finish their tasks, workers continually choose whether to work on more than one platform (multihoming) or change their location (repositioning) based on what will give them the most benefit in terms of their overall satisfaction (utility). 

When making these decisions, they consider the potential increase in earnings from switching platforms or locations, and compare it to the personal costs associated with multihoming and repositioning. These costs can be seen as obstacles that stop workers from always making the perfect choice about which platform to use and where to be.

This is how it looks:

You may find the actual mathematical model described in detail in the paper. For anyone interested, I recommend reading it. For the rest … 

The Results

We find that drivers find multihoming particularly difficult and costly, such that they  rarely switch platforms even when switching would significantly raise their earnings. Repositioning costs are relatively smaller, as drivers reposition frequently. Behaviorally, our evidence also suggests that drivers find multihoming much more difficult to undertake or less salient than repositioning.

Graphically, this is how multihoming cost varies across the population of our drivers:

As you can see, a significant number of drivers never multi-home, and our model attributes a very high multi-homing cost for that group. 

There is a small, almost equally sized group that almost always multi-homes when the opportunity is there. The low cost indicates that even for a small benefit, they will switch (if, indeed, there is a benefit). 

As for repositioning, most drivers do.

Regarding the drivers’ forward-looking approach, we find  that there’s significant heterogeneity. We measure the “discount factor” they assign to the future. 

Drivers assign little value to the future when the discount factor is low. As you can see, this is evenly distributed, with the majority of the population in the middle. Some drivers are concerned about the future while others don’t care much and take what they believe is the best offer at that moment. 

The first conclusion of this study is that while gig work offers significant flexibility (and a way to combat discriminatory behavior), not all workers take advantage of that. Many do. Maybe even the majority does. But not all. 

Another question we explore is whether it can be beneficial to make it easier for workers to switch between multiple platforms (reducing friction). We found that if the cost of working on multiple platforms (multihoming) is reduced to zero (by making it easier to switch between apps, for example), drivers’ earnings increase slightly to $31.2 per hour. At the same time, the percentage of time spent actively driving (utilization) decreases to 63%, and the number of completed trips (service levels) increases significantly to 75,000 visits.

Even when drivers face more competition for well-paying tasks, they earn more money and work longer hours to complete more rides. However, the increase in hourly earnings is not due to workers being busier; instead, they are positioning themselves to take on higher-paying rides. In other words, the benefits of being able to work on multiple platforms easily outweigh the downsides of competing for the same high-paying trips.

Implications Beyond the Gig Economy

What about us non-gig workers? 

Many of us re-position. 

Many of us multi-home. 

I came to the US from Israel to study. I stayed and took my first academic position (at Kellogg), and then moved to Wharton. In each step, I gave up a future known for a future unknown, but one with new opportunities and significant costs and risks. This is the ongoing question of the optimal level of switching and repositioning. 

The reality is that we don’t do this enough:

“The interstate migration rate has fallen 51 percent below its 1948–1971 average, and that number has been falling steadily since the mid-1980s. Or, if we look at the rate of moving between counties within a state, it fell 31 percent. The rate of moving within a county fell 38 percent. Those are pretty steep drops for a country that has not changed its fundamental economic or political systems.”

Not everyone can move, and the inability to move was driven by the 2008 financial crisis (and home pricing) as well as issues related to health insurance and the risk of losing established benefits. 

But the discrimination mentioned in the article, which sparked this post, is much more pervasive than in the gig economy, and as the major layoffs in tech firms have shown, professional life can be very fickle. 

One of the main lessons I try to teach my students is that they have to engage in anticipatory behavior. 

You don’t have to plan too many years ahead. 

You can’t. 

But don’t focus on the best job you can take now, instead, ask yourself, “What’s the best job I can take now that will position me best for the next job I really want?” And if you don’t know what your next job will be, that’s ok. Pick the job that will maximize the optionality once you are done. 

In a few words:

Try to reduce your repositioning cost. Try to reduce the cost of switching jobs. And try to find a way to account for future options. This is my three-sentence graduation speech.