Hacker News new | past | comments | ask | show | jobs | submit login
Artificial intelligence, algorithmic pricing, and collusion (voxeu.org)
200 points by mgulaid 36 days ago | hide | past | web | favorite | 51 comments

This reminds me of the "Tit for Tat" results of simulations of the prisoners' dilemma[1][2], where cooperating strategies won over competing ones.

It also reminds me of Colossus: The Forbin Project[3], a science-fiction movie from 1970 in which a supercomputer in the US and one in the Soviet Union learn to communicate with one another in ways incomprehensible to their human creators and together rule the Earth.

[1] - https://en.wikipedia.org/wiki/Tit_for_tat

[2] - https://plato.stanford.edu/entries/prisoner-dilemma/

[3] - https://en.wikipedia.org/wiki/Forbin_Project

This reminds me of what happened in a game I used to play called 'RoboWar'... it was a game where you programmed robots to fight each other in simulated battles. They held these big tournaments every year.

However, at some points players realized that the scoring gave 1 point for a victory, but also 1 point for a tie (you got a point for surviving a round, whether it ended in a draw or a victory). This lead to a 'pacifist alliance' where robots wouldn't fight unless the other robot also fought. Since both robots would get a point every round, the non-pacifist robots had no chance to compete.

Really interesting outcome to a scoring rule. (Which they fixed by giving an additional point for a kill)


And that observation makes the article doubly weird:

1) Nobody needs an AI to figure out Tit-for-Tat is a good strategy. It isn't complicated.

2) The technical economic term for this collusion is probably something like 'efficient market price', where the sellers have agreed on what the fair price is to offer their service.

Collusion can't mean that the sellers all have an implicitly coordinated price, because the market is always going to settle on an implicitly coordinated price if conditions aren't changing.

Collusion is something like companies agreeing to strategically offer and not offer services to control their competitors. Like, maybe FedEx has a deal where they refuse to carry online goods that aren't bought from Amazon (don't ask me how to implement that), that would be collusion. I'm not getting a service for my online store because there is an exclusionary deal amongst the big players.

> The technical economic term for this collusion is probably something like 'efficient market price'

That's definitely untrue. I mean, clearly an order book shows many sellers with different opinions about what a fair price is, but it would seem reasonable that the price in between the bid and the ask is a close approximation of the efficient market price.

> Collusion can't mean that the sellers all have an implicitly coordinated price, because the market is always going to settle on an implicitly coordinated price if conditions aren't changing.

That's also pretty untrue, or so pedantic as to be meaningless. Clearly when orders are fulfilled on a proper bid/ask market, conditions aren't changing in any fundamental sense but the price definitely does.

You're really stuck applying Econ 101 in one of the worst places, retail. Your line of thinking has been completely co-opted by a very specific kind of consumption--buying all your things from the same few stores--where producers benefit from collusion, meaning coordinating pricing, at the zero-sum expense of consumers.

Ordinary consumers could buy their goods at auction, or from artisans, or in pre-sales. Those are much better examples of situations where secular trends could sometimes show the symptoms of collusion without actual collusion. But sellers on Amazon? C'mon man, sellers coordinate prices via implicit information exchange on Amazon in a way that, if they were allowed to, they would just collude straightforwardly to do so.

> That's definitely untrue.

I think he meant Efficient Market Hypothesis, which could be posed as a strategy.

> That's also pretty untrue, or so pedantic as to be meaningless.

It's still collusion.

> You're really stuck applying Econ 101 in one of the worst places, retail.

This is Game Theory, not Econ 101. I also think it's one of the best places.

> But sellers on Amazon? C'mon man, sellers coordinate prices via implicit information exchange on Amazon in a way that, if they were allowed to, they would just collude straightforwardly to do so.

I've seen it happen. I've even seen collusion with online reviews on there.

Most coordinated equilibria results depend on facing the same competitors repeatedly - partly to learn an equilibrium that is influenced by each player's private preferences, and partly because repeated play introduces consequences for defecting from that equilibrium. The more you expect to see the same opponents again, the more you have to gain for colluding, and the more you have to lose from defecting.

So, the most interesting aspect of AI pricing might not be the AI - humans could probably learn to collude with the same setup. It's the fact that AI enables orders of magnitude more pricing decisions, so bots spend less time discovering or defecting from equilibrium, and more time sharing maximum collective profit from customers.

> Nobody needs an AI to figure out Tit-for-Tat is a good strategy. It isn't complicated.

That's true for the iterated prisoner's dilemma. It's also pretty non-intuitive. It was formalized in 1980 which is cutting-edge for mathematics.

I think collusion can also be using real-time pricing updates as a communications channel. Collusion could even be the Nash equilibrium -- in which case, yes, that's the efficient market (I believe).

> Collusion can't mean that the sellers all have an implicitly coordinated price.

Why? Obviously explicitly coordinating prices is collusion. Implicitly coordinated prices depends on nobody else defecting to a lower price point. If enough players have a tit-for-tat strategy, sellers can converge to a higher than market price, trivially. This is far from a 'efficient market price' derived from perfect information.

I’ve heard the line between collusion and conscious parallelism, viewed from the outside, described like this: A group of people leave a building and all open their umbrellas at the same time. If it’s raining, that’s conscious parallelism. If it’s not raining, that’s collusion.

This explanation misses the key point almost entirely. Obviously, if there is an external reason for everyone to do the same thing, it is probably not collusion.

But there doesn't have to be an external cause for it not to be collusion. If all kebab shop owners in a neighbourhood charge exactly the same price for a kebab, because each of them has decided individually that charging more or less than the others would put them at a disadvantage, then that is not collusion.

Everyone opening their umbrellas when it's not raining doesn't have to be collusion either. It could just as well be fashion for instance.

Collusion requires an agreement to act in a certain coordinated way. Drawing the line is difficult when there is a silent agreement based on tradition or fear or whatever.

For the sake of a more informed discussion, here is a copy of the actual paper:


(Couldn’t find it on Sci-Hub, so I paid $5 for it.)

Edit: in particular, it addresses a question I had after seeing the original article and graph… or tries to:

> On the face of it, one may wonder whether the algorithms are effectively punishing the deviation, or whether instead the price cuts simply serve to regain market share. Looking only at the non-deviating firm’s behavior, in fact, it is hard to tell these alternative explanations apart. But if you focus on the behavior of the deviating firm, the difference becomes perfectly clear. Given that in the deviation period (i.e., period τ = 1) the rival has stuck to its old price, in period τ = 2 the deviating algorithm, which meanwhile has regained control of the pricing, has no reason to cut its price endogenously unless it is taking part in the punishment itself. If its only concern were to maintain its market share, the deviating algorithm would cut its price only in period τ = 3, i.e. after observing the rival’s price reduction in period τ = 2. Actually, however, in period τ = 2 the deviating algorithm prices almost exactly the same as the other. This clearly shows that the deviating algorithm is responding not only to its rival’s but also to its own action. Such self-reactive behavior is often crucial to achieve genuine collusion, and it would be difficult to rationalize otherwise.

The problem with this explanation is that the “AI” algorithm they’re using is ridiculously simple. They talk about its input being “the set of all past prices in the last k periods”… and then, for their main experiment, they set k to 1! So unless I’m misunderstanding something, the input at each round literally consists of both players’ chosen prices from the immediately previous round; the algorithm has no memory beyond that. So all it knows is that its price is lower than its competitor’s; how is it in any way surprising that it would decide to increase the price for the next round?

And yes, I mean increase – the authors seem to claim it “cut its price endogenously”, but on the graph, the price clearly increases at τ = 2 compared to τ = 1. It does keep its price “cut” compared to two rounds ago, before the intervention, but again that’s not surprising since it only has memory of the last round.

Am I missing something?

I don't think you are and thanks for the link. I need to read it again more carefully, but on a first read the results seem completely unsurprising. Seems to be a problem with labeling this behavior "collusion" in that it gives more agency to the agents than they deserve.

This. Is how an informed discussion should be. Based on original references and nuanced positions. For hyperboles, we always have the mainstream media :)

I think we are there for some of the lower demand products even though it's not collusion in the strictest sense.

If an algorithm's goal is to match a competitor's price then it's not possible to find a better price for any item.

In theory, those item's prices don't move constantly, therefore, all prices will be the same across the board. As a vendor, all I have to do is hike my product's price and wait for all the competitors to match it and then hike it again. If I'm a large vendor like Amazon then it's only a matter of time until I can sell a product at my designated profit without worrying that a competitor will beat me.

A smaller vendor has no incentive to lower the price since it knows that Amazon can always win the pricing war so they compete on service or other ways.

Before algorithms and the net, this was not feasible on lower priced items so vendors had to set prices independently. Now, it's easy to just match your competitor no matter the price of the item.

It's not a conspiracy by definition but it has the same ultimate result of algorithms rasing prices.

Somewhat related, there's a great book, Twenty Lectures on Algorithmic Game Theory:

The book is free to download on the author's website, per-chapter, eg. first one:

Highly recommended. After reading the first 20 pages I understood why I innately hate buying/selling cars and real-estate in the real world [and why I'd prefer to do it on Ebay].

It does not even have to be AI. If there is an area where the bulk of humans contact a single agency to decide their prices for them and that agency then tells every caller the same price (adjusted for quality differences), then you will get this style of collusion.

The agency can merely claim it is doing research on competitors to set prices, which is almost true since it looks at the prices it has told others to charge before giving you a figure.

Which is exactly what is happening with rental housing, the vast bulk of which use a shared revenue management system saas.

Also what is happening with wages. Companies relying sharing their wage data with each other through in organizations like Options Impact, and effectively colluding to match each others’ comp.

> Companies relying sharing their wage data with each other

If the collusion was only accidental, then an enterprising company could _just_ increase their price a tiny bit, and attract better people than their competitors.

For this to work, potential employees would have to know about it while competing employers don't.

Even without agency, for e.g. fees determination during a blocks duration in bitcoin. It seems as though it is another name for 'free market'??

Imagine if the big world powers built computers with control over the nuclear arsenals as well as nuclear defense to have superhuman first strike and retaliation capabilities. Imagine if those computers realized that the biggest threat to peace is not the other side but the fact that humans, a warmonger species, control them. Imagine them colluding and trying to establish world control using their nuclear arsenals. Guess the movie title. It's a must watch!


It's a fantastic movie I ought to rewatch one of these days...

Surprised it hasn't been remade. Seems quite appropriate given the fear of machine learning lately.


Sounds like War Games + Terminator + Dr. Strangelove

Dynamic pricing developer here.

The bigger problem than collusion is the arms race between higher margins and higher ad costs. In the end Facebook and Google will take almost all the margin as consumers see prices rise until they receive zero net value from their purchases.

Is there any published evidence that Goggle and Facebrick ads are actually effective? Genuine question.

I don't know anybody who clicks them or takes any notice at all of them but maybe these people exist. I don't know of any company that is wholly reliant on some kind of advertising who uses them. Eg Chanel, Coca Cola etc. I do see a bunch of goog/face advertisements on tv and posters at bus stops where I do also see coke ads.

I don't know of any companies and products who road the early google advert trend to prominence. Do they exist?

My strong suspicion is that the ad industry is drowning in BS and always has been. Data cuts through that so I'd love to see anything that anyone has. Maybe it's not all a giant mountain of con?

> Is there any published evidence that Goggle and Facebrick ads are actually effective?

Effective means ROI positive which is pretty easy to demonstrate.

This is a link to published evidence?

First, consider that no rational person or business would spend money on any marketing that isn't ROI positive (because then they would just be losing money).

Google and businesses aren't amenable to just sharing all of their confidential marketing data, but here's a birds' eye view:


Colour me unimpressed with a marketing document, unaudited that is literally selling the service. Without saying google are or aren't a con or having any view on it in the absence of real data, Bernie Madoff had marketing documents. There's a good reason to want, actual data.

Businesses do irrational stuff literally all the time. There is no business that has ever existed that was always wholly rational.

No data is no data...

Neat work, but doesn't behavior require colluding to be considered collusive? Otherwise this is just another neat equilibrium.

Well, it's not just a legal question. The problem with colluding is that it involves effectively taking money from the non-colluding parties. And this causes those parties to not want to play anymore. If a small or medium investor know the algorithms are going to cheat them, that investor would not want to play the game and this, illegal or legal, prevents investing.

You have to keep in mind that markets are both a "game" and way that capital can be mobilized by society for productive purposes and one doesn't want to kill this latter aspect.

But people will compare prices, and alert the media/friends when they find discrepancies.

I've seen the narrative change in the last few years, Fast food was deemed low cost, but a Calorie Per Dollar comparison showed that eating fast food is significant more expensive.

If a single company doesn't play by the rules, the world will be aware of the price difference.

This assumes competitive markets which allow them to do anything about it. For example (in the U.S. at least), when your local cable company raises prices for Internet services that effectively signals to the local telco that they can too. Sure, the market notices this but has few, if any, other options. (a big assumption is that you even have the option of both cable and telco ISP service in your area.) If all it takes to end up in this situation are NN's learning that implicit collusion is good for profitability, I'd argue that it really isn't a competitive market.

True, heavily regulated government sectors struggle from this(medical for instance), but online is nearly anarchy. I don't see Amazon having ability to screw customers (forever).

And those sectors you are talking about need serious reform that the incumbents DONT want.

Regulated sectors are indicators of markets that aren't competitive since regulation tends to come in after it's well past obvious that markets have failed. When Amazon gets to their end game (which they seem to be pretty close to IMO) I think you'll be surprised at how much and long they'll be able to screw customers... after all, most of their competitors will be long gone and the few that remain will be weak so who's going to be able to offer an alternative in the short to intermediate term?

I'd like to see this done, but over a larger scope where pricing agents are continually improving. I might be willing to believe that non-communicative collusion is possible if both agents are at the same "skill." In the real world, agents will be improved periodically to gain an edge over each other.

"these autonomous pricing algorithms may independently discover that if they are to make the highest possible profit, they should avoid price wars. That is, they may learn to collude"

No, sorry, not what collusion is. It's pretty much business school 101 that price is a terrible way to compete unless you actually have a sustainable comparable advantage on cost or supply -- most people already do this. Collusion involves an actual agreement of some sort. Even if parties are using the same software that operates with the same strategies, that's hardly different from relatively common knowledge about price competition.

edit: I love the concept, sci-fi-wise, though!

Different people can have different versions of the word “collusion” but the important definition is that in the law, and in most countries it does not require an explicit agreement between parties.

I’m not aware of any definition of the word collusion, legal or otherwise, that doesn’t require an agreement between parties. I can’t see how you can have collusion without colluding.

'Tacit collusion' is a real concept; as far as I can tell (as very much a non-expert) it is legal in the US, but there's some ambiguity in the EU.

This (paywalled unless you have academic access) journal article looks very relevant: https://academic.oup.com/jeclap/advance-article-abstract/doi..., and there's a blog post by the author here: https://www.law.ox.ac.uk/business-law-blog/blog/2019/02/tack...

Tacit collusion still requires a tacit agreement. From your linked article:

>In an article published in the Journal of European Competition Law & Practice, I argue that such an approach would amount to a reversal of the burden of proof under existing case law, in particular on the prohibition of coordinated facilitating practices. Indeed, following the current position of the Court, in order to establish the existence of a collusive agreement solely on the basis of circumstantial evidence (such as the uniform adoption of an alleged facilitating practice), and without proof of reciprocal contact through direct or indirect communication, the existence of any plausible and legitimate alternative explanation must be ruled out.

Based on the graph, it looks to me like the agents are just doing some sort of average between the previous prices, and the "ideal" price. This is evidenced by the fact that they both update to the same value immediately post-intervention, and continue to set to about the same values on every subsequent step. The article claims "This kind of self-reactive behaviour is a distinctive sign of genuine collusion," but that seems like a reach.

This is not new, it was already discovered (and also theoretically explained) by Steve Keen & Russel Standish (2006):"Profit Maximization, Industry Structure, and Competition: A critique of neoclassical theory". You don't really need AI to that.

Also, Alan Blinder's book Asking About Prices confirms this (and other things) in practice.

I heard from a friend Amazon dolling out tons of cash to their team of PhDs so they can design automated pricing for certain customers

Second that.

In the end does this matter - my cost function is to get the most possible with the least effort, but there are case statements in the function labelled "do not rob a bank".

If your AI program commits a crime in the pursuit of your goals, both of you are going to jail, just as if your agent / lawyer / employee does.

Side not: Is it illegal for independent contractors to collude on freelance platforms like Fever?

Or could this be considered a trade union?

Things like this are how Soylent Green will happen without anyone actually realizing it.

Applications are open for YC Summer 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact