Computer bot sentences to ponder

…as pricing systems become ever more autonomous, aspiring monopolists like Mr Topkins eventually will not even need to speak to their competitors to fix prices. Computers will do the colluding for them, either by using the same algorithm or learning from their interactions with other machines — all without leaving behind trails of incriminating emails or voicemails.

“Finding ways to prevent collusion between self-learning algorithms might be one of the biggest challenges that competition law enforcers have ever faced,” said a recent paper by the OECD, the Paris-based club of mostly rich nations.

These digital tools automatically calculate prices based on instantaneous assessments of supply and demand and a seller’s own instructions, such as specific profit or price targets.

…It [the OECD] added: “Particularly in the case of artificial intelligence, there is no legal basis to attribute liability to a computer engineer for having programmed a machine that eventually ‘self-learned’ to co-ordinate prices with other machines.”

That is from David J. Lynch at the FT.  Will this prove more or less stable than traditional, human-based collusion?  Here are comments from Henry.  Can the bots send buyers “we are breaking the collusion now” alerts?  Will monitoring third party bots perform that function?  Or will collusion reign supreme?


As a computer programmer with an economics background, I am somewhat skeptical of this claim. In a simple model (typical supply and demand curves, equivalent products, homogeneous consumers, at least two sellers), a seller can always increase profits by defecting from the equilibrium monopoly price, so any algorithm that profit-seeks will not collude. For the monopoly equilibrium to occur, every seller must sacrifice a potential profit and continue to do so in order to maintain the equilibrium, which seems like a hard state to arrive at without every seller's bot having some code tailored to that end instead of profit-seeking.

I have seen bot pricing do stupid things, e.g. two bots both try to have a price slightly above the other, sending the price into the stratosphere. That might look like monopoly pricing to a naive observer, but all it results in for the sellers is zero sales.

As an economist with a computer programmer background, I agree with you.

There might be some really advanced computer sciencey/mathy argument as to a set of conditions whereby a computer will not defect in an AI pricing game, but I can't fathom what it would be.

Having never thought about this prior to this moment, I'll float this:

Airlines used to (or still do?) collude implicitly by raising prices in the middle of the night and seeing whether competitors did the same. If they did, the higher price was left in place. If not, the lower price was reinstated. (This was ruled collusion by the courts, even though it was never explicit.)

So a simple algorithm would be:

1. Periodically raise price at random times with probability P1. If competitors follow within some period of time, keep the higher price; otherwise revert.

2. When you observe competitors raising their prices, match the raise with probability P2.

I would expect there to be multiple equilibria, some of which would result in stable cartels.

I don't think this would work. A lot of people buy tickets or products at night, which means the vendor is going to lose a lot of sales. To the extent that they can ever compete on anything other than price they will also have a negative reputation; I can't fly on any Star Alliance planes since United over charged my parents for a ticket when I was five. If there is more than one firm, there can't be any exclusive IP, and the Chinese can make anything for the same cost as any Internet vendor.

Market segmentation, dude.

Instead of the argument being "this will always happen," try instead "this could easily happen, and be unprosecutable."

If two algorithms seek an equilibrium not of always increasing profit but rather of stable profits, it's not hard to see this happening.

If you set the right parameters for preference for stability, assume price wars are unstable, and restrict it to two algorithms playing against each other who have vast amounts of past data to learn dynamics, then maybe it's possible.

In reality with less data, rapidly changing environments, new algorithms, new competitors, and additional chaos, it's challenging for me to imagine a scenario where two algorithms enter a stable anti-competitive relationship. I doubt reality will be a closed off common-knowledge two-player infinitely repeatable game.

I could imagine bots being more likely to effectively collude if they're set to play a long term game, but i still doubt this will actually happen. The reason is that there will be buying boys, too. Just as buyers keep sellers in check now, mostly, the same should happen with bots. If anything, there will be an incentive to provide bots for small buyers who are currently less sophisticated, so if anything the ability to procure at fair prices may increase. The area is still worth studying, but I don't see the bot equilibrium being much different from the human equilibrium.

But that argument rests on the idea that there are profits to increase in the short run by defecting. If all your competitors instantly notice your rate decrease and, before you conceivably get even one sale (or perhaps more reasonably, before you get substantial numbers of sales) match your price, then when you cut your prices you ONLY cut your own profits.

If your bot is smart enough to understand the landscape, it then does not cut its own price, right?

(There are many plausible exceptions. If your margin is better than your competitors, perhaps you can increase your own profits by reducing your price to a level that they simply can not match. But that's probably only possible if your margins are MUCH better. If you can somehow hide your price from your competitors, then you can beat them -- but how do your hide your price to competitors but show it to customers?)

I've always strongly surmised that FICO scores also enable systematic collusion. For instance, why would repeatedly moving credit balances to a lower rate card be a bad sign for creditworthiness? That indicates a willingness to pay the debt. However, moving balances to a lower rate may be bad for the joint profits of credit card issuers.

What FICO model are you using that says moving credit to lower balance rates lowers your score?

Recently opening a new account can lower your score, but that effect goes away in a month or two.

You can already use google to search a specific item and compare prices at various local stores. Of course, many stores don't have their products easily searchable. But I can forsee an image recognition app, take a picture of a price tag or even the item itself and the app will spit out the prices for competing stores. This will be used as an excuse for stores to set prices to very similar numbers.

The problem, as stated, yields a solution in which gas supply trucks are almost always full when delivering. Supply transaction rates yield the 'no hedge' condition when the delivery rate is minimized, which occurs supply trucks are mostly full. At that point, price is a bell shaped distribution, unchanging.

The price distribution is not known to commuters as stated. So commuters filling their tank will be selecting the quantity as a white noise distribution. However, Gas Buddy offers the same algorithm to commuters as well as gas supply trucks. The commuter technology lets the commuter know relative price at any station, so the commuter fills his tank with an amount adjusted by price.

When this happens then supply drivers and commuters agree, the best bet is to expect a delivery or fill up to about 2/3 of a tank size, that leaves inventory capacity which supports pricing liquidity. .

So the regulator's job is then to protect competitive entry. No good can come from a regulator in the weeds of these algorithms. No regulator will accept either of the two previous sentences.

Isn't this what HFT algos are already doing? You don't even have to program in the behavior, let evolutionary algorithms figure it out.

@davidwright: algorithms with profit maximization utility functions can easily achieve collusion. Any algorithm built by someone who knows game theory will allow for it. However, such an algorithm would need to signal collusion (mainly through pricing patterns) to other participants, and other algos will need to signal back the collusion. So the owner or programmer can be held liable because they built in the signaling and therefore 'intended' to enable collusion.

That said, it might also be possible for algos without any allowance for active collusion to settle into a collusive state. Complex adaptive or autonomous agent models environments have long shown such or similar emergent phenomena. Tit-for-tat algo is also fundamentally a collusive algorithm.

Overall, nothing interesting here. Just stick 'self-learning' into any article these days and everyone thinks its something fresh

here may be a role for middlemen after all.

Unless the government is to take charge of prices, it's hard to see how else the mass of consumers would collaborate to balance against such asymmetric information and information processing speed and quality.

I don't think you can go very far down that road until some middlemen see major profit to compete as an intermediary. But, maybe not.

For example, a Walmart supplier could turn to direct service to consumers. But then if they go very far at all, it's not difficult for Walmart to either buy from them again or source from another supplier.

In principle, most conusmer goods are not that difficult to make (assuming you've got an appropriate amount of money to attract the needed materials, talent, etc.), and there are close substitutes to most of them.

More likely, this will increase the ability of suppliers to offer products which better match the preferences of consumers at a price that they are willing to pay, with price discrimination plausibly offering better prices for price-sensitive consumers and better quality for quality-sensitive consumers.

Obviously, that is not a sure thing. We might end up cows for the milking as well. As a matter of principle, I think in the long run most surplus should accrue to consumers, considering that, for most stuff "you didn't build that". When the opposite logic applies (new stuff), then it will apply ...

But the incentives to break with the cartel remains the same regardless of technology, right? If the price is artificially high, it doesn't take a very smart algorithm to see that you can increase market share by lowering it.

Using price changes as a signal to other people in the cartel only works if there is a cartel. And it's actually worse to do that with an algorithm than with people, because the algorithm's code is discoverable.

With version control, the exact dates of any changes to the code are recorded. Even the stupidest prosecutor in the world is going to figure out what it means when you all come back from the strategy meeting and code up the "raise price by two cents, then wait to see if the others respond" convention in your separate algorithms at the same time.

We are all missing a couple of things.
First, if price information is available only to suppliers, then they get an absolute hedge over consumers. And, as stated, the problem precludes a sub-grouping of suppliers colluding, it is not in the problem. In fact,each supplier only gets a probability distribution that is inaccurate because all suppliers have round robin access to the prices and can change their own prices asynchronously.

Even more importantly, we cannot apply game theory with prices until we include the game that makes pricing work. Just announcing that price, a ratio, can be compared, added and subtracted, implies you have savings and loans up and running in your model. All games with price must include the possibility of buying less and saving more; or buying more and borrowing more.

222@22222222222222222222222222222@@@@@@@@@@@2222@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@22222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnmmmmmmmmmmmmmmmmmmmmmmmmmzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzz9999999999999999999999999999999999999999999999999999999999999999999999999999999oooooooooooooooooooooooooooooooooooooooo8888888888888888888888888888888888888888888888888888888888888888888888877777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777776666666666666666666666666666666666666666666666666666666666666666666666666666666666666666666665555555555555555555555555555555555555555555555555555555555555555555555555333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111```````````````````````````````dedbgtjdeskilejsbvg5rtk4jj45kl5ui4yuh43oi4lnsghewhasbhjt4gyuyheuty7urthiruy3wyrv3uwdQEGFrf53ZSWCVBZAWZESDVGBQASDFGH CTFDYHUFASBHUGYGASVGFSVGTGXXTFSTFVSGDAYRTFQCRVFCRFSWRF5RFEWCDRF

Collusion requires a mechanism to punish defection, as it is the interest of every colluder to defect. What is the postulated punishment mechanism?

Everyone else matches your lower price instantly, so you don't get any increased market share for your price lowering, you only get reduced unit profits.

So perhaps as pricing becomes more program- and algorithm-based, we can make oligopolistic coordination illegal, and require that the programs and algorithms incorporate features that make supracompetitive pricing equilibrium difficult to maintain (by holding that the absence of such features indicates a presumptive intention and invitation to competitors to coordinate prices).

“Particularly in the case of artificial intelligence, there is no legal basis to attribute liability to a computer engineer for having programmed a machine that eventually ‘self-learned’ to co-ordinate prices with other machines."

I believe that depends entirely on the legal standard. Even if we were talking criminal liability, it seems possible that a negligence or recklessness standard could be satisfied: if a reasonable engineer knew or should have known that the machine would eventually self-learn to coordinate pricing with other machines, and didn't take reasonable steps to prevent this from happening, I think that could give rise to liability. But I am about seven years out from Criminal Law at this point.

What would happen to the micro model of supply if dynamic pricing was pushed to the theoretical max?

In other words, what if dynamic pricing made it possible to charge each customer at any given time a unique price equal to the demand curve? All consumer surplus would be eliminated and transferred to supplier(s). I think competition would still exist but it would be on the basis of cost. The lower a supplier's cost, the more profit they could make. The result would be for any given market look at how big a supplier is at full economies of scale. If the market can fit lots of suppliers all at large size, then it will be closer to 'pure competition versus a monopoly model. But in all cases dynamic pricing would mean it would function like a monopoly but worse because you'd never enjoy any consumer surplus.

Since consumer surplus shifts to suppliers, however, society could offset it by increasing taxes on supplier's income. What consumers used to enjoy from surplus could now come from lower taxes and income subsidies.

Might this be more efficient for society as a whole since it will ensure everything that can be created at lowest cost will be?

I used to work with a guy in a quantitative trading group at a hedge fund who got fired for writing an algorithm that ended up trading futures contracts with itself, which is market manipulation, which is illegal. Basically the algorithm had the wrong parameters. If someone implements a machine learning algorithm that they know could potentially "learn"/select/fit such parameters, and they don't do anything to stop it from doing so, then aren't they just as culpable?

Let's say someone wrote an algorithm that was explicitly designed to collude but then let the computer decide whether to truly use the algorithm or not based on the expected profitability of the algorithm in a simulation. So, effectively, all the computer would be doing is making the binary decision to use the algorithm or not; only one parameter, one bit, would be learned/selected/fit by the computer. All of the other parameters were explicitly initialized in advance by the human. Can the person who implemented the algorithm then blame the collusion on the computer while knowing that there was a good chance of the computer choosing the collusion algorithm beforehand? Is this any different from the case of a more complex machine learned model with less explicit initialization? Is the level of explicit initialization really something we want to try to define legally?

Part of what needs to be defined is what it means to learn. Machine "learning" is nothing more than orchestrated parameter selection. Just because we call it "learning" doesn't mean that it is something more mystical that we can't govern, and it certainly doesn't mean that its creators shouldn't be held responsible for it when it goes awry.

Comments for this post are closed