3 ms·

Though a fun paper, Technology Review's write-up of it is worthless. At most, all you should take from reading this paper is that "cartels" potentially can emer

by pash 15y ago

Though a fun paper, Technology Review's write-up of it is worthless. At most, all you should take from reading this paper is that "cartels" potentially can emerge without coordination, not that they do or that they are likely to. Or that this phenomenon has anything at all to do with the dynamics of prices at your neighborhood gas-station. As anyone who has studied complex systems understands, the evolution of interactions among even very simple agents is enormously sensitive to initial conditions. (That's often taken as the very definition of chaos in deterministic systems.) So it is almost always wrong to draw strong conclusions about a complex part of the real world based on your computational model of it: vary one parameter a little bit and you end up with something that looks completely different. This is why agent-based models, like the one employed in this study, are widely despised in economics and the other social sciences despite offering a way to do experiments that would otherwise be impossible in these fields. We simply don't know enough about people and how they interact to build meaningful models, and we probably never will. For example, an economist might point out that the model presented in this paper doesn't account for the spacial distribution of gas stations and consumers, for buyers' and sellers' expectations of future prices, or for any number of conceivable and inconceivable factors—almost all of which would affect aggregate outcomes in the model, I would bet, just as the authors found that the speed of reacting to price-changes did. (That was the only parameter the authors varied, by the way.) And of course we could model the problem in many fundamentally different ways to start with. In other words, this is much different (and harder) than building an N-body model of the dynamics of Alpha Centauri, where's there's basically one force (gravity) that operates in the same way on a bunch of things that can all be described adequately by a single parameter (mass).

eternalban 15y ago
I agree with your general critique, but these guys have a structural problem in their model: Their description of their model[1]: Our model is defined as follows. We consider a population of N agents, where each agent has two simultaneous roles: donator (e.g. a buyer) and rewarder (e.g. a seller). This simply does not correspond to the reality: we have locally interacting agents (gas-station owner, vehicle owner) that exist in a (relatively) 'flat'/'simultaneous' information space (e.g. the news of troubles in oil region x) and the fact that one set of agents (e.g. gas-station owners) are themselves engaged in affinity based relationship[s] with global agents (e.g. oil companies). "No need for conspiracy: Self-organized cartel formation in a modified trust game" -- sure, in a purely locally bounded space of action and information. But that is not the reality of life on this planet. (Not for the past 200 years, for sure.) [1]: http://arxiv.org/pdf/1201.3798v1.pdf http://arxiv.org/pdf/1201.3798v1.pdf
dingfeng_quek 15y ago
Unless we compare simulated models with reality via empirical studies? (: I believe the same issues plague subjects like theoretical physics. Then again, we do have the preference-utility and utility-maximization core of microeconomic theory, which appears quite well-accepted. Accepted, in spite of it's inability to reconcile with some of the non-rational parts of human behaviour. Having said that, I'm quite partial to some of the PDE models (i.e. the works of Dr Yang Xiao Kai). And PDE models have at least one corresponding agent-based model simulation. I think that at some point, a leap of faith based on intuition is taken enmasse. Although this really goes into the philosophy of science...
- pash 15y ago
  Yes, and of course economists do test their models empirically. But many economic models are inherently or practically untestable because the quantities they employ are inherently or practically unmeasurable. And even when you can measure most of what you need, models of complex economic phenomena just haven't worked. The lesson most economists took from the failure of efforts to build useful CGE [1] and DSGE [2] models in the 1960s–1980s was that this approach was a dead-end. Agent-based models, though built on better-founded micro theory, are arguably even harder to get right because agents are adaptive in unexpected ways. And because it's possible to make many models give results that "look right" if you fiddle with the parameters enough, you're just not going to convince anyone that you've built a model that represents the way the world really works. Nobody trusts a black box. I think that's why most economists still build little mathematical models that try to give some targeted insight into some small aspect of some economic phenomenon. If you can understand how a model works (and, again, you usually don't "understand" simulations—all the interesting stuff is emergent), you feel like you've learned something about how the real world might work. Then you can get down to bickering about assumptions, which effects dominate, etc. 1. http://en.wikipedia.org/wiki/Computable_general_equilibrium http://en.wikipedia.org/wiki/Computable_general_equilibrium 2. http://en.wikipedia.org/wiki/Dynamic_stochastic_general_equilibrium http://en.wikipedia.org/wiki/Dynamic_stochastic_general_equi...
  - dingfeng_quek 15y ago
lnguyen 15y ago
That's why this study was done by physicists who tend to not have that particular issue with not completely understanding the system and interactions or reducing things down to "toy" models to as a starting point to build up that understanding. Yes, they are starting with a simple and straightforward model and limiting the parameters they vary. But they can produce what looks to be emergent cartel behavior. That's the important takeaway: a simple model is unexpectedly able to do that. AKA you don't need a complicated model with any number of variable parameters. [Also, what they've done here is study the statistical properties of the system. You're not going to be able to "predict" or replicate the exact behavior/configuration of one particular instance unless you do know its starting parameters in detail. But looking at the statistical properties should smooth that out.]
- pash 15y ago
  > That's why this study was done by physicists who tend to not have that particular issue with not completely understanding the system and interactions ... That's the important takeaway: a simple model is unexpectedly able to do that. Yes, I know. That's why my criticism was directed at the Tech Review article that totally misinterprets the paper, and indirectly at all of the comments in this thread that make the same mistake. The paper does not describe how gas-pump prices change; it does not describe how any actual market actually works. It does not identify time to react to price signals as the key to well-functioning markets; that is an artifact of the highly stylized model employed. As I wrote in the first line of my comment, the take-away is that cartel-like behavior is possible without direct coordination, not that it is actually happening, ever has happened, or ever will happen. The rest of my comment was meant to point out why we can't draw those conclusions. EDIT: And, by the way, there are lots of other ways to show that cartel-like behavior is possible without concerted price-fixing. Several authors have shown this in a game-theoretic framework, and there is a whole literature on indirect communication among economic competitors that has arisen out of law-and-economics analysis of anti-trust regulation.
  - lnguyen 15y ago
    I view the mention by Technology Review and the paper of gas-pump price changes as just giving an example and context that most readers will understand or at least believe they understand. If the assumption is that the cartel-like behavior is coordinated, then that a model that is clearly not coordinated exhibits similar qualitative behavior should give pause to that assumption. I agree that it's not an actual explanation or a model of any actual market. How both the article and paper are worded is a bit ambiguous on that point most likely for purposes of attracting a bit of attention and grant money. But if you look closely, there's no outright claim in either. As for seemingly ignoring existing literature in another field, it's kind of par for the course. A lot of "discoveries" made early on in complexity research on sandpiles were already well-documented in civil engineering.

by pash 15y ago

eternalban 15y ago

I agree with your general critique, but these guys have a structural problem in their model: Their description of their model[1]: Our model is defined as follows. We consider a population of N agents, where each agent has two simultaneous roles: donator (e.g. a buyer) and rewarder (e.g. a seller). This simply does not correspond to the reality: we have locally interacting agents (gas-station owner, vehicle owner) that exist in a (relatively) 'flat'/'simultaneous' information space (e.g. the news of troubles in oil region x) and the fact that one set of agents (e.g. gas-station owners) are themselves engaged in affinity based relationship[s] with global agents (e.g. oil companies). "No need for conspiracy: Self-organized cartel formation in a modified trust game" -- sure, in a purely locally bounded space of action and information. But that is not the reality of life on this planet. (Not for the past 200 years, for sure.) [1]: http://arxiv.org/pdf/1201.3798v1.pdf http://arxiv.org/pdf/1201.3798v1.pdf

dingfeng_quek 15y ago

Unless we compare simulated models with reality via empirical studies? (: I believe the same issues plague subjects like theoretical physics. Then again, we do have the preference-utility and utility-maximization core of microeconomic theory, which appears quite well-accepted. Accepted, in spite of it's inability to reconcile with some of the non-rational parts of human behaviour. Having said that, I'm quite partial to some of the PDE models (i.e. the works of Dr Yang Xiao Kai). And PDE models have at least one corresponding agent-based model simulation. I think that at some point, a leap of faith based on intuition is taken enmasse. Although this really goes into the philosophy of science...

pash 15y ago
Yes, and of course economists do test their models empirically. But many economic models are inherently or practically untestable because the quantities they employ are inherently or practically unmeasurable. And even when you can measure most of what you need, models of complex economic phenomena just haven't worked. The lesson most economists took from the failure of efforts to build useful CGE [1] and DSGE [2] models in the 1960s–1980s was that this approach was a dead-end. Agent-based models, though built on better-founded micro theory, are arguably even harder to get right because agents are adaptive in unexpected ways. And because it's possible to make many models give results that "look right" if you fiddle with the parameters enough, you're just not going to convince anyone that you've built a model that represents the way the world really works. Nobody trusts a black box. I think that's why most economists still build little mathematical models that try to give some targeted insight into some small aspect of some economic phenomenon. If you can understand how a model works (and, again, you usually don't "understand" simulations—all the interesting stuff is emergent), you feel like you've learned something about how the real world might work. Then you can get down to bickering about assumptions, which effects dominate, etc. 1. http://en.wikipedia.org/wiki/Computable_general_equilibrium http://en.wikipedia.org/wiki/Computable_general_equilibrium 2. http://en.wikipedia.org/wiki/Dynamic_stochastic_general_equilibrium http://en.wikipedia.org/wiki/Dynamic_stochastic_general_equi...
- dingfeng_quek 15y ago

lnguyen 15y ago

That's why this study was done by physicists who tend to not have that particular issue with not completely understanding the system and interactions or reducing things down to "toy" models to as a starting point to build up that understanding. Yes, they are starting with a simple and straightforward model and limiting the parameters they vary. But they can produce what looks to be emergent cartel behavior. That's the important takeaway: a simple model is unexpectedly able to do that. AKA you don't need a complicated model with any number of variable parameters. [Also, what they've done here is study the statistical properties of the system. You're not going to be able to "predict" or replicate the exact behavior/configuration of one particular instance unless you do know its starting parameters in detail. But looking at the statistical properties should smooth that out.]

pash 15y ago
> That's why this study was done by physicists who tend to not have that particular issue with not completely understanding the system and interactions ... That's the important takeaway: a simple model is unexpectedly able to do that. Yes, I know. That's why my criticism was directed at the Tech Review article that totally misinterprets the paper, and indirectly at all of the comments in this thread that make the same mistake. The paper does not describe how gas-pump prices change; it does not describe how any actual market actually works. It does not identify time to react to price signals as the key to well-functioning markets; that is an artifact of the highly stylized model employed. As I wrote in the first line of my comment, the take-away is that cartel-like behavior is possible without direct coordination, not that it is actually happening, ever has happened, or ever will happen. The rest of my comment was meant to point out why we can't draw those conclusions. EDIT: And, by the way, there are lots of other ways to show that cartel-like behavior is possible without concerted price-fixing. Several authors have shown this in a game-theoretic framework, and there is a whole literature on indirect communication among economic competitors that has arisen out of law-and-economics analysis of anti-trust regulation.
- lnguyen 15y ago
  I view the mention by Technology Review and the paper of gas-pump price changes as just giving an example and context that most readers will understand or at least believe they understand. If the assumption is that the cartel-like behavior is coordinated, then that a model that is clearly not coordinated exhibits similar qualitative behavior should give pause to that assumption. I agree that it's not an actual explanation or a model of any actual market. How both the article and paper are worded is a bit ambiguous on that point most likely for purposes of attracting a bit of attention and grant money. But if you look closely, there's no outright claim in either. As for seemingly ignoring existing literature in another field, it's kind of par for the course. A lot of "discoveries" made early on in complexity research on sandpiles were already well-documented in civil engineering.