An Evolutionary Theory of Economic Change by Richard R. Nelson and Sidney G. Winter
There’s a somewhat straw-man view of economics where we take all the mathematical assumptions at face value. We imagine, for instance, Apple designing the next iPhone. First, they take all of their collected data and build detailed probabilistic models of their customers. Then they consider every possible iPhone design, plug each one into the customer-model, and simulate customers choosing whether to buy the phone or whether to make purchases on it. The company calculates their expected profits and risk, plug those into a model of their longer-term finances, and finally choose the design with maximum long-term expected profit. This is how idealized companies behave in economic models: they calculate a probabilistic model based on all the information available, then choose the option with maximum expected long-run profit. We’ll call this the “idealized profit-maximizer” model.
It has been observed, from time to time, that actual companies don’t really seem to behave this way.
What actually happens is that Apple takes a bunch of noisy data, builds some crappy models that everybody is correctly skeptical about, then some executives eyeball a few graphs and argue a lot and make a largely-random decision about the final design.
One of the economists’ main responses to this observation is a selection argument. In the context of companies maximizing profit, the argument says that companies whose behavior just so happens to maximize profit will end up dominating the industry. They’ll have more money to expand faster, invest more, advertise more. It’s not that any particular company is a perfectly rational idealized profit-maximizer, it’s that the dominant companies act like profit maximizers would, because that’s what the market has selected for. It’s like fitness-maximization in biology: bees and mushrooms are not perfectly rational fitness maximizers, but they act like fitness-maximizers would, because that’s what evolution has selected for.
It’s not that Apple figured out which design will maximize profits. It’s that Apple stumbled on a design which just-so-happened to roughly maximize profits, and as a result they expanded rapidly and that profit-maximizing design took over the industry.
We’ll call this picture the “evolution” model, in contrast to the pure “idealized profit-maximizer” view. The key hypothesis of the model is that evolutionary pressure produces (approximate) profit-maximizing behavior, even when companies themselves are not idealized profit-maximizers.
The central question of this review is: does this evolutionary hypothesis actually hold? What are the hidden assumptions, the approximations, the edge cases? When, and to what extent, will market selection pressure actually result in an industry which behaves like profit-maximizers? How well does the evolution picture hold up when we actually do the math?
Based on Nelson and Winter’s 1982 book An Evolutionary Theory of Economic Change, when we do the math the evolution picture holds up remarkably well. An evolutionary model produces (approximate, local) profit-maximizing behavior surprisingly reliably, under fairly weak assumptions. Nelson and Winter demonstrate this via both proofs (under at-least-sometimes-realistic assumptions) and simulations (under models intended to capture more real-world complexity).
We’ll briefly talk about their results - both their main proof and some simulation results - then zoom back out and talk about the broader approach and worldview behind those results.
On the “proofs” side, Nelson and Winter’s core argument goes like this: * Businesses change their behavior if-and-only-if profits are low (relative to the rest of the industry) * Profit-maximizing behavior will never have low profits (relative to the rest of the industry) * Therefore, any company which happens to try whatever actions are profit maximizing will continue to perform those actions indefinitely: they’ll never have low profits, so they’ll never change. * Conversely, any company with less-than-average profits will try new things until their profits are no longer low. This will push average profits up over time, until the industry converges to profit-maximizing behavior. Some of this sounds unrealistic, but as with any proof we can add more realism with a few simple patches, at the cost of more complexity. The main part which can’t really be patched is the idea that a successful company will never abandon their successful strategy. That strains credulity, and relaxing the assumption to something like “businesses mostly don’t change behavior unless profits are low” completely breaks Nelson and Winter’s proof. That said, just because a more realistic assumption wrecks their proof doesn’t mean that businesses won’t still end up as profit-maximizers, it just means more complicated math… which is where the simulations come in.
For people skeptical of the relevance of proofs to complicated real-world systems, the simulations are the real evidence in this book, and they are the bulk of the technical results presented. Nelson and Winter simulate companies with many production techniques, various policies on which technique to use, various research and reinvestment rules, etc, with parameters chosen to roughly match reality. All of the simulated companies follow randomly-chosen rules, without any explicit profit maximization. Do the companies which happen to adopt profit-maximizing rules end up dominating the industry? Does the industry overall end up behaving as though companies are profit-maximizers? In chapter 9, Nelson and Winter fit an idealized profit-maximizer model to the simulation results to find out. Their answer: “by and large it seems that [an economist testing the idealized profit-maximizer model] would believe that his theory had performed well”.
That’s a rather spectacular understatement - they fit an idealized profit-maximizing-firm model to the evolutionary simulation data and find an R^2 value of 0.99. In other words, companies subject to selection pressure approximate profit maximization ridiculously well; evolution does indeed induce near-ideal profit-maximizing behavior. The evolutionary hypothesis holds extremely well.
But is this result really that impressive? Presumably Nelson and Winter set out to build a foundation for orthodox profit-maximizer-style economic models; of course they would choose parameters for which the evolution-induces-maximization model works well.
Except that Nelson and Winter weren’t trying to build a foundation for profit-maximizer models. What makes the results really believable is that they were quite explicitly trying to attack those models.
Habit and Routine, Not Calculation and Optimization
Nelson and Winter have a vision of how economic theory ought to look. It’s a vision in a similar vein to James Scott (Seeing Like a State, Against The Grain), Michael Polanyi (The Great Transformation), or Joseph Schumpeter. Knowledge is not explicit, it’s tacit. Local and contextual, not global and general. Habits, not calculation. Sometimes the steel mill’s process only works because of an impurity in their ore which neither the mill nor the supplier knows about. One worker always leaves tool A in bin B, another worker always looks for tool A in bin B, and nobody ever thought about it because it just works. Maybe someday a new worker will come along and leave tool A in bin C, and other workers will be confused for a bit until they settle into the new normal. People mostly keep doing what works, and if anything goes wrong they’ll look around for something else which works and then do that instead.
Businesses, in this vision, are the same thing at a larger scale. They hire people, adjust processes, introduce new products or pull old products, generally try stuff out and see what works. They don’t know that the sudden uptick in sales of an old sunblock product is due to the opening of a new waterpark, they just know that sales are suddenly up so they better make more of it. If the waterpark shuts down for some reason, the company won’t know that they need to cut back production until a backlog of sunblock builds up in the warehouse. The business doesn’t really know what aspects of the environment are crucial to their profitability, which employees are performing crucial tasks that nobody else thinks about, which accidental features of their products are crucial to downstream consumers.
In a static world, it wouldn’t matter, at least for predicting the behavior of businesses. The results from the previous section tell us that businesses will randomly try things until they hit on (approximately) optimal choices, and then mostly keep doing that. They end up acting like profit-maximizers, regardless of their actual decision processes.
But this breaks down when things change quickly.
Take a worker from one company, hire them to do nominally the same job at a different company, and they’ll spend a month or two largely confused and ineffective while they figure out where things are, which people they need to talk to when, and all those other little things that don’t directly translate from one company to another. Take a moderately successful thirty-year-old family restaurant and throw it into the environment of COVID and delivery, and the restaurant will thrash around desperately trying different menus, prices, hours, delivery and payment options until they either find a combination which works or shut down. There is no methodical enumeration of all available options, no Bayesian update. In retrospect, the restaurateurs may realize that they had all the information needed to identify the best option - but hindsight is 20/20.
In other words, when things change quickly, the evolutionary viewpoint says that businesses’ behavior should be drastically suboptimal for some time. Not indefinitely; over time those businesses will figure out what works (or go out of business and be replaced by others). But we shouldn’t expect profit-maximizing behavior on short timescales in response to major changes.
Thus the need for an evolutionary theory of economic change, specifically. Idealized profit-maximization models, in which we pretend that companies make the best-possible decision every time, might be able to predict a long-run equilibrium. But they can’t properly predict the path to that equilibrium, because in the real world that path involves companies making less-than-ideal decisions and learning from their mistakes. And in a sufficiently rapidly-changing world, we may never get close to the equilibrium anyway; the path is all there is.
Nelson and Winter expound on such shortcomings of “orthodox” (i.e. idealized profit-maximizer) economic models for about half of the book.
Economics Has Come A Long Way
Unfortunately, while they talk at length about the inability of idealized profit-maximizer models to handle high uncertainty, fast change, tacit contextual knowledge, and the like, Nelson and Winter’s math doesn’t really back it up.
This is probably more clear in hindsight than it was in 1982. When Nelson and Winter talk about things which are “fundamentally incompatible” with idealized profit-maximizer models, a fairly large chunk of the commentary is about nonequilibrium dynamics (i.e. fast change) and high uncertainty. The tools to handle those (mainly dynamic programming, Bayesian probability, and game theory) were around in the ‘80’s, but they hadn’t been integrated into the “core” economic models to nearly the extent that they are today. For instance, the more recent models in Ljungqvist and Sargent’s Recursive Macroeconomic Theory are very clearly in the same idealized profit-maximizer family which Nelson and Winter criticize, yet they directly handle dynamics and uncertainty. The core models of economics just aren’t dependent on static environments and perfect information to the extent that they used to be; those shortcomings turned out to be largely independent of the idealized profit-maximization assumption.
Conversely, the evolutionary models which Nelson and Winter themselves present mostly seem to provide a solid foundation for the idealized profit-maximization viewpoint: despite involving plenty of dynamics and uncertainty, their results mostly look like the simulated companies are quite close to idealized profit-maximizing behavior most of the time. Much as they want to overthrow the idealized profit-maximizer models, they don’t really show evolutionary models behaving dramatically differently from how idealized profit-maximizer models would behave. They expend a lot of ink arguing about how evolutionary models are superior to idealized profit-maximizer models and should largely replace them, but their actual results don’t really back it up.
That said, to a large extent I think this is a failure of Nelson and Winter rather than a failure of the evolutionary approach. Just as orthodox economic theory has come a long way over forty years, evolutionary economics has also come a long way - a two-minute perusal of this 2004 paper, for instance, reveals a much more substantive mathematical foundation than Nelson and Winter used, even in a paper which isn’t really doing much sophisticated modelling. That mathematical foundation matters if we want to do things like, say, back-of-the-envelope estimates to figure out whether and when the predictions of an evolutionary model will diverge from the corresponding idealized profit-maximizer model.
I very much doubt that Nelson and Winter’s commentary was entirely wrong; we should expect the predictions of idealized profit-maximizer models to diverge from reality under conditions of rapid change. But without a sufficiently solid mathematical foundation, we have no idea how fast things need to change, or which things, in order to see novel testable predictions from evolutionary theory. What precisely are the conditions under which the idealized profit-maximization approximation breaks down? That’s the sort of foundation I was hoping to see in Nelson and Winter, and which I still hope to see - but I’m not sure where to look.