What drives economic development? Or more precisely, what constrains economic development? An emerging consensus on this question surrounds the role of locally embedded productive capabilities and the idea that countries build on their existing capabilities to move into new economic activities. In new research published in Nature Communications, Neave O'Clery, Muhammed Yildirim and Ricardo Hausmann develop a mathematical model based on capability accumulation of countries and use this model to construct a directed network of products, the Eco Space. They uncover a modular structure in the network and show that low- and middle-income countries move from product clusters dominated by few capability products to advanced (many capability) products over time. They also show that the network model is predictive of product appearances in countries over time. In this Growth Lab podcast, Research Analyst Ana Grisanti interviews the authors of this new research Neave O'Clery, Muhammed Yildirim, and Ricardo Hausmann, to learn more about their findings.
Read the Productive Ecosystems and the arrow of development paper, published in Nature Communications.
Learn more about the authors Neave O'Clery, Muhammed Yildirim, and Ricardo Hausmann.
Explore the Product Space via the Atlas of Economic Complexity.
Katya Gonzalez-Willette Hello and welcome to another episode of the Growth Lab podcast series.
Katya Gonzalez-Willette What drives economic development? Or, more precisely, what constrains economic development? An emerging consensus on this question surrounds the role of locally embedded productive capabilities and the idea that countries build on their existing capabilities to move into new economic activities. In new research, published in Nature Communications, Dr's Neave O'Clery, Muhammed Yildirim and Ricardo Hausmann develop a mathematical model based on capability accumulation of countries and use this model to construct a directed network of products called the eco-space. They uncover a modular structure in the network and show that low and middle-income countries move from product clusters dominated by few capability products to more advanced products over time. They also show that the network model is predictive of product appearances in countries over time. In this episode of the Growth Lab Podcast series, Research Analyst Ana Grisanti interviews the authors Neave O'Cleary, Muhammed Ali Yildirim, and Ricardo Hausmann to better understand their new research on productive ecosystems and the Arrow of Development.
Ana Grisanti Thank you all for being here. We're here to talk about the Productive Ecosystems and the Arrow of Development paper that you just published. So I wanted us to start with what is the framework of your paper and what is the research question to be answered here? How did this research question come about and what discussions based? Is it relevant?
Neave O’Clery So I guess I'll start with how the paper came about. I arrived at the Growth Lab in 2013 as a former Ph.D. student in network science, and I knew very little about the field of economic complexity. And one of the key tenets of the economic complexity theory, which I'm sure Ricardo will elaborate on in future questions, is the idea that countries are endowed with capabilities and it's the presence of these capabilities that constrains their ability to grow right. There's some kind of directed process by which countries accumulate capabilities, and it enables them to move into new economic areas, new products, new industries, etc.. And so when I started to think about this from a network perspective, it seemed to me that this was very much a directed process, right. That there was a process by which countries accumulated capabilities, but that the modeling tools that were currently being used were not necessarily directed. In essence, a specific network approaches, right. They were using a static, whereas a cross-sectional network which didn't describe this process of capability accumulation over time. So the real motivation for the paper and for the research was to better capture using a model and using data this directed process of capability accumulation, which if you were to explain how a development work through capability accumulation, that was very much present in the literature at that time, but the modeling techniques didn't quite capture that process.
Ana Grisanti Great. Thank you.
Muhammed Ali Yildirim The area that you're thinking about, product space, which is a cross-sectional look at the capability correlation, I would say, but we're also thinking about this dynamic process, which products should be there to grow other products because it how we see capabilities are through the products that we have. So that was the basic idea behind the paper by looking at what products are often there. When we see a jump of a country to a new product, we can identify this ecosystem. That's what we call this space instead of a product space, and to mathematically show that this ecosystem captures the capability of overlap between the products. So we can say that this product has that much overlap with other products. And this leads to for us to understand this arrow of development in terms of capability accumulation.
Ricardo Hausmann Yes, I like both descriptions. To me, the idea of arrow of development implies that, you know, you're more likely to first make a car and then make an airplane. It's very unlikely that we will start making airplanes and then move to making cars. The idea being that an airplane is something harder to make, requires more things so that we are more likely to start with the simpler things and then move to more complicated things. So first you have to learn from something that's happening over time. But I think in this paper there's that idea and there's a different idea which says that instead of asking yourself which other product is there when this product is just product or product, it's really a relationship of product and all the other products that tend to be already present in this product. So it's a relationship between a product and a basket of products. And on the whole, that's an appliance in the business because we talk about business ecosystems. Right. That's the typical Karlan's in business groups. Economists don't like that too much, but it is sort of like what are all the other species that are around when this species arises? And I think that idea is a core idea in this group. And the second idea is the arrow of time. Some things happen before others.
Ana Grisanti Great. Thank you. Can you explain the way in which you redefined capability and relationships between products in this paper?
Muhammed Ali Yildirim So what we do is basically the capabilities are somewhat elusive. We know that that we can taste and we can smell them, but we cannot really identify all the capabilities extensively. So this is the nature of capabilities. So you need to look at how you can quantify the relationships between these capabilities. So in our work, when we write this capability, in terms of mathematical models and relate the probability of jump to the capability overlaps between products, we think that we can identify the extent of the length of the capability because we don't still identify what the capabilities are exactly, but we kind of capture the extent of the capabilities required for a product by looking at this arrow of development and by capturing this extensive margin of the capabilities, we can say that these products are more complex. So it arrives at a different definition of complexity because if you require a lot more capabilities, it means that this product is very sophisticated and it becomes a kind of ladder of development. So that's the new way that we are measuring capabilities here.
Neave O’Clery So I think we have two components to the paper, right? We have a model which as Muhammed explained if I can write a product in a country that are endowed with capabilities and we have a model that tries to capture the gap in capabilities between what a country has, and what a produce requires, and what we do as we relate this model of capabilities to an expression for product presences and appearances, right. And that's something that we can empirically measure and data. So we use international trade data to estimate the capability overlap between products. So we take this quite similar approach to the product space in the sense that we use international trade data to infer the capability overlap between products. But what is quite different is the way in which we actually calculate that overlap, because what we do is we use the time dimension in this data, the sequence of products presences, products appearance's for countries, and we sort of mine a much greater amount of data because we use it all the way back to the 80s in order to try to infer these capability overlaps to create instead of a cross sectional, undirected network, we create a direct network. So that's one of the key differences between what we do in this paper to construct the network and what was done in previous work.
Ana Grisanti Great. Thank you. And this question kind of stems from what you were just talking about, Neave. What contribution does this method make to the theory of how countries move forward and diversification? And how does the ecosystem space add to what is learned from the product space?
Muhammed Ali Yildirim So I think what we see is the title of the paper says, the area of development. So we see countries jump from some type of products to others. And often this is from that low complexity to high complexity products, for instance, what we observe in the data. So one of our findings in the paper, we show that this relationship goes from low complexity, low PCI, that from the Economic Complexity Network to the high complexity products in terms of diversification in the future. One might ask whether the countries that achieve these terms today have different type of properties and other things that facilitate these jumps. It's an open question, but in terms of diversification, what we observe is it's often the case that we see this pattern of diversification repeatedly and on top. It's generally from the store complex. It's of a complexity nature.
Ricardo Hausmann So essentially, I would say that what you're doing as Neave was saying is you look at a country's production of all products and you're learning from the whole history of diversification of all the countries in the world over all the previous years to infer what is likely to be a jump from this country, given all the things that it's is currently making. And the model allows you to make that prediction and sometimes to choose to identify which are the most likely next steps for this country.
Muhammed Ali Yildirim I would add in the product space, think of you look at all the products the country makes and how close they are to this particular product divided by how close all the products are to this particular product. So it looks at the monkey and the forest analogy that Ricardo devised. It looks at how many trees that you have monkeys on and how close those monkeys are to the tree that you want to jump to here. What we do is we basically come up with a different type of density measure that is a direct result of the mathematical model that we developed.
Ana Grisanti So what are some of the advantages of calculating complexity and density in this way rather than the way the product space does it?
Ricardo Hausmann Well, the paper shows that you get more predictive power.
Muhammed Ali Yildirim And it's also dynamic. That's the direction to the relationship in the product space. The distance between products are symmetric. Here, you can make it asymmetric. So it's a choice to minimize the noise that we do in the product space. But what we do is we come up with something that has an arrow that goes from product A to product B, and it's generally if the arrow from product A to product B is really straight in one direction it is not as informative in terms of jumps.
Neave O’Clery As we were saying before, because we include so much more data from the past. It's unsurprising in many ways that we get a better predictive power because we are learning from what has happened in the past and that can be quite a powerful addition to the model.
Ana Grisanti Great, thank you. So let's dig a little deeper into your results. Can you talk about how you created the product communities and how they differ in terms of the ecosystem size and ecosystem input? And what are some of what you call stepping stone communities?
Neave O’Clery So for any network, we can first ask ourselves, what is the community? So in a network of communities, a partition of the nodes into groups, and those groups are typically characterized by high internal connectivity. Right. So you have groups of those that are highly connected, but those groups are less connected, in a sense, externally or to other groups. And so when we think about our network, a group of nodes is a group of products and they are connected by some kind of high degree of capability overlap. So these are groups of products that acquire similar capabilities. So when we apply an algorithm which is based on a sort of random walk around a network, so this random walk around the network detects areas of the network with high density of edgeways, we find I think it's five communities and we study the characteristics of the nodes, the products in these different communities. And we have some communities that are full of things like food products and low complexity type activities, things that very much developing countries might be active in. And then we have other communities that are full of sophisticated, complex products such as pharmaceuticals and electronic manufacturing, etc. So we can characterize these communities based on the size of the ecosystem. This is going back to what Muhammed was saying about being able to characterize the vector of capabilities. In a sense, for a product the size of an ecosystem of a product is really how many other products tend to be present before this product appears. A product with a large ecosystem size tends to require many products that were present in the past and is a complex, in a sense, in the lingo, a product that has a high ecosystem input score tends to be part of the ecosystem of many other products. So it tends to be almost a springboard products that you might start from to move into other products in the future. So those would be the low complexity, things like food and sort of first games and things like that. So we have the different communities we characterize in terms of we see quite some distinction between communities in the paper. For example, what we term the yellow community is full of these low complexity products, which are mainly ecosystem input products, and then the blue community, which is very complex and has manufacturing in pharmaceuticals. This is characterized by-products with a large ecosystem size. So they require many other existing capabilities and products to be present before those products tend to emerge in countries.
Ana Grisanti We also would like to know a bit about how ecosystem size is related to wealth of countries.
Muhammed Ali Yildirim So while we are building this ecosystem measure or other measures all the information that we used is which countries export which products. So we don't put any price information or any other measure that can capture the wealth of the nations. But surprisingly, if you look at the mean ecosystem size of a country's products and it's gross GDP per capita, we see a super high correlation. So because the large mean ecosystem size means that these products are requiring a lot of capabilities and countries making these products are going to be making in many other products, like the economic complexity measures that we do. And not surprisingly, we see that kind of relationship. So the ecosystem size of a country is highly correlated with GDP per capita of the country. On the other hand, we have this system input capacity of a product. So the product that's more fundamental can lead to many other products. So the product could be a source to many other products. And when you see that look at the mean ecosystem input size, we see a negative correlation instead of a positive correlation. So it means the countries that have only fundamental products, those really unsophisticated products, they can go to many places they can. We can devise them many different parts. But its current status if. You have predominantly products that are inputs to many other products, then generally you have less GDP per capita.
Neave O’Clery Yeah, so, I mean, the next thing we did was we used the aforementioned communities to characterize, in a sense, the development path, the arrow of development of different countries. Right. So we looked at how countries change their share of products in each of these different communities. So we thought up, for example, we could see this kind of clear trend that countries over time would move out of this sort of low complexity community and they would move into the blue and purple communities which were characterized by these large ecosystem science complex products. And so we actually tracked for many countries. At which point did they kind of transition from having more products in yellow to having more products in blue? And we saw some really interesting patterns. We saw countries like Singapore transition in the 80s and we had a raft of middle-income countries transition around the US and Malaysia, Mexico, China transition and much more recently. And India is still on its way to transitioning. So it was kind of an interesting way of capturing and visualizing, in a sense, the development of the kind of export baskets of countries over time in the network.
Muhammed Ali Yildirim And these countries are super stable. Right. So it seems that they haven't diversified like they have in their direction towards higher complexity products or high input ecosystem sized products. So it means that these countries like to transition or get out of the poverty trap they need to transition, but they haven't done that yet.
Ana Grisanti That's really cool how you can see the history of countries there, like what they're exporting and how they're moving from less developed to more developed Ricardo would you like to add anything?
Ricardo Hausmann In another metaphor we like to use is that products are like words that are made by putting letters together, and if you want to know if a country can make a product, you are sort of asking the question, do they have the letters? Since we cannot observe the letters, but we actually observe is all the other products that would require those letters. And since every product requires a collection of letters, maybe those are going to be present in some of the products that you're making in different products are made. So a very long work is going to require many letters. Those are going to be expressed in many of the other products that a country makes. So that's this ecosystem of that product. It's all the other products that use more or less the same letters. And because those products already preexist, it means those letters are there. It means that this new product can appear. And so rich countries would be countries that have a lot of letters. Consequently, they make many products. And among them they make products that require a lot of letters. They make long words. And so these products at the end of the process are these very long words. They have these big ecosystem requirements and poor countries start with these short words. So they don't have big ecosystem requirements because it's just a short word, very few letters are that you need to get into those industries. And the whole challenge is to add letters in a way that can be transformed into more words and longer words. That process, if you send them to a very poor country or nuclear engineer, most likely that nuclear engineer is not going to be very useful because he needs a whole lot of things to be there for him to do his thing. These ecosystem requirements tell us, is this transition likely to happen, given that this new product that you're trying to make requires all of these letters? But if you have them, then all of these things should be also present and they're not. So maybe that transitions happen. And the fact that transitions that we say are likely still don't happen means that this process of capability accumulation must be challenging.
Ana Grisanti Thank you. So what are some of the limitations, if any, of this model?
Ricardo Hausmann I think that the fundamental limitation that we're struggling with is that we wish we knew what these capabilities were. We wish we could observe them, that we could identify them explicitly. We this metaphorically when the difference between the genotype and the phenotype, genetics started by just looking at the properties of beans or rain. And Mendel had no idea about where these properties came from. But he found some relationship between what happens if you make seeds from a tree with beans of this shape, with sex organs of trees that create beans of this other kind. What happens when you mix them and you put infer the properties of these mixes just from the beans because Mendel had no idea of the genes. He did not know about DNA. he did not know about these things. So he could only measure things at the level of the species and not at the level of the genetic code. So we are a little bit trusting that we can see very easily the products. We have some very broad categories of the requirements of those products. We can look at the labor inputs that are used in making the products according to some classification. We can look at the input-output characteristics of this product, but many of these things that a product requires are things that, you know, you don't have them. It doesn't matter because you can just import them. So which ones are the ones that really are critical that they have to be in place for this product to appear? That's something that we still don't know and that we are making progress in different papers on that question. There is a current paper by Dario Diodato and Ulrich Schetter which they are working on trying to say, well, let's assume that these inputs are really occupations or the occupations that have to be present and are those occupations present in this country and try to predict from that? You don't get as good predictions from a purely predictive point of view as this paper, because you're not trying to maximize the predictive capacity of trying to maximize understanding of what may happen or what may be the mechanism. And so I think that that's a little bit this research has to go.
Neave O’Clery If we look at the limitations of this specific model in terms of if you were to try to replicate it or apply it in your own context, you mentioned a couple of limitations in the paper. The predictive power is good, but we still have quite a lot of false positives. So things that our density metrics that should appear that don't. Right. So this is of course, there's a lot of things going on in the economy. There are a lot of things we don't observe in this data. And so certainly that's something to keep in mind when policymakers will be looking at a granular level. And so one of the things that we suggest is that this could be quite interesting way of identifying market failures, for example, trying to understand why things that seemingly have the perfect ecosystem in place don't end up appearing at all. And the other limitation that we mention in the paper is that, of course, the technological requirements of products and policy overlaps evolve over time, not very fast, but they still evolve. And so when one is looking to the future and thinking, what could you use this type of metric for when you're predicting we suggest that a sort of five-year time frame is probably roughly appropriate, that if you were looking further into the future, you possibly could neglect this change in the network structure that would occur over that time, even though we do find that it is really quite stable over the time period that we.
Ana Grisanti Great, thank you. Let's end with how is the ecosystem electric helpful for policymakers? What would be the next steps for policymakers to use this metric in a practical way?
Ricardo Hausmann I would guess that this paper will be of interest to anybody who's interested currently in the atlas of economic complexity in the product space because it is an improvement figuring out which things are likely to happen in your country if you push a little bit if you figure out these market failures that he was talking about. So that includes, you know, people who are in the interest of promoting investment, whether it's at the national level, at the state level, at the city level. It also is useful for anybody who's planning to invest in a particular industry in a particular place. It helps, you know, if the ecosystem that is in place is appropriate for the appearance of this product. So I think that's the activity activities. So investment promoters, investors, firms, et cetera, are trying to make allocation decisions. Those are the two that come to mind.
Muhammed Ali Yildirim In the paper we mentioned what Neave said about identifying market failure because I assume the product that should be there is not appearing remeasure indicates that with a high likelihood that product should be that gives you this identity of the products that you can go. And we also identify which products are the closest ones, have the closest capability over next. And you can go to that industry and ask what limits the jumped to this other product. So it provides you with a concrete tool to go and survey the business people and so forth in terms of identifying the market leaders and policymakers could use different policy tools such as those papers and also in the paper to say that this application of the idea of this ecosystem or the capability overlap is not just limited to products and countries. That's why we published this paper with general interest scientific journal Nature Communications, because we think that this idea could be brought to the ecology methodology and some other literature. And on top, we think this idea of genotype versus phenotype, as Ricardo alluded to, is an important concept and we are approaching different problems in many different fields.
Neave O’Clery I would echo those sentiments. I think the path forward for us, in a sense, is to think about many of the product space metrics have been implemented on a regional and urban level using other types of data. So export data, when we look at countries primarily because it's a very clean, standardized, reliable data set, when you go to the national level, you have use of all sorts of different types of data sets that capture a lot more activity in an economy. So employment in industries which captures domestic activity as well as exporting activity. So I think there's a lot of scope to develop this model in that direction and try to create similar metrics that can be used for urban development experts.
Ana Grisanti So thank you all for being here, so this paper is published in Nature Communications and the link to the paper can be found in the podcast notes.
Katya Gonzalez-Willette If you want to learn more about the Growth Lab's latest research and events, please visit www.growthlab.cid.harvard.edu.