What is the Price of a Skill? The Value of Complementarity

The global workforce is urged to constantly reskill, as technological change favours particular new skills while making others redundant. But which skills are a good investment for workers and firms? As skills are seldomly applied in isolation, we propose that complementarity strongly determines a skill's economic value. For 962 skills, we demonstrate that their value is strongly determined by complementarity - that is, how many different skills, ideally of high value, a competency can be combined with. We show that the value of a skill is relative, as it depends on the skill background of the worker. For most skills, their value is highest when used in combination with skills of a different type. We put our model to the test with a set of skills related to Artificial Intelligence (AI). We find that AI skills are particularly valuable - increasing worker wages by 21% on average - because of their strong complementarities and their rising demand in recent years. The model and metrics of our work can inform the policy and practice of digital re-skilling to reduce labour market mismatches. In cooperation with data and education providers, researchers and policy makers should consider using this blueprint to provide learners with personalised skill recommendations that complement their existing capacities and fit their occupational background.

1 Introduction Technological change is not "skill-neutral".Innovations create new jobs but these jobs are characterised by new tasks that require new sets of skills.As a result, the landscape of skills and occupations changes.Despite recurring fears of mass unemployment as a result of automation, the current literature suggests that firms are using technologies to automate specific tasks rather than entire occupations (Autor, 2015).While some occupations will indeed disappear, the remaining ones will change, and entirely new jobs will emerge (Acemoglu & Autor, 2011;Brynjolfsson & Mitchell, 2017;Frey & Osborne, 2017).Logically, the work that is eliminated has different skill requirements than the newly created jobs, resulting in the paradoxical situation of simultaneous unemployment and labour shortage (Autor, 2015).In other words, workers risk being pushed out of employment while companies struggle to find suitable employees to pursue new types of jobs.To stay in employment, workers need to learn new skills and combine existing skills in new ways.To stay competitive, employers need to invest in reskilling their workforce and talent acquisition.However, for many of the newly emerging jobs, precise skill requirements are unclear and constantly evolving.So, how can policy makers, businesses, and workers decide which skills to invest in?
In this paper, we introduce a method to measure the market value of skills that contribute to human capital formation.We propose that the value of a skill is, mostly, determined by its complementarity, that is, by how easily it can be combined with other skills of high value.In other words, the value of a skill depends on its complements, with skills being most valuable when applied in combination with other similar skills.We test our assumption with the use of rich and near-real-time online labour market data from one of the most popular online freelancing platforms.
We perform this analysis against a background of widespread uncertainty among policy-makers, employers, and workers of how to address the growing skill mismatch and resulting labour market inefficiencies, in the face of technological change.The conventional policy responsealigning training programmes with changing labour market demand -is becoming increasingly ineffectual as technological and social transformation outpaces national training systems (Collins & Halverson, 2018).Likewise, large employers struggle to keep the skills of their workforce up to date (Illanes et al., 2018).Employers, workers and education providers seem uncertain about which new, often digital, skill is the first step towards a successful re-skilling trajectory -should workers be "learning to code", and if so, should they be concentrating on Python or Java?Or on something else entirely?To provide reliable information on which skills are most marketable and have a sustainable demand, we propose an evaluation of skills based on complementarity as core features that determine a skill's market value in the real world.
To this end, we use online labour market data from one of the most popular online freelancing platforms.The data on platform transactions (the "projects" or "gigs" undertaken by workers) contains both skill requirements and price information.Using this information and methods from network science we construct a network of co-occurring skills, providing us with an endogenous categorisation of skills and illustrating the context dependency of human capital.
The monetary value of freelance projects (in USD per hour) allows us to statistically assess the market value of the 962 most popular individual skills.We can then build a regression model to test our assumptions about what drives the variance in skill premia, focusing on the value of skills that are frequently used in the domain of AI work.Using this example, we show how the value of a specific skill depends on supply and demand, skill community, and -most importantly -complementarity.Most importantly, we show that the value of a skill is relative, as it depends on the capacities it is combined with.We show why "AI skills" are particularly valuable, as they are frequently combined with other high-value skills.Finally, we contrast our skill premia with automation probabilities and show that some skills are very susceptible to automation despite their high market value.
Our paper adds to four related streams of literature.Firstly, we contribute to an established debate on how to measure human capital.We attach an interpretable economic price to individual skills and we highlight the relevance of complementarity for a skill's value.Secondly, we contribute to scholarly work on how technological change and automation changes the demand for skills.Here, our work suggests a pragmatic approach to the near real-time monitoring of demand and value of individual skills in fast-changing labour markets.Thirdly, we contribute to the growing body of research that uses digital trace data to study labour market developments.In contrast to recently suggested data sources that examine either the market's supply or demand side exclusively, our approach takes both demand and supply into account and, crucially, allows us to attach price tags to individual skills.Lastly, we build on existing attempts to model the complex relationship between skills using network science.We show that a skill's network centrality (i.e.how related it is to other skills) captures its complementarity, which is strongly predictive of its market value.To guide our empirical analysis, we synthesise these scholarly debates into four hypotheses about what determines the value of a skill.
Our findings are applicable in several ways.They allow us to identify and track over time what skills and what combinations of skills are in demand and successful (as an hourly wage premium).Conceptualising the relationship between skills as a network enables us to identify the most valuable complement to add to any existing skill bundle.Thereby, we could support workers in building individually tailored, data-driven reskilling trajectories.We illustrate our method for skill evaluation with online freelancing data for U.S.-based workers, however, the method can be extended to other data sources, such as online job ads or profiles from career building webpages, that cover larger parts of the traditional labour market.

Human Capital Formation
The conceptualisation of human capital has been one of the most prominently discussed topics in economic sociology and labour economics over the last few decades, as novel concepts and metrics have emerged (Angrist et al., 2019).Conventional measures of human capital often rely on counting years of experience, training, or education, or divide workers into categories, e.g., of labourers and management (Willis, 1986).However, a growing body of literature suggests that training duration and broad worker categories fail to address the importance of skill specialisation, diversity, and recombination in knowledge generation (Aggarwal & Woolley, 2013;Hong & Page, 2004;Lazear, 2004;Ren & Argote, 2011;Woolley et al., 2010).The rise of the knowledge economy (Powell & Snellman, 2004) has also sparked new interest in a more nuanced measure of skill composition.In this context, several papers have taken skill diversity and individual cognitive abilities into account to estimate their effect on wages (Altonji, 2010;Autor & Handel, 2013;Borghans et al., 2008;Bowles et al., 2001;Heckman et al., 2006;Ingram & Neumann, 2006).
A central conclusion of past contributions on skill diversity is that the relationship between wages and skills doesn't just depend on a worker's individual skills, but also on how they are combined.The question of skill complementarities therefore arises (Allinson & Hayes, 1996).
For some skills (e.g., programming in JavaScript and visualisation techniques) it is fairly obvious that skill complementarities emerge.The bundle of JavaScript and visualisation skills is more valuable than the sum of its parts, given that jobs in data journalism, data analytics, and data visualisation, for example, would typically require both.Oftentimes, skill complementarities are likely to be limited to their respective occupational domains, e.g., programming in Python and translating Russian should have little complementarity, while being valuable skills in other contexts.Certainly, the value of additional skills depends on the skill portfolio that the worker already possesses (Altonji, 2010).Although the existing literature recognizes the importance of skill complementarities, there have been few attempts to model this formally at the level of skills.One notable exception is Neffke (2019) who investigates the value of having complementary co-workers using a large Swedish dataset on educational specialisations.Introducing educational synergy and "educational substitutability" allows Neffke (2019) to measure and model co-worker complementarity, showing that the value of what a person knows is relative, as it depends on the skillsets of those with whom they work (Neffke, 2019).
Our work focuses on the complementarity between individual skills.We model this complex relationship using network science, and observe that skills of similar market value cluster together.Our results confirm the importance of complementarity in two ways.First, we find that the (absolute) market value of a skill is higher if it has the potential to be more frequently combined with other high-value skills.Secondly, skills that match the overall skill background of workers are of higher value to them than skills that are distant to their skill domain.This illustrates the relative value of a skill, as it depends on a worker's existing skill bundle.

Automation and Technological Change
Modelling human capital at the level of skills represents a promising avenue to analyse the impact of technological change on labour markets.The periodic warning that automation and new technologies are going to replace large numbers of jobs is a recurring theme in economic literature, most recently in the case of digital and AI technologies (Acemoglu & Autor, 2011;Brynjolfsson & McAfee, 2014;Frey & Osborne, 2017).However, measuring the effects of new technologies on the future of work is not trivial.Looking at education and wages at the aggregate level has proven to be insufficient (Acemoglu & Restrepo, 2018;Beaudry et al., 2016;Brynjolfsson & McAfee, 2014), and treating occupations as homogeneous entities with a certain automation probability risks being overly simplistic and could even lead to false conclusions (Frank et al., 2019).Skill requirements of occupations are dynamic, because technological innovations change the demand for specific skills and thereby the skill composition of occupations -aggregating skills into occupations therefore obfuscates the differential impact of new technologies (Frank et al., 2019).Acemoglu & Autor (2011) propose a task-based model in which occupations are classified into routine or nonroutine and physical or cognitive, based on their task content.This approach has proved to be powerful in explaining the hollowing out of middle-class jobs, such as manufacturing.
However, even this level of aggregation might still be too coarse to understand technology's complex impact on labour markets, and further disaggregation to the level of specific skills might be necessary.Indeed, if we conceptualise occupations as "dynamic containers of skills'', and machines as replacing individual elements within these containers, then we need to move the analysis to the level of skills to fully grasp how technology is redefining the demand for human skills.However, reliable data on workplace skills is not widely available.In fact, Frank et al. (2019) identify sparse skills data as the main barrier to forecasting the future of work.
Recent studies (e.g.Acemoglu & Autor, 2011;Alabdulkareem et al., 2018;Frey & Osborne, 2017) rely on survey-based data such as O*NET for the U.S. (or PIAAC for the OECD).While valuable in many ways, O*NET data is relatively static and therefore not ideal to study a fast changing economy (Frank et al., 2019).Indeed, the skills included in O*NET and the corresponding skill taxonomy have tended to be outpaced by the dynamic frontier of technological innovation time and again.
Instead, we propose a new data source and methodology to help overcome the roadblock of sparse skills data.Using project-level skill information from a leading online labour platform, we study skills at their most granular level and at almost real-time.The actuality of our data allows us to monitor ongoing changes in the demand for and the market value of skills.Together with our historical time series data from 2014 to 2021, we are able to determine whether the value of a skill is on the rise or in decline.Moreover, applying tools from network science, we build a dynamic, bottom-up and data-generated classification of skills based on how they are actually used in the market.We can thereby observe how the relationship between different skills changes as they are recombined in new ways.Lastly, we break down the occupation-level automation probability reported by Frey & Osborne (2017) to the level of specific skills and illustrate how automation risk relates to the market value of skills.Taken together, these elements could help policy makers, education providers, and workers identify skills that are worth investing in during times of fast-changing labour market.

Understanding Labour Markets with Digital Trace Data
With the rise of digital platforms as economic intermediaries more and more data about labour market processes has become available online.Indeed, as automated retrieval of large amounts of this economic transaction data has become more feasible in recent years, the study of labour markets with digital trace data is gaining momentum.Besides relying on traditional survey data, researchers have thus started to investigate skill developments with these new digital data sources as well.
While the idea proposed here of using online labour market data for skill monitoring is novel, a number of scholars have explored other sources of online generated data to investigate skill formation.De Mauro et al. (2018), for example, examined the skill complexity of the new profession of data science with data retrieved from various job boards.Similarly, both Börner et al. (2018) and Calanca et al. (2019) demonstrate the increasing relevance of soft skills based on online job vacancy data.Bastian et al. (2014), on the other hand, make use of data from LinkedIn to compare the relevance of certain "hard skills" across industry domains.While the methodology we propose in this paper to study skill development and detect the emergence of novel occupational domains is applicable beyond online labour market data, to things like data from online job vacancy portals (such as Indeed or Glassdoor) or data from professional social network sites (like LinkedIn), these different data sources have particular advantages and shortcomings, as summarised in Table 1.
Table 1.Suitability of online data sources to the study of the demand and supply sides of work.Compared to data from online job vacancy sites and career portals, online labour market data allow for the study of both the demand and supply side of work, including relevant information on prices.Source: Stephany & Luckin (2022).

Online Job Vacancies
Networking Sites

Online Labour Platforms
While online job vacancies cover a large segment of the labour market, including many industry sectors and also potentially non-digital and manual work, they seldom include information on price levels and give no indication of the possible supply in the targeted population.Data from professional social media sites, like LinkedIn, on the other hand, allow for an in-depth analysis of skill compositions in the population.However, no price or income information is revealed, and matching efficiencies cannot be evaluated in the absence of demand-side data.Online labour market data, e.g. from platforms like UpWork or Fiverr, only covers a small segment of the labour market, namely digitised tasks from jobs in the professional service sector.However, these data have the major advantage of containing information on both the demand and supply side of skills.In addition, it is possible to observe the matching process and price, e.g., hourly rate, for each job with a particular skill bundle attached to it.

The Network Value of Skills
The availability of novel data sources representing manifold aspects of interconnectedness has enabled complexity scientists to capture some of the network-related value of skills.Waters & Shutters (2022) examines a U.S. skill network, in which skills are connected if performed by the same worker, focusing on the relationship between network centrality of skills and economic performance.They find that occupations with higher skill centrality are associated with greater annual salaries, and metropolitan areas with higher skill centrality have higher productivity rates.These results suggest that the application of traditional network metrics to this view of cities as complex systems can offer new insights into the dynamics of regional economies.
When it comes to the complementary value of a skill, Dave et al. (2018) show that the relative positioning of a skill in a skill network contains information on how likely it is for a person to acquire it.In their simulation of skill acquisition they show that a skill is more likely to be acquired by learners if it is close to the bundle of their already existing capacities.
Similarly, del Rio-Chanona et al. ( 2021) use network metrics to track worker transitions between occupations due to automation shocks.They show that transitions between occupations are more likely if skill similarities are high.This leads to the assumption that certain "bridging" skills are particularly valuable, as they enable a transition from one (less profitable) occupation to the next.This complementary aspect of skills is outlined by Alabdulkareem et al. (2018) who argue that much of the polarisation across industries and occupations can be described by skill polarisation, which describes groups of workers that are "stuck" in a low-value segment of the skill network with little proximity to the central skills that would allow them to shift into more profitable occupations.
While all of the above contributions highlight the importance of understanding the complexity of skills via network methods, they do not explicitly seek to measure and describe the market value of a skill.Our contribution is to return to the assumption that a skill's network centrality matters, as it captures its complementarity, which should in turn be reflected in its market value.
Moreover, modelling the relationship between skills as a network allows us to map skills by their similarity, with skills that sit closely in the network being more similar.Via the network structure, we can also categorise skills and workers (based on their skill profile) into communities or domains.This can be used to approximate how close or distant a given skill will be to the existing skill bundle of a particular worker.

What Determines the Value of a Skill?
As a theoretical contribution, our work aims to describe the market value of a skill.Based on the theoretical considerations outlined in the previous sections, we argue that three main factors influence the market value of a skill: 1) Hypothesis 1 -Supply and demand.Our assumption rests on an intuitive understanding of market dynamics.We perceive the value of a skill as the price it commands, which is conventionally determined by forces of supply and demand.We expect a skill that is often requested and seldom offered by workers to be of high value.
2) Hypothesis 2 -Community.Past literature has pointed towards a premium on belonging to a specific community of skills (Weeden, 2002).The community membership is an approximation for signalling various aspects of human capital, e.g., software or legal skills are prestigious and difficult to learn.This should lead to systematic-level differences in skill values across communities of skills.
3) Hypothesis 3 -Complementarity.Lastly, reviewing the network science literature on skills, we assume that a skill's market value is determined by its complementarity.
Complementarity captures the potential of a certain skill to be combined with other (high value) skills.Using a network-based approach (Anderson, 2017;Dave et al., 2018;Waters & Shutters, 2022), we measure a skill's complementarity by measuring its (weighted) centrality in the network of skills.

Complementarity from a Worker's Perspective
However, in practice, skills are never applied in isolation.Workers add new skills to their existing skill portfolio combining them into individual bundles of skills.In this scenario, the literature suggests that the value of skills (e.g.Anderson, 2017 or Neffke, 2019) depends on how they are combined -implying that a particular skill will not hold the same value for all workers.For this fourth, explorative hypothesis, we postulate that a particular skill can have a different value depending on the skills it is combined with, as proposed by Stephany (2021).
This builds on the intuitive idea that there are larger (or smaller) synergies between skills, depending on which skills are being combined.For example, compared to the average worker, a

Method Online Labour Platform Data
The data for this analysis stems from one of the most popular online freelancing platforms2 , also referred to as online labour markets as introduced by Horton (2010).These platforms are websites that mediate between buyers and sellers of remotely deliverable cognitive work (Horton, 2010).The sellers of such work are either people in regular employment earning additional income by moonlighting via the Internet as freelancers, or they are self-employed independent contractors (Stephany et al., 2020).The buyers of work range from individuals and early-stage startups to Fortune 500 companies (Lehdonvirta & Corporaal, 2017).Online labour markets can be subdivided into microtask platforms, e.g, Amazon Mechanical Turk, where payment is on a piece rate basis, and freelancing platforms, such as Fiverr, where payment is on an hourly or milestone basis.Between 2017 and 2020, the global market for online labour grew by approximately 50% (Stephany et al., 2021).Online labour market data allow us to monitor skills in a global workforce on a granular level and in near real-time.The data include worker and employer location, project wages, previous income, and project-level skill requirements.

Building a Network of Skills
This work conceptualises the relationship between skills using a network approach.The data consists of a sample of 49,884 freelance projects posted between 2014 and 2021 with multidimensional skill requirements completed by U.S.-based workers.We constrain our analysis to U.S.-based workers to limit variation in wages driven by structural differences between countries (e.g.local price levels) and other unobserved heterogeneity.Using the information on skill requirements contained in each project, we construct a network in which 4,583 skills are represented as nodes that are connected by a link if a worker applies both in a freelance project.The links are weighted according to how often two skills co-occur, as illustrated in Figure 1.

Measuring the Value of Skills
The "premium" of each skill is calculated by comparing the mean rate of projects that require the respective skill with the mean rate of those projects that do not require it: Most projects require more than one skill.In this case, the project is counted for each skill it requires.By calculating this premium, we derive the market value of the 962 most popular skills, that is skills that occur in at least 20 projects. 3It is likely that other factors than the presence of a specific skill determine project wages, such as the experience of the worker.To test the influence of these characteristics, as much as data availability allows, we run a regression model estimating the "price" of a skill, as described in detail in the Appendix.Skill price and premium are strongly correlated, with a Pearson correlation of 0.69, indicating that they both consistently measure a similar concept.It is worth noticing that the correlation between the two metrics is lowest for skills from Finance & Legal work.One could assume that factors other than skills alone, such as official certificates and degrees, matter significantly more in the legal domain.While the price metric allows us to control for differences in worker experiences and time effects, it is less intuitive to interpret compared to the skill premium.The strong correlation between the two metrics encourages us to continue working with the price premium as our preferred measure, as it is easy to interpret.However, we use the price metric in our regression model for additional robustness checks.
Lastly, with our metric of skill value at hand, we investigate how the variance in values can be described by the features proposed in our hypotheses.We build a regression model that allows us to test our assumptions about what drives skill prices.We include the following features: We take a critical stance on potential endogeneity problems with this model.While we think that complementarity and community are exogenous to the other characteristics, modelling supply and demand usually suffers from issues of multicollinearity.However, we do not observe market supply matching demand in the particular online labour platform we analyse here, given there are many workers registered without completing any project (Kässi et al., 2021).
We use a weighted pagerank centrality measure as our proxy for a skill's complementarity, borrowing this concept from various network science applications.For the examples of aviation markets (Chung et al., 2020) and citation networks (Ding, 2011) it has been shown that the complementarity of nodes, such as relevant airport hubs or popular scholars, can be described by a weighted form of eigenvector (often pagerank) centrality.In our scenario of estimating the complementarity of skills, we assume that both the eigenvector centrality and the respective market value of adjacent skills that can be reached from a specific skill node matter for our complementarity.We modify the pagerank centrality formula:

Creating a Network of Skills
We begin modelling the complex relationship of skills by creating a network, as described in the Method section.Figure 2A illustrates the resulting network of skills, i.e the "skill space".We use the Louvain method for community detection (Blondel et al., 2008), with skills clustering into seven communities according to their co-incidence in a project.We label the clusters based on the type of most prevalent skills in each of the communities: "Finance & Legal, "Software & Tech", "Marketing", "Design", "Audio & Video", "Writing", and "Admin".In addition to the community labels, Figure 2A   In contrast to the platform's official occupation taxonomy, our clustering allows us to fully group all skills into communities based on their actual application.In the next stage of the analysis this will help us to evaluate the contribution of community membership to a skill's market value.

Revealing the Complementary Value of a Skill
In this second step, we estimate the individual value of the 962 most popular skills on the online labour market as outlined in the Method section ( -with the node size representing the premium of each skill -are not distributed at random across the skill space.Instead, they seem to have distinct characteristics and a distinct positioning in the skill space.Furthermore, valuable skills are not equally distributed across skill communities, as Figure 1B shows.Skills in the domain of Finance & Legal have, on average, a significantly higher premium that skills in Marketing, which have in return a higher premium than skills in Admin.This finding points towards a confirmation of our second hypothesis, that there are systematic-level differences in skill values across communities of skills. To further test our hypotheses, we run multiple regression models describing the skill premium (and price), as described in the Method section.In a stepwise fashion, our models include supply (number of projects), demand (number of workers), skill communities, and centrality measures (degree and weighted pagerank centrality) of the respective skill.The results of the regression analysis are shown in Table 2.

Table 2. (A) The value of a skill, measured by skill premia in models 1-5 is determined by supply, demand, community, and complementarity. (B) The out of sample R-Squared of each model increases consistently when adding new controls.
Model 1 and 2 show the effect of supply and demand on the value of skills.Here we see, in alignment with our first hypothesis, that skills with lower supply and higher demand have higher values on average.In model 3 we add dummies for the seven skill communities, while excluding the domain Admin as a reference group.We see that the value of a skill clearly depends on the community it belongs to, as all skills apart from Writing work have a higher premium than our reference group, with Software & Tech and Finance & Legal as the communities of highest skill value on average.Accordingly, skills from Admin work have the lowest values on average.In models 4 and 5 we add various network metrics.First, in model 4 we add the degree centrality (logged) of the skill to the model, which reveals that more centrally connected skills have higher premia on average.It is worth noticing that this effect is not driven by the overall occurrence of a skill, which is controlled for by introducing the number of projects.Finally, in model 5, we add the weighted pagerank centrality of the skill.This significantly improves the model performance, with the complementarity of a skill playing a crucial role for its value, as proposed in our third hypothesis.Interpreting the coefficients, we see that supply and demand still have a significant influence.On average, a one percent increase in the number of workers decreases the value of a skill by 30 percentage points while a one percent increase in demand elevates the value of a skill by 22 percentage points.Model 6 repeats this full model setting for skill prices (and complementarity based on price information).
It shows that our assumptions hold regardless of whether skill values are measured by price or premium (see Appendix Section "Calculating the price of a skill").Table 2B shows that the out of sample R-Squared of each model increases consistently when adding new controls.
Interestingly, when introducing the pagerank centrality the coefficients for the skill communities change noticeably -that is, the community premium on skill values diminishes, and disappears entirely for Software & Tech work.We think that this finding is driven by the fact that much of the premium of skills in Software & Tech is determined by their complementarity rather than by the signalling value of the community.This assumption is confirmed when turning to Figure 3, which plots the 962 skills across their complementarity and premia.19 Here, we clearly see that the most valuable skills are skills with a strong complementarity; that is, those skills that are connected to many other skills of high value and centrality.These skills are not necessarily the most in demand, which is represented by the size of the node.However, we clearly observe that skills from specific communities, such as Software & Tech or Finance & Legal, exhibit significantly stronger complementarity and higher skill premia.Accordingly, skills such as Python (Software and Tech) or Contract Law (Legal and Finance) could be described as hubs in network terminology, as they allow to be combined with many other skills of high value.
At the same time, the results of our regression analysis confirm our assumption that -besides market forces of supply and demand -the value of a skill is, influenced by both its membership to a specific skill community (e.g.Software and Tech) and by the complementarity of the skill (e.g.Python).However, our analysis also indicates that the benefits of strong skill complementarity are not distributed equally across skill communities, as illustrated in Figure 3B, which depicts the relationship between skills premia and their complementarity for each of the seven skill communities.We see that both the baseline of skill values in each community (intercepts) and the benefit of a stronger complementarity (slope) are different across skill communities.In particular, the community of Finance & Legal and Marketing skills seems to be different to the other groups, as we see that the additional benefit of a stronger complementarity is significantly higher than in the other five communities.In Finance & Legal, furthermore, we observe a high initial level of skill values, which might indicate the strong signalling value of this community of skills, as they are highly formalised skills from a profession.

Understanding the Complementarity of Skills and Workers
In summary, our analysis indicates that the degree of complementarity of a skill influences its value in the marketplace.The higher the potential of a skill to be combined with highly connected and high-value complements, the higher its own value.While one can explain parts of the value of a skill in absolute terms, regardless of who applies it, our fourth hypothesis proposes that the value of a skill also depends on the skill background of the worker.Specific combinations of skills have specific benefits as they work in a synergistic relationship (Stephany, 2021).For example, Python is required to programme Natural Language Processing or Adobe Photoshop Pro can be used effectively for Logo Design.The skill network that we created already captures the result of these beneficial complementarities in action, as it shows us which skills are most frequently combined with each other.One could use the structure of the network as a piece of evidence to suggest that there is no single best, "most valuable", skill for everyone to learn, as we observe local hubs and clusters of skills rather than a centralised structure in which the entire network evolves around one "superstar" skill or group of very central and valuable skills.
In contrast to the absolute value of a skill, described in the previous section by a skill's demand, supply, community, and complementarity, here we elaborate on its relative values, based on the assumption that the value of a particular skill for a particular worker depends on its complements.To showcase this relative value, we examine skill-worker combinations and the value of a particular skill in different such combinations.To do so, we have grouped the workers in our dataset, just like skills, into seven domains, which share the same labels as the skill communities.Workers are assigned to the domain with which they share the largest number of skills.(This quite clearly distinguishes workers, as for 78% percent of them at least half of their skills stem from one specific community.)We then measure the premium of each of our 962 skills depending on the domain of the worker who is applying them, to investigate beneficial complementarities across skill domains. 4The findings of this analysis are summarised in Figure 4. with an average premium of 26%, and most profitable for workers from Admin with a premium average of 34%.
machines or algorithms in the near future might not be an optimal destination for reskilling interests.However, a skill with low likelihood of automation but little economic reward might be similarly unappealing.For a comparison of our proposed skill evaluation metric, we calculate the automation probabilities of various skills using the automation probabilities developed by Frey & Osborne (2017), who use four-digit SOC codes to identify occupations.We use the matching table developed by Braesemann et al. (2021) to map the four-digit SOC codes to the 98 platform occupations identified in our dataset.After the matching, we calculate the automation probability for each skill via the following formula: Here the automation probability of is given by the average of the automation   probabilities of all occupations that contain weighted by the share of projects  = 1,...,    that are listed under the respective occupation.Our new metric allows us to compare both  premium and automation probability for each skill, as illustrated in Figure 5.Our comparison shows that skills do not group randomly in the space of profitability and automation risk.Broadly speaking, we can see a tendency that profitability and automation risk are actually negatively related, that is, that the least profitable skills tend to have a higher risk of automation.In Figure 5A, we divide the space into four quadrants: high value and low risk (top left), high value and high risk (top right), low value and low risk (bottom left), and low value and high risk (bottom right).Skill domains are distinctly distributed across this space.Many skills from the domain of Software & Tech.are in the top left quadrant, including most of the AI skills.Writing and Design skills are of low market value but also show low risks of automation.
Skills from Admin on the other hand have low market value and high risks of automation.A group of skills that does not follow the negative relationship between premium and automation risk are certain Finance & Legal skills, which are scattered across different premium levels but all with high susceptibility to automation.Our group of AI skills are clearly different from the rest of all other skills, as they have above average premia and complementarity, and below average automation probabilities (Figure 5B).Lastly, we compare the development of skill values over time.We use the example of AI skills to illustrate that the premia of skills are not stable over time, as they are influenced by changes in supply, demand, or complementarity.Figure 6 illustrates this change by comparing the premia of selected skills between the periods 2014-2017 and 2018-2021 (average premia for each period).In Figure 6A, we contrast the change in skill premia for two popular programming languages, Python and Java.While both languages start with more or less the same premia in the timespan of 2014-2018, they develop differently over time.While Java loses its premia significantly, Python gains in market value.A driver of this trend could be a change in popularity, as Python

Conclusion
Technological change does not affect all tasks and occupations equally.New technologies require new skills while making others redundant.In other words, technological change is not "skill-neutral".As a result, the skill composition of occupations changes.To respond effectively, workers need to reskill, however, the precise skill requirements of newly emerging jobs and the economic benefits of learning a new skill are often uncertain and constantly evolving.It is therefore difficult for workers, employers, and policy-makers to build profitable and sustainable reskilling pathways.To address this uncertainty, we propose a method that attaches a market value to skills based on market demand and supply as well as their relationship with other skills.This expands the understanding of human capital formation in at least two ways.First, we reveal the market value of individual skills and track their demand and price over time.We also investigate what drives the value of individual skills.We find that the value of a skill depends (obviously) on market forces of supply and demand, but we find that it is also largely determined by the relationship of the skill to its complements.Skills that are frequently combined with many other valuable skills (i.e., which have high network centrality) tend to be of high value -we can think of these as hubs.According to our analysis, such skills include programming languages like Python or generalistic legal skills like Contract Law.In addition to the absolute value of a skill, we also reveal a relative value that emerges from beneficial combinations of skills.Often skills are most valuable if they are applied together with a similar type of skill bundle.
For a set of 17 AI skills we illustrate how to apply our evaluation metric.We show that skills associated with AI, which is widely considered to be a major breakthrough technology, have a significantly stronger complementarity in the skill space and, hence, higher skill premia, than the other skills in our dataset.In addition, we are able to contrast skill premia with automation probabilities -another relevant metric when assessing the sustainability of skills.This two-fold categorisation adds relevant nuance to debates about the future-readiness of learning a new skill.
Ideally, reskilling efforts should focus on teaching skills that are both economically profitable and less exposed to risks of computerisation.AI skills seem to fall into this category.Lastly, we track the development of skill values over time.We see that AI skills, such as Deep Learning and Python have been gaining in value significantly in recent years.Our model allows us to ascribe these changes to an increase in demand relative to supply.
This study has certain limitations.Firstly, we measure the market value of skills relative to each other in a specific setting.This allows for comparisons between skills, but statements on the absolute value of skills are not possible, and in any case would not be easily transferred from the particular context of our dataset to a different one.That said, while a statement such as "learning deep-learning will increase my hourly wage by 1.31 USD per hour" (see Table A2) would be problematic, stating that "deep-learning might be a more valuable skill to learn than data-analytics" is in line with our approach.
The second limitation relates to explanations.In this paper we find that the market value of a skill is relative, as it depends on how it is combined with other skills.While this makes intuitive sense and is highly relevant for reskilling purposes, it is outside the scope of this paper to develop a theoretical framework and empirical strategy to explain the how and why of these mechanisms.Finally, this work is based on data from online freelancing which comes with several advantages but also a number of pitfalls (see Background section for details).Most notably, online freelancing is limited to fully digital professional services, meaning that many (predominantly manual) occupations are missing entirely.Moreover, relative to the size of the overall workforce, rather few people work via online freelancing platforms.However, with adequate data access, our methodology could be extended to other data sources, such as online job advertisements and online career platforms, which cover large parts of the traditional labour market.

Policy Implications
Our work on categorising and evaluating skills allows for multiple advances in understanding labour market developments.It can help to establish a taxonomy of skills, understand their application and individual complementarity, and enable automated, individual, and far-sighted suggestions on the value of learning a new skill in a future of technological disruption.Hence, policy implications and applications are manifold.workers with a need to reskill could insert their current skill profile, be located in the landscape of skills, and receive targeted reskilling advice.This would allow them to switch to more sustainable occupations that are closely related to their existing skill set with minimal reskilling effort.Via these individualised reskilling recommendations education providers and vocational training organisations could address the urgent need for individualised solutions in adult reskilling.Furthermore, the continuous "pricing" of skills over time enabled by our approach allows reskilling practitioners to monitor the development of skill values and advise workers on which reskilling to "invest" in.
Secondly, official occupational and skill taxonomies could be improved with near real-time online generated data.As technology creates a demand for novel skills, new occupational clusters can quickly emerge and pull away from official taxonomies, such as the European Skills, Competences, and Occupations (ESCO).This will be bad news for both firms and workers, if professional training providers find it hard to "speak" in the same language as market demand.Online generated data, on the other hand, stems from most recent market developments and allows for identification of new occupational clusters, including in-demand skills.These data-driven, near-real time taxonomies could complement conventional classifications.An immediate contribution to current policy efforts would be the continuous (re-)classification of AI and "green" skills or jobs, as the "twin-transition" has been identified as a catalyst for active labour market policies (OECD, 2021b).

Outlook
This work uses data from online labour markets.The major advantage of this data source is the availability of granular information on demand, supply, and wages -and hence on the value of skills.On the other hand, this data covers only one segment of the labour market.That said, the methodology presented here could easily be adapted to analyse other data sources covering larger parts of the labour market, such as job vacancy platforms (e.g.Indeed) and social career platforms (e.g.LinkedIn).Our results have implications for the debate on successful reskilling strategies in times of dynamic technological change.We find that in-demand skills pay off, and that they can be identified.This suggests that individualised skill-centred reskilling pathways could represent a promising avenue to mitigate skill mismatches -assuming access to timely and granular data.
The European Commission has recognised the need and potential of a data-driven approach to closing the skill gap by bringing forward various legislative and policy proposals.The Pact for Skills (European Commission, 2020), launched in 2020, for example, aims to maximise the impact and effectiveness of skills investment, with a particular focus on upskilling and reskilling in the vocational training sector.For a successful implementation of the Pact, two aspects are crucial.Firstly, industry needs for specific skills must be made explicit, and secondly, the unique training histories of individual workers need to be acknowledged.Online generated data of worker profiles presents a promising approach to monitoring occupation taxonomies and skill requirements via online labour platform data.This is very much aligned with Europe's interest in further building out their skill foresight (French Presidency of the Council of the European Union, 2022) via skill anticipation and support for career transitions.It can offer targeted and near-real time reskilling advice to workers, regarding both industry needs and the worker skills required to fulfil them.This action could support the Commission's proposals (European Commission, 2021) for recommendations on individual learning accounts and micro-credentials, fostering the skill-by-skill learning of workers instead of traditional certification.
Similarly, the Commission's 2022 Data Act (European Commission, 2022) has identified the importance as well as the complications of accessing business (and platform) data in the interests of the public, while acknowledging the protection of business interests.However, the retrieval and usage of private sector data, such as online labour market or job vacancy data, by public body institutions, is not necessarily enabled under the new legislation.Enforced sharing of private sector data requires the ex-ante proof of a "public emergency", and the current modes of automated data retrieval, such as web-scraping, could even be prohibited by the Data Act if they were to be interpreted as coercive or deceptive according to Article 11.In light of this well-intended but potentially contradictory proposal to current EU data legislation, future amendments need to land on a mechanism that gives public bodies acting in the interest of the public the right to access data.
Our investigations on the complex ecosystem of skill formation show that online data can be a valuable tool for designing sustainable reskilling policies.To leverage the full potential of this resource future legislation needs to make public interest its focal point, allowing data access via web-scraping, while enabling strategic public-private partnerships to release the full potential of online generated data for the benefit of society.

Calculating the price of a skill
The 962 linear regression models -one for each skill -describe a project's rate with Each of the 962 most popular skills is considered individually as an explanatory feature in the linear regression.The beta coefficients of each of the 962 skill regressions allow us to determine the added value of an individual skill according to the following formula: As the dependent variable, the project wage in USD per hour, is log-transformed, we need to perform an exponential transformation on the coefficient of the skill dummy to calculate the

Figure 1 :
Figure 1: Derivation of a unipartite network of skills (C) from the bipartite network connecting freelancers and skills (B), where two skills are connected if a worker applies both in a particular freelance project (A).

Firstly
, this skill network provides us with an endogenous categorisation of skills and illustrates the context dependency of human capital.Secondly, the value of each project, measured in USD per hour, allows us to statistically assess the value of individual skills.
Value of SkillsFor the representation of skill complementarities, we employ a network structure of how skills are combined with each other.This structure also tells us something about the relationship between skills and workers.Skills that are very distant to each other in the network are not frequently applied by the same worker.This representation should give us some indication about the costs and benefits of complementarity.Accordingly, we assume that it is most beneficial and least costly, in terms of market value, to combine skills that are close-by in the skill space.To test this hypothesis, we represent skills and workers in a simplified fashion.Just as we attribute skills to specific communities, we use the same structure to group workers by their most frequent type of skill, and then calculate the value of each skill for that particular group of workers.This allows us to compare the value of a particular skill in the context of different other skills, e.g., the value of Python for workers (and skill bundles) from the domain of Design with the value of Python for workers from the domain of Admin.
also highlights three of the most prevalent skills for each of the communities, revealing that the network is based on a continuous skill similarity that allows for application domains of skills to overlap.For example, it is worth noticing that the skill Creative Writing is adjacent to the Design community, while it has a very high distance from the writing skill French Translation.Another example is the Audio & Video skill Motion Graphics, which is much closer to the Design skill Drafting than to the same cluster skill Music Composition.To a certain extent these skill communities overlap with the occupation taxonomy provided by the freelance platform (see FigureA2in the Appendix), i.e., the majority of skills in the Design community also fall into the occupation category Design & Creative.However, other communities, such as Admin distribute more broadly across the occupations of Writing, Admin & Support, and Legal.This is in line with findings byAnderson (2017) and matches the categorisation of online labour market skills byKässi & Lehdonvirta (2018).

Figure 2 (
Figure 2 (A) Skills cluster into seven groups according to their field of application.We label the clusters based on the type of most prevalent skills in each of the communities: "Finance & Legal, "Software & Tech", "Marketing", "Design", "Audio & Video", "Writing", and "Admin".(B) Valuable skills -the node size represents the premium of each skill -are not distributed at random across the skill space.We use the software Gephi and the Force Atlas algorithm to layout the network(Bastian et al., 2009).

Figure 3 (
Figure 3 (A) The most valuable skills have high levels of complementarity -We clearly see that skills from specific domains, such as "Software & Tech" or "Finance & Legal", exhibit significantly higher complementarity and higher skill premia accordingly.(B) The relationship between a skill's value and its complementarity is different across skill communities.For Finance & Legal or Marketing skills, a better complementarity translates much more strongly into higher skill values than in other skill communities.

Figure 4 (
Figure 4 (A) The premium of a skill depends on the skill background of the worker.In general, the skill premium is highest for workers that have mostly skills from the same domain.(B) One exception is Software and Tech skills, which seem to be more profitable for workers with skills from other domains.(C) The premium of a skill depends on the skill background, or domain, of the worker.Skills from the worker's domain are usually profitable, as the diagonal in the heatmap shows.Software and Tech skills are an exception, as they are profitable for workers from all domains

Figure 5 (
Figure 5 (A) How does a skill's value relate to its risk of automation?We can compare these two metrics, which are relevant for assessing a skill's sustainability.AI skills, like many other Software and Tech skills are mostly located in the top-left quadrant of the plot.(B) As the t-tests confirm, AI skills are valuable and have below average automation probabilities, while having high complementarity.

Figure 6 .
Figure 6 .For a set of AI and non-AI skills we observe that changes in the value of a skill are closely aligned with changing supply and demand.As demand (relative to supply) for Python (A), Deep Learning (B), and TensorFlow (C) has increased (lower panel), the premium for these skills has risen (upper panel).

First
, reskilling institutions, like the European Centre for the Development of Vocational Training, could be the main beneficiaries of this highly individualised data.The high granularity of online generated data allows us to describe the skill profiles of individual workers and track their development over time.It also enables reskilling institutions to assess the individual complementarities of learning a new skill.Building on online generated labour market data,

Fabian Stephany :
Figure A1 (A) The minority of workers (4%) has had more than 10 projects and more than half of the workers had only two projects or less.(B) For skills, we see that the largest share of skills only occurred in a few projects, only 22% of all skills were applied to 20 projects or more.

β 4 addedFigure A2
Figure A2 Skill application clusters largely overlap with the occupation taxonomy provided by the freelance platform, i.e., the majority of skills in the application cluster of "Design" also fall into the occupation category "Design & Creative".
To further test this assumption, we compare the market value of individual skills in combination with different skill bundles.Thereby, we examine how the value of a skill depends on the type of skills that it is combined with.For reskilling policies, this aspect of our person who is experienced in Logo Design with Adobe Photoshop Pro might benefit disproportionately when learning about Brand Strategy, as this skill has strong complementarities with her existing skill set -in fact Logo Design is often part of a Brand Strategy.investigation is crucial, as workers do not start from scratch but build on existing skill sets.It is thus important for workers to understand which skills are the most valuable complements to their existing individual skill set.
Table A1 in the Appendix shows the top 20 and bottom 20 skills by premium).The premia attached to each skill in Table A1 are derived from actual market demand.While the exact monetary values are specific to one online labour platform in the U.S. they allow us to make quantitative comparisons between skills.We see skills from the domain of Admin work with relatively low skill premia, while skills from Software & Tech or Finance & Legal show higher values.Figure1Areveals that valuable skills

Table A1
The top and bottom 20 skills are ranked by their premium.Many of the top ranked skills stem from Finance & Legal while bottom ranked skills are from Admin work.

Table A2
Of the 42 initially identified AI skills 17 remain after filtering for a minimum of 20 occurrences.They are ordered by their premium.