analysis of decomposed value chains
submitted in partial fulfillment for the degree of master of science Gideon Gabriël Alexander Mooijen
10686290
master information studies data science
faculty of science university of amsterdam
2021-03-25
Internal Supervisor Second Supervisor
Title, Name dhr. prof. dr. P.M.A. (Peter) Sloot mw. dr. V. (Valeria) Krzhizhanovskaya
Affiliation UvA, FNWI, IvI UvA, FNWI, IvI
A methodology to assess the resilience of industrial ecosystems
through the analysis of decomposed value chains
Gideon Mooijen
ABSTRACT
Industrial ecosystems are complex systems that display a high degree of interwovenness. One of the properties that can be observed when regarding the system on a large scale, is their collective resilience. More specifically, the ability of the ecosystem to collectively adapt to perturbation in order to maintain form and function. This research describes an approach that involves target removal of node simulations to analyze the cascading effects of perturbations regarding the availability of raw resources. A case study demonstrates that a propane deficit will resonate downstream and affect the production of both propylene as polypropylene, both deriva-tives of propane. This approach is designed to incorporate micro-level production statistics rather than a connected net-work, in contrary to conventional network-based approaches.
1
INTRODUCTION
A large percentage of the companies in the Dutch industrial sec-tor are located close to each other, forming industrial clusters. In order to accurately assess the resilience of industrial ecosystems, extensive and intrinsic knowledge on the internal mechanisms of industrial clusters is crucial. High-density industrial clusters are networks with strong interwovenness and inter-dependencies: resources traverse through and are processed by multiple nodes before reaching their final state of end-product. A disturbance af-fecting one industry (or node in the system) may lead to a domino effect, resulting in cascading impacts on the rest of the network [1, 2].
Complex Adaptive Systems (CAS) is a framework for studying, explaining, and understanding systems of agents that collectively combine to form emergent, global level properties [3]. Complex systems are highly non-linear, show complex multi-scale behav-ior, have a structure spanning several scales, and evolves and self-organizes through a complex interplay of its structure and func-tion [4]. Earlier research has considered industrial ecosystems as complex-adaptive self-organizing systems[5].
This research evaluates decomposed value chains as a CAS. A decomposed value chain consists of a resource, an endproduct and any intermediate commodity. This paper maintains the following decomposed value chain as an example:
propane → propylene → polypropylene
The dominant technology for the first conversion is steam cracking, an intricate petrochemical process. The subsequent conversion involves polymerization. This second process is, too, complicated and requires demanding and advanced equipment. The resulting polypropylene is, arguably, a more convenient resource. It is the resource for many plastic products, such as jerrycans and soda bottles. This example of a decomposed value chain is unidirectional: the reversed conversions of commodities are physically not possible.
The value chains are not derived from the data: they are required as an input for this approach and can be provided by domain experts. The decomposition of a value chain yields salient information as it exposes which companies need to collaborate to collectively manufacture an endproduct. A value chain emerges through inter-action of individual converters without central organization, each pursuing their own economic goals and bound to market mecha-nisms such as supply and demand. The decomposition allows for analysis of the role and significance of all involved converters.
Statistics Netherlands maintains extensive commodity-level mi-crodata on the financial production statistics of all companies in the Netherlands. It is possible to construct the monetary commodity-associated in- and output per company. In other words: how much money companies earn and spend per resource. This does not trans-late to volume directly: companies in the retail industry that dis-tribute resources add value to the resource, but do not increase the volume of said production. As an effect, it is difficult to pinpoint which companies produce or consume through financial statistics. With these limitations in mind, the data serves as a robust proxy to determine which companies are possible of converting resources. An additional database contains information on the RegKol (RK) of companies. This new information functions as a coarse filter that yields nodes for which it is deemed reasonable that they physically produce commodities, rather than adding value through distribu-tion.
Any commodity can be transformed into a undirected bipartite network. One partition contains all producers of said commodity, the other contains all consumers. This bipartite network forms one layer of the decomposed value chain. The ensemble of all layers form a multiplex network, in this research referred to as the de-composed value chain. Nodes may hold several positions in this multiplex network, as they can produce and consume multiple relevant commodities. Multiple occurrences of nodes in different layers allow the determination of how perturbations propagate downstream.
Conceptually, the position of nodes in this multiplex network roughly translates into their function in the ecosystem. A node that buys resource 𝑎 and sells resource 𝑏 is likely to convert the former into the latter - given a sensible combination (𝑎, 𝑏). The other nodes in that partition of the bipartite network are referred to as neighboring nodes and are attributed with the same conversion ability. These nodes collectively are responsible for the production of 𝑏. If any node fails to produce, the neighboring nodes need to increase their production to meet the required resources. Their capacity to adapt through compensation of failed production is re-ferred to as adaptability. More specifically, this research determines the maximum amount of perturbations that the ecosystem can re-sist. If this amount is exceeded, the ecosystem can not produce sufficient resources to perform all required conversions, resulting in diminished production.
This methodology is performed on the chemical industrial sector in the Netherlands and uses the Production Statistics 2014 (PS) of Statistics Netherlands.
1.1
Research question
Can financial commodity-level production statistics function as a satisfactory proxy for the physical de-composition of value chains?
What insights do targeted removal of node simula-tions on decomposed value chains provide in the con-text of industrial resilience?
1.2
Related work
1.2.1 Decomposition of value chains. The decomposition of value chains is not a novel field of research. Many academic papers de-scribe a methodology that decompose global value chains (GVC) for analysis of international trade [6]. The foundation of the majority of these papers is the intellectual courtesy of Leontief’s model:
. . . which describes the indirect ripple effects among the industry sectors of the economic system - to eval-uate the impact of a disruption in one infrastructure and the cascading effects on all other interconnected and interdependent infrastructures [7]
Wei et. al used this as building blocks to develop the Inoperability Input-Output Model (IIM) of Supply Chain Networks[8]. The authors replicate Leontief’s approach but perform it on a domestic level: using individual producers and consumers as nodes rather than countries. This procedure appears to be a good method to analyze the resilience of industrial ecosystems in the Netherlands, were it not that it requires a full and accurate picture of the interaction of nodes (exchanges of materials). Although Statistics Netherlands has developed an advanced methodology to derive these interactions[9], this research has been designed to neglect these derivations in order to rely on interpolated data as little as possible. In addition, recent research shows that the application of the IIM is problematic and tends to overestimate the subset of impacts that the model is able to quantify[10].
1.2.2 Resilience in the context of CAS. There is no general agree-ment on the definition of resilience in the context of industrial ecosystems. More strongly, resilience by definition is a dynamic concept. Specifically, because resilience is a property of dynamic systems, it is important to focus on system attributes and the dy-namic structure of supply chain networks[11]. Previous research has reviewed 92 works from various research fields that all main-tain a unique definition of resilience[12]. Despite lack of consensus, most authors agree that resilience is a multifaceted property with, amongst others, the following dimensions.
Adaptability. The adaptability of a system reflects its capacity to adapt (not just resist) to novel and unexpected changes[13]. Adaptability, therefore, considers the capacity of a system to self-organize and reconfigure its structure and behavior to satisfy new conditions, such as the changing demand of final consumers[14, 15]. Dependency. Nodes that are strongly dependent on resources and can not use substitutes, decrease the overall resilience.
Monopolization. The production of a commodity is considered to be monopolized if it is unevenly distributed. As a result, the resilience of the ecosystem decreases.
Supply chain length. The resilience of industrial ecosystems de-creases as the supply chains grow: upstream failure can lead to downstream consequences, known as the bull-whip effect [16].
Vulnerability. This property describes how fragile individual nodes in the network are. Any factor that potentially prevents the node from functioning increases the vulnerability. These factors can be exogenous (floods, earthquakes), or endogenous through failure in a part of the production process.
1.2.3 Targeted attacks. The removal of nodes, to simulate random or targeted attacks, is a common technique in the field of Infor-mation Systems to analyze the impact and consequences on an ecosystem[17–19]. It is performed in a wide variety of domains to gain more insights in the resilience of networks. The fundamental difference between the conventional approach and this research are the characteristics of the network: this research aims to assess resilience without knowledge on the interaction of nodes. Because of this lack of flows of resources between nodes, the simulation of targeted attacks is limited. The simulations are used to deter-mine whether the ecosystem is capable of producing all required (half)products of value chains when faced with perturbations. It is not relevant which specific nodes exchange resources to measure the effects on the availability of (half)products.
2
METHODOLOGY
2.1
Data
The network, as generated by Statistics Netherlands, is the outcome of an advanced procedure. This section describes the building blocks and processes that are required to generate the inter-firm trade network.
2.1.1 A: ABR general company register. This dataset is the outcome of the annual tax reports that all Dutch companies are required to submit. It yields the BEID (company identification number), the gross revenue of the company and general characteristics (location, number of employees).
2.1.2 B: AGT (industry-scale supply and use tables. This procedure builds on Leontief’s model and allow the researches to calculate the domestic production and consumption per industry.
2.1.3 C: IOT (industry-scale input and output tables. These tables hold information on collaborations between different types of in-dustries. For example: the chemical industry and petrochemical industry are associated with a high degree of exchange.
2.1.4 D: PS (production statistics - microdata. Large companies (> 20 employees) are obligated to submit an annual survey. These surveys contain questions about their costs and revenues per com-modity. They are referred to as the Production Statistics (PS) and maintain a financial unit (e ) due to the nature of the questionnaires.
industry BEID volume gross revenue (mil) ind A 101 6000 102 5200 103 3900 104 900 105 720 · · · · 121 200
total volume of ind A total revenue of ind A
industry BEID volume gross revenue (mil)
ind B 201 700 202 580 203 400 204 200 · · · · 221 24
total volume of ind B total revenue of ind B
Table 1: Combining ABR and AGT to obtain all companies that are associated with a commodity.
2.2
Joining and manipulating the data
It is important to address the fact that the Production Statistics do not provide a complete picture of the consumption and production of commodities. In other words: the sum of all production statistics microdata is smaller than the total volume of an industry as given by the industry-scale supply and use tables. This makes sense, there exist companies that do not partake in the survey but do consume or produce resources. It is paramount to acquire a complete picture of the production and consumption of commodities to develop a macro-scale interfirm trade network. Hence, an interpolation step is incorporated to derive the volumes of ’unknown companies’. This is done by constructing two matrices for every commodity: one for consumption and one for production. The rows are filled with all industries that are known to process the commodity. Subsequently, the rows are unfolded and filled with all companies that belong to that industry - as provided by the ABR.
The industry-scale supply and use tables function as the marginal totals, yielding the total volumes of the commodity. Subsequently, the microdata production statistics can be loaded. The outcome of these steps can be seen observed in Table 2
The final step is to perform a procedure called Iterative propor-tional fitting (IPF). This algorithm estimates the volumes of the unknown companies. All companies with a revenue smaller then e 10.000 are neglected. The algorithm attributes that these com-panies do not process the commodity. This exact value of this parameter is not justified and manipulation of this arbitrary num-ber is currently under evaluation for novel research at Statistics Netherlands. The following step is the calculation of the industry significance. This is a fraction, and is simply the revenue of the
industry BEID volume gross revenue (mil)
ind. A 101 610 6000 102 530 5200 103 3900 104 85 900 105 720 · · · · 121 200 4320 44000
Table 2: Further enriching of table 1 with microdata PS (red), total industry revenue (green) and total commodity-associated industry volume (yellow).
company divided by the total revenue of the industry. Conceptu-ally, it defines the dominance of a company in its industry. This parameter allows for an educated and weighted distribution of the unaccounted volumes over all unknown companies with a revenue overe 10.000. After this step, all initial values are set. The IPF pro-cedure iteratively adjust the cell values such that the marginals remain satisfied. In this case, the marginals are both the total vol-ume of the industry and the revenue per company. It is important to address the fact that the microdata PS are not fixed, ie., these values can be modified in order to achieve a more coherent matrix. This is by design, as the industry totals are a more significant statistic than individual and manually submitted questionnaires.
The resulting dataset is a list of all producers and consumers of commodities. They require additional manipulation in order to yield an interfirm trade network. The data comprises of two tables per commodity: consumption and production. They are sorted on volume. The next step estimates the connectivity of the nodes.
First, the number of (commodity x company combinations) in the use table is calculated. This is a lower bound for the total indegree, corresponding with the theoretical situation where every using company has only one supplier [9]. This lower bound is raised with an adjustment factor, which is based on assumption (1.3). By definition, the total indegree is equal to the total outdegree. The outdegree is distributed over all producing companies, following findings in the literature that the degree distribution follows a power law distribution related to the relative size of the company. More specifically, research by Watanabe et al. demonstrated the following relationship between the turnover 𝑆 of a company 𝐼 and its number of connections 𝑘:
𝑆𝑖 = 𝑎𝑘 𝑦 𝑖 𝑦=1.3 𝛼0= ( Í 𝑆 1 𝑦 𝑖 𝑘𝑠 )𝑦 (1) [9]
This procedure yields two tables per commodity, one for produc-tion, one for consumption. Furthermore, they include all companies, the corresponding volumes and their estimated connectivity.
2.3
Matching producer and consumer
The last section matches producer and consumer to yield the de-rived network per commodity. This is achieved by a metrics that is referred to as the overall score. It utilizes three different parameters, which all can take any value in between 0 and 1. The overall score is a parameter that is calculated for every consumer. Given consumer 𝑥 ∈ 𝑋 and candidate producer 𝑦 ∈ 𝑌 , the combinations between consumer 𝑥 and all candidate producers 𝑦 ∈ 𝑌 are calculated. The (consumer, candidate producer) pair with the highest overall score is considered a link.
2.3.1 Company score. The first parameter is conceptually similar the the industry significance of a company, but is not a fraction. It is converted to a score ranging from 0 to 1 for convenient incorpo-ration in the overall score.
log score𝑥= 𝑙𝑜𝑔(net turnover𝑥) − 𝑙𝑜𝑔(net turnover𝑚𝑎𝑥)
company score𝑥=1 −
log score𝑥
𝑚𝑎𝑥(log score
(2) 2.3.2 Industry score. This parameter can be destillated from the industry level input/output-tables, as described in Section 2.1.3. This is a (industry x industry) matrix, containing numerical values if the industries are known to collaborate. If the industries are known to collaborate, the industry score is +1. If they are known to not collaborate, the industry score is -1. If there is no information, the industry score remains neutral.
2.3.3 Distance score. The is the Euclidean distance between two companies, given their longitudonal and latitudonal coordinates. This unit is divided by the largest distance between two companies to transform it to a value between 0 and 1.
2.3.4 Overall score. . The three parameters are processed into a singular metric through the following formula:
overall score(𝑥,𝑦) = 𝛼∗ company score
𝑦+
(1 − 𝛼) × distance score(𝑥,𝑦)+
industry score(𝑥,𝑦)
(3)
Without further modifications, this algorithm requires the calcu-lation of all possible (consumer, candidate producer) combinations. As a result, the runtime complexity of the matching algorithm is in 𝑂 (𝑥2). However, the sequentiality of the tables hold salient in-formation: volumes are sorted from high to low. As a consequence, the companies with a company score will be found in the upper regions of the tables. Hence, the calculation of the overall scores for (consumer, candidate producer)-pairs can be terminated as the value continue to decline. The final stap is the attribution of a link, after which the in- and outdegree of the consumer and producer respectively has been updated. This process is executed iteratively untill all nodes are satisfied with regards to their connectivity.
This procedure yields a multilayered network, one layer for each commodity. The final step is to add weights. This is achieved by the distribution of the volume of the supplying company over all receivers. The total volume is distributed over the receivers in such a manner that it follows a power law distribution.
The volumes of producers and consumers are in financial units. They are used as a proxy for physical production of resources or usage for conversion. This is no direct translation. The PS are ought
to be interpreted as import or export in terms of value. Any com-pany that imports a commodity and sells it for the double price, appears to be a producer of said commodity when the initial, unmod-ified data is interpreted as a proxy for volume. The mechanism is straight-forward: adding value does not imply production, it can be achieved by distributing or packaging as well. More specifically, the retail industry is able to exist because of this mechanism. For this research, it is crucial to identify which companies produce and con-sume elementary chemical resources, rather than the distribution of resources. To assess which companies are able to convert resources, the company RegKol (RK) codes are used. This classification de-scribes the core activities of a company. Examples of RK-codes are: ’21000: Farmaceutical industry’, ’10813: Coffee/Tea/Sugar’ and ’46290: Retail’. This classification allows for manual filtering to obtain only the companies of which it is deemed reasonable that they perform industrial conversion of resources. As a result, the aggregated vol-umes of these filtered companies is lower than the industry totals but is a more representative proxy for physical conversion. The rationale behind this procedure is that producers and distributors hold a fundamentally different function in industrial ecosystems. Even though they both add value to resources, distributors are de-pendant on the initial producers to perform their activities. The initial PS do not reflect this difference. The importance of this differ-entiation arises for the simulation of removal of nodes. It assumes the ability of upscaling production of er neighboring nodes. This ability is reasonable for initial producers, in contrast to distributors. The PS cover 711 commodities. A hand-picked selection of 156 relevant commodities are selected and included in the Appendix.
2.4
Method
2.4.1 Decomposition of value chains. The producers and consumers of all commodities in the value chain are transformed to an undi-rected bipartite graph. The first partition contains the producers, the second partition contains the consumers.
Figure 1: Visual representation of an undirected bipartite graph for commodity c. This example contains 5 producers and 3 consumers
This process is repeated for every commodity of the value chain, after which the layers are joined. Subsequently, nodes with multiple instances in the multiplex network are linked. These links can be either intralayer (blue) or external (green). It is important to emphasize that, in the field of Information Systems, intralayer links usually resemble an interaction between different nodes. In this research, this is not the case. Intralayer links indicate that a node is
Figure 2: Full representation of the multiplex network of value chain (𝑟, 𝑖1, 𝑖2, 𝑒)
both a supplier and a user of a commodity. External links indicates that the company is associated with more than one commodity of the supply chain. The result is a multiplex network, with one layer per commodity and two partitions per layer. The decomposed value chain is the ensemble of partitions.
2.4.2 Simulate removal of nodes. The effect of the removal of a producing node can be determined and is used to predict how the ecosystem will be disrupted if any node is deleted. Deletion entails any event that prevents the node from producing: bankruptcy, a terrorist attack, or any other. If a node is removed, all linked nodes are removed as well. They resemble the same physical company but hold multiple roles in the value chain. It is reasonable to assume that the neighboring nodes of the removed node are able to upscale their production processes to compensate for the removed node. This research assumes an upscaling capacity of 20% for every node. In reality, this property is domain-specific and unlikely to be homo-geneous. Given the removal of a node and the assumed upscaling capacity of its neighbors, it is possible to measure the diminished production for every layer of the value chain. It is worth to address the difference between the direct effects of the perturbation - the targeted node is removed from the network, and the indirect effects of a perturbation - a resource deficit results in failed production that cascades downstream.
2.4.3 Resilience assessment. The expression of the resilience of the value chain is achieved by iteratively removing a random node from the producers of the first commodity: the start of the supply chain. Failure of production propagates throughout the network in the downstream layers if:
(1) The removed node holds multiple positions in different layers of the multiplex network.
(2) The diminished production of the removed node causes a shortage for the consumers.
The effects of type (1) can be determined directly through adjust-ment of total production and consumption for any inflicted layer. Type (2) requires a generic procedure that determines which nodes are victimized, i.e., which nodes are excluded from consuming a resource to restore the equilibrium between production and con-sumption. In reality, a resource deficit would increase the price. The victimized nodes would be those actors that are not willing or able
Figure 3: Example of a simulation of producers removal for one layer of the decomposed value chain
to meet this new price. The data does not provide an opportunity to capture this market-mechanism in order to accurately determine the victimized nodes. To resolve this, the procedure simply selects the smallest consumer that solves the supply deficit. The rationale behind this mechanism is that the removal of that node has the smallest impact on the form of the original network.
Figure 3 demonstrates a simulation for a layer of the value chain in which initial producers are removed. The crosshairs (orange) indicate the breakpoint: upon the removal of the 4th node the neigh-boring nodes are not able to upscale their production processes to meet the initial quantity. However, this does not result into a resource deficit: the removed nodes are also consumers of the com-modity. Upon removal of the 5th node, the negated production exceeds the negated demand. The height of the resource deficit determines which nodes are victimized for the lack of resources. This nodes will be deleted in all partitions.
Figure 3 displays merely the first order effects of removal of producers. The resource deficit that occurs after the breakpoint indicates that there is a shortage of resource r. There are fewer input resources available for the producers of 𝑖1than that is used
in the current situation, without perturbations. Hence, a deficit of resource 𝑟 results in a trailing deficit of 𝑖1, 𝑖2and endproduct 𝑒. It is worth mentioning that from the second layer and upwards, upscaling of the neighboring nodes is not an option as the root of the inability of production is lack of availability of input resources. Even though some scientists argue that it is unreasonable to capture resilience of ecosystems in a singular metrics [7], this re-search does aim to capture this property in a quantified manner. More specifically, the resilience of the decomposed value chain is expressed in the curvature of diminished production of downstream commodities as effect of upstream failure. This can be interpreted as the capacity of the ecosystem to upscale production to counter the loss of neighboring nodes in the removal of nodes simulations. Conceptually, it describes how many hits the ecosystem can endure before it starts decomposing - unable to maintain overall form and function.
Many researches that include simulations in which nodes are iteratively removed to measure the consequences. Oftentimes, the researchers repeat the experiment frequent to measure the average effects and the standard deviation. In general, these findings give a complete and reliable notion of the expected outcome of the experiment. In this case, the targets - the removed nodes - are highly heterogeneous. Many commodities have a monopolist, or a small number of nodes that account for the majority of the production of said commodity. In that case, selecting a random number of nodes to delete to measure the effects will not provide significant information: if the targeted nodes are small, negligible producers, it will have no effect on the ecosystem. Hence, the random sample of targeted nodes define the outcome of the experiment, even though there is no plausible explanation for the sudden absence of a group of nodes. A scenario in which malicious actors purposefully try to attack a specific industry, however, is a possible scenario that is worth to investigate. To resemble this situation, the algorithm examines the worst-case scenario, in which the node that is targeted first, is the largest producer of all. The second target is the second highest producer, and so onward.
2.4.4 Resource deficit. This approach artificially removes consumers of resources in case of a deficit to resemble the ecosystems capac-ity to adapt in order to maintain form and function. This is not a straightforward procedure. If all companies in the Netherlands produce 200 units of oil, and they consume merely 100 units of oil, an equal part is assumed to be used for export. Consequently, in case of an oil deficit, this approach assumes that the deficit is equally distributed over the designated users. In this case, 10% of the domestic oil consumers must be artificially removed in order to achieve a new equilibrium. Inversely: if there is a resource deficit in the initial state, the assumption is that the ecosystem relies on import of resources to begin with. To summarize: if the node is an exporting hub, the resource deficit is inflicted on both domestic and international consumers. If this is not the case, the full deficit is inflicted on domestic consumers.
Algorithm 1 Simulate impact of removal of nodes
1: Calculate propagation of removed initial producers
2: for for producer p in partition0do 3: for partition in partitions do: 4: delete producer p
5: calculate new supply, new demand
6: net loss = new supply + upscaled production
7: diminished demand = demand - new demand
8: while net loss > diminished demand do victimize node
9: end while
10: end for 11: end for
12: return statistics
3
RESULTS
Figure 4 demonstrates the cascading effects of targeted attacks on the largest propane producers. The figure on the right shows that the intermediate commodity propylene is moderately affected by the perturbations with regards to polypropylene, the end product. This is consistent with the hypothesis that the upstream conversion are more strongly affected by the perturbations than the final, down-stream effects. The red line, indicating the production of polypropy-lene remains moderately horizontal. A completely horizontal line would indicate that there are absolutely no cascading effects. In that case, a value chain is highly resilient.
Figures 5 and 6 demonstrate that there exist two large consumers of both propane and propylene. After the first two nodes are re-moved, the demanded resources decline a substantional volume: indicating that a large consumer is victimed in order to overcome a resource deficit.
Figure 4: Absolute (left) and relative (right) reduction of commodity production.
4
DISCUSSION
One salient finding is the difference of the lines in figures 5 - 7. The black curves describe the demand of commodities and has a choppy, non-smooth curve. This is in contrast to the supply curve, which are smooth curves. This is in agreement with the expectations: each newly removed node affects the availability of resources. However, only when the demand exceeds the supply, the largest demanding node is victimized in order to guarantee that supply and demand are in equilibrium. This yields a ’buffer’, a volume of demand deficit that allows for a supply reduction without the need of a new victimized node.
4.1
Limitations
4.1.1 The inability to address international nodes. Statistics Nether-lands does not have access to international Production Statistics. As of such, the current implementation of matching supplier and user does not incorporate possible transactions between the high-density industrial region of Zeeland and the port of Antwerp. Interviews with domain experts have provided the notion that these collab-orations are very likely. Due to the nature of the algorithm, all suppliers and users are to be matched to satisfy the conditions of volume and connectivity. Because this prohibits external links, this introduces an inevitable bias.
4.1.2 The simplification of the retail-sector. One of the assumptions of this research is that the retail sector does not hold a meaningful position in the context of the resilience of the ecosystem. Accord-ing to the ABR, companies are not multidisciplinary. Therefore, companies belong to exactly one industry, even though they can be both producer and retailer of a commodity. It is impossible to differ-entiate between retailers that add value through distribution and retailers that factually produce materials, because of the monetary nature of the data. The current approach neglects these multiplex roles and simplifies the real life situation.
4.1.3 Granularity of data. The categorization of resources into 711 commodities, which are reduced even further to 156 commodities
for this case research, is a simplified representation of the real situation. In order to accurately analyze value chains in Dutch industrial regions, it is desirable to be able to differentiate between very specific comomdities, rather than the aggregate form in which they are processed in this research. As an example: the commodity group 2014400 is defined as Aminozuren (Amino-acids), whereas more specificity on the exact type of resource would allow for a more extensive analysis of different value chains.
Figure 5: Changes in supply and demand of propane after removal of the 50 largest propane producers
Figure 6: Changes in supply and demand of propylene after removal of the 50 largest propane producers
Figure 7: Changes in supply and demand of polypropylene after removal of the 50 largest propane producers
5
CONCLUSION
The procedure yields interesting results that grasp the downstream effects of upstream perturbations. However, the applicability of the research is limited as it does not evaluate the actual, real life, inter-firm trade network. The procedure that derives the network utilizes many steps of interpolation, assumptions and estimates. Rather than utilizing that noisy network, this research assesses the adaptability and thus resilience of the state space of value chains. The case study demonstrates that the ecosystems capacity to pro-duce propylene declines as the propro-ducers of propane cease to exist. Furthermore, the production of polypropylene is also influenced by the diminished production of propane, but to a slighter degree. The added value of the analysis of a singular value chain is limited, but the comparison of different value chains with identical resource 𝑟 can be of aid in comparing how strongly downstream conversions are connected to said resource. It is difficult to capture a complex and multidimensional property as resilience in a single metric. How-ever, the resulting graphs demonstrate the snowballing effects of upstream perturbations. For further analysis, it is crucial to improve the derived inter-firm trade network by incorporating international links and validate this novel network. An accurate inter-firm trade network provides a solid foundation for more extensive analyses.
5.1
Further research
This case study evaluates a value chain with one endproduct. The most obvious feature is the comparison of two different endprod-ucts. If there was an alternative endproduct (endproduct X), it would be interesting to compare the adaptability of:
propane → propylene → polypropylene with the adaptability of:
propane → propylene → endproduct X.
This would allow for a direct comparison between the downstream effect of upstream perturbations. This allows for an objective quan-tification of the resilience of a singular conversion in the context of a larger value chain.
ACKNOWLEDGEMENTS
I would like to express my very great appreciation to Peter Sloot and Valeria Krzhizhanovskaya, for their guidance and useful critiques during the development of this research. Also, Gert Buiten at Sta-tistics Netherlands, and Carolina Mattsson were of great aid during our numerous video-meetings. I am thankful for the Ministry of Economic Affairs and Climate Policy for this interesting research problem. I want to thank Joyce ten Holter, Eelke Heemskerk, Cees Diks and Frank Takes for their collaboration and for making me part of such an interesting research group. Finally, I want to thank my girlfriend Sophie for her never ending encouragement and support.
REFERENCES
[1] Brad Allenby and Jonathan Fink. Toward inherently secure and resilient societies. Science (New York, N.Y.), 309:1034–6, 09 2005.
[2] Frank Boons and Wouter Spekkink. Levels of institutional capacity and actor expectations about industrial symbiosis. Journal of Industrial Ecology, 16(1):61–69, 2012.
[3] Ted Carmichael and Mirsad Hadzikadic. The Fundamentals of Complex Adaptive Systems, pages 1–16. 06 2019.
[4] Amit Surana, Soundar Kumara, Mark Greaves, and Nandini Raghavan. Supply-chain networks: A complex adaptive systems perspective. International Journal of Production Research - INT J PROD RES, 43:4235–4265, 10 2005.
[5] Junming Zhu and Matthias Ruth. Exploring the resilience of industrial ecosystems. Journal of Environmental Management, 122:65 – 75, 2013.
[6] Zhang Wei and Chen-Chen Yong. Decomposition of global value chains for equipment exports of china. International Journal of China Studies, 9:79–96, 04 2018.
[7] Yacov Y. Haimes. On the definition of resilience in systems. Risk Analysis, 29(4):498–501, 2009.
[8] Hairui Wei, Ming Dong, and Shuyu Sun. Inoperability input-output modeling (iim) of disruptions to supply chain networks. Systems Engineering, 13(4):324–339, 2010.
[9] Sjoerd Hooijmaaijers and Gert Buiten. A methodology for estimating the dutch interfirm trade network, including a breakdown by commodity. https://www. oecd.org/naec/new-economic-policymaking/Buiten_Hooijmaaijers.pdf, 2019. [10] Jan Oosterhaven. On the limited usability of the inoperability io model. Economic
Systems Research, 29:1–10, 03 2017.
[11] Henry Adobor and Ronald McMullen. Supply chain resilience: a dynamic and multidimensional approach. The International Journal of Logistics Management, 07 2018.
[12] Rodrigo Reyes Levalle and Shimon Nof. Resilience in supply networks: Definition, dimensions, and levels. Annual Reviews in Control, 43, 03 2017.
[13] Simon Levin. Complex adaptive systems: Exploring the known, the unknown and the unknowable. BULLETIN (New Series) OF THE AMERICAN MATHEMATICAL SOCIETY, 40, 11 2002.
[14] Hau Lee. The triple-a supply chain. Harvard business review, 82:102–12, 157, 11 2004.
[15] Edward Hearnshaw and Mark Wilson. A complex network approach to supply chain network theory. International Journal of Operations & Production Manage-ment, 33, 03 2013.
[16] Hau Lee, V. Padmanabhan, and Seungjin Whang. Information distortion in a supply chain: The bullwhip effect. Management Science, 43:546–558, 12 2004. [17] Giannis Moutsinas and Weisi Guo. Node-level resilience loss in dynamic complex
networks. Scientific Reports, 10(1):3599, 2020.
[18] Marco Janssen, Örjan Bodin, John Anderies, Thomas Elmqvist, Henrik Ernstson, Ryan Mcallister, Per Olsson, and Paul Ryan. Toward a network perspective of the study of resilience in social-ecological systems. Ecology and Society, 11, 11 2005. [19] Alessio Pagani, Guillem Mosquera Doñate, Aseel Alturki, Samuel Johnson, Stephen Jarvis, Alan Wilson, Weisi Guo, and Liz Varga. Resilience or robust-ness: Identifying topological vulnerabilities in rail networks. Royal Society Open Science, 6, 01 2019.
500000: Steen-/bruinkool 610000: Aardolie&-condensaat 610110: Aardolie ruw 610120: Aardgascondensaat
620000: Aardgas 700000: Metaalertsen 710000: IJzererts 720000: Non-Ferro ertsen
800000: Overige delfstoffen 811000: Natuursteen 812110: Zand 812120: Grind
812200: Klei 891000: Mineral.v.chem.ind. 892000: Turf 893000: Zout
899000: Delfstof neg 900000: Dnst. tbv delfst.win 1910000: Cokesovenproducten 1910209: Ov.olie-/kolenprod.
1920211: Benzine 1920231: Nafta’s 1920241: Jetfuel 1920249: Bunker. jetfuel
1920261: Gasolie grondst. 1920262: Diesel 1920263: Gasolie verwarming 1920269: Bunker. diesel
1920270: Petroleum 1920280: Stookolie 1920289: Bunker. stookolie 1920290: Smeerolie
1920311: Vloeib.PropaanButaan 1920312: Autogas (lpg) 1920320: Overige gassen 1920490: Briket&ov.aardoliepr 1920930: Ov.vloeib.Brands/LPG 2010900: Anorg.grst.basischem 2011000: Industriële gassen 2012000: Kleurstoffen
2013100: Splijt-/kweekstof 2013240: Zuren 2013890: Overige zouten 2013990: Anorgan.grondst.e.d.
2014000: Organ.grondstof. 2014113: Ethyleen 2014114: Propyleen 2014118: Koolw.open.
2014122: Benzeen 2014125: Styreen 2014129: Ov.koolw.ges 2014199: Halogenen/Fenolen
2014220: Alcoholen 2014300: Carbonzuren 2014400: Aminozuren 2014529: Caprolactam ed
2014639: Ethers ed. 2014730: Aromaten 2014745: Alcohol >80% 2014990: Ov.organ.grondst.
2015100: Kunstmest 2015900: Ov.stikstofverbind. 2016000: Kunstharsen 2016100: Polyetheen
2016200: Polystyreen 2016300: Pvc 2016400: Polyacetaten 2016510: Polypropyleen
2016520: Overige polymeren 2016540: Polyamide 2016550: Polyurethaan 2016599: Overige kunstharsen
2017000: Synthetische rubber 2020000: Bestrijdingsmid. 2030000: Verf/vernis/inkt ed 2030100: Verf/vernis
2030240: Drukinkten 2030299: Ov.verfproducten 2040000: Zeep/poets/parfum ed 2041000: Was-/reinig.mid.ed
2041300: Zeep/poetsprod. 2042110: Parfums ed. 2042120: Huidverz.mid 2042160: Haarverz.mid
2042199: Ov.kosmetische prod. 2050600: Ov.chemische prod. 2051000: Vuurw/Springstof/Luc 2052900: Lijmen/gelatine 2053000: Etherische oliën 2059100: Fotochemische prod. 2059910: Biobrandstof e.d. 2059920: Chemische prod. neg 2060000: Kunst-/synth.garens 2110000: Farmac.verbind/grstn 2110910: Farmaceut.verbind. 2110920: Farmaceut.grondst.
2120000: Farmaceut.producten 2120100: Geneesmiddelen 2120210: Sera/vaccins 2120240: Gaas/verband
2120290: Ov.farmac.prod. 2120299: CocaïneHeroïneXTC 2210000: Rub.band&ov.prod.rub 2211010: Autoband ed.
2211099: Ov. Banden 2219000: Ov.Rubberprod. 2221000: Kunststof plaat e.d. 2221290: Staaf/slang v.kunst
2221300: Plat.ongecel.v.kunst 2221400: Ov.platen v.kunst 2222000: Verpakking v.kunst. 2223000: Bouwart.v.kunst. 2229000: Ov.product.v.kunst. 2309000: Overige bouwmaterial 2310000: Glasproducten 2312199: Vlakglasproducten 2313199: Glaz.FlesPotVaas ed. 2314990: Ov.bewerkte glasprod 2323400: Ov. Keramische prod. 2323456: Cement&Keramis.prod. 2339000: Keram.Bouwmat/Tegels 2341000: Keram.sier/huish.art 2351900: Cement/kalk/gips 2360900: Producten van beton
2361110: Stenen van beton 2361199: Overige betonwaren 2361900: Bouwelem.v.beton 2363400: Beton/mortel
2370000: Bewerkte natuursteen 2390000: Bouwmaterialen neg 2410120: Ferro primair 2410345: Fe gewal.onb
2410512: Fe gewl.bekl 2410600: Ferro gewalst, rond 2410790: Ferro profielen 2412300: Ferromet.&plaat/buis
2420900: Ferro buizen 2432120: Plaatststaal 2439000: Overig staal 2440000: Non-ferromet&halffab
2442110: Aluminium, ruw 2442120: Aluminiumoxyde 2442200: Alumin.halffabrik. 2443120: Zink, ruw
2443220: Zink halffabrik. 2444100: Koper 2444200: Koper halffabrik. 2449190: Ov.non-ferrometalen
2459000: Gieten metal 2509000: Ov.metaalproducten 2511000: Metal.constructiewerk 2512000: Metal.deuren/ramen
2515000: Met.Constr./DeurRaam 2521100: CV-ketels/radiatoren 2521900: Metal.tanks/reserv. 2530000: Metal.stoomketels
Table 3: All implemented commodity groups (156/711)
