Abstract
This work introduces a multiagent model for informal communication about cheating merchants among investing Genoese traders in the 12th century. The model builds on previous, game theory focused work of Avner Greif and extends it by enabling communication between Genoese traders. A trust-based cooperation model is tested across different network topologies as well as two different communication modes: (a) a reactive one representing Genoese trader interrelations and (b) a more proactive one that is based on accounts on the relationships among a North African traders collective, known as the medieval Maghribi Traders. The simulation shows that even for high levels of initial trust among Genoese, their mode of communication would not have sufficed to collectively identify cheating merchants and prevent them from future transactions. The results obtained from the tested network topologies are further discussed in the light of literature accounts of both historical societies.
Keywords
Introduction
This work addresses the formation processes of economic institutions. Institutions in the context of economics represent “the rules of the game” (North, 1990) and encompass a variety of human behaviors, ranging from behavioral conventions, progressively to norms, and on to legally codified rules. The importance of institutions is widely acknowledged, but the processes surrounding their emergence continue to be puzzling. Recent notable works that venture into this area are Acemoglu and Robinson (2012), North, Wallis, and Weingast (2009), and Greif (2006).
Greif’s work is especially noteworthy in that he undertakes a comparative study of historic occurrences, based on the method of analytic narratives (Bates, Greif, Levi, Rosenthal, & Weingast, 1998) that combines rational choice theory with game-theoretical analysis, to make convincing hypotheses about the role and emergence of institutions grounded on concrete examples inspired by historical circumstances.
The work described here adopts his theme but applies a different method. While Greif’s work is inherently based on game theory, which implies the interpretation of institutions as equilibria, and pioneered by Menger (1963) and von Hayek (1945, 1967), we employ agent-based simulation to emphasize the dynamics that underlie such a development. Instead of making strong assumptions about the existence of an institution, we shift the perspective and look into aspects between the lines of this work, in particular the strong distinction between societies based on formal institutions versus those reliant on informal institutions. We wish to emphasize a more dynamic perspective on the fuzzy institutional environment one encounters in complex social systems and thus test Greif’s assumption that informal institutions would not have been sufficient to sustain cooperation in an individualistic society. This way, we avoid a concentration on static–comparative analysis of the fixed social configurations presented by Greif, and thereby provide a less rigid base model that incorporates environmental aspects stronger than a rational two-player game can capture.
We do that by using Greif’s (1989, 1993, 2006) Maghribi Traders’ Coalition, which describes the North African Jewish trader collective that operated as long-distance traders along the North African shore between the 10th century and early 13th century. Based on specific features of the group—its closure to outside entry, its collective rule set, and a contrasting social environment characterized by Islamic rule—Greif suggests that their ability to sustain compliant in-group behavior was based on informal institutions, such as the reciprocal ṣuḥba relationship, 1 rather than on available formal institutions such as the “qirad algoyim,” which represented a formal contract that indicated rights and obligations of either party (Pryor, 1977; Udovitch, 1962; van Doosselaere, 2009). 2
In contrast to the “Maghrebis” [sic], 3 traders in medieval Genoa could not rely on completely informal networks, which gave rise to a use of formal institutional mechanisms to govern economic behavior.
Both societies’ (Genoese’ and Maghribis’) business was long-distance trade involving a combination of a stationary investor/trader who owned goods to be sold (the commendator of a commenda) and a traveling merchant (the commenda’s tractator) who would undertake the long-distance journey (often lasting several months) to sell those goods at remote trade locations far beyond the reach of the stationary investor. The realized profit was then to be returned to the original owner of the goods, who would then pay his traveling partner (in the case of Genoese trading) or grant reciprocal favors for future business transactions (in the case of Maghribis). The respective institutions by which cooperation was maintained (as mentioned before) set the two societies apart.
The contribution of the work presented here is 2-fold. As a primary aspect, we intend to translate the communication aspects of Greif’s scenario from a game-theoretic perspective into an agent-based approach, in order to make the dynamics of this scenario accessible. As a second aspect, we relax some of the strict assumptions of the original analysis to arrive at a better understanding of why informal communication may have sufficed to sustain cooperation in the Maghribi case, and why it did not for the Genoese trader society.
To present our case, we first outline Greif’s modeling approach in the second section and suggest that even though the Genoese society relied on contracts to sustain cooperation, they were still likely to have also relied on informal information exchange (though this was probably not sufficient to overcome the problem of misreporting earnings to investing partners or other means of cheating). In the third section, we introduce a base model of informal communication among Genoese. After presenting results for this model (fourth section), we include the more proactive Maghribi communication pattern into the same model (fifth section) to show its built-in potential to exclude cheaters from trade participation. The final sections summarize the results (sixth section) and situate this work in a multidisciplinary literature context (seventh section).
Motivation
In his work, Greif (1989, 1993, 2006) sharply contrasts two societies, the Maghribi Traders Coalition and the Genoese traders, as archetypes for individualistic and collective societies, which are, respectively, supported by the predominant use of formal versus informal institutional mechanisms. This is, at least in part, relativized by more recent work (Gonzalez De Lara, 2008) that puts stronger emphasis on the complementary role that informal private order enforcement (which Greif analyzed and discussed, but did not incorporate into his model [Greif, 1994, 2006]) played in relation to formal legal mechanisms to regulate long-distance trade in medieval Genoa. Following these formal and informal protocols, investors would not only rely on sanctioning of contract breaches by commercial courts but could also extend threats to local relatives of violating merchants, should those withhold or misreport profits, or worse, fail to return to Genoa.
Moving From Game Theory to Multiagent-Based Simulation
The game-theoretical approach taken by Greif (1989, 1993, 2006) effectively contrasts differing societies. Thus, Greif differentiates Maghribian and Genoese traders in a dichotomous manner that sharply contrasts both groups, leaving the reader with a perception of perfectly matching counterparts.
In this historical situation, there were two only slightly overlapping geographic domains of this trade. Genoese traders largely served along the northern Mediterranean coastline reaching from Spain to the Black Sea (and later extended to the Northern European coastline), while Maghribis, generally based in Fustat, the capital of the Fatimid empire, traded their goods along the North African coastline, Sicily, and reaching up to Andalusia in the west and Constantinople in the east 4 (see map in Figure 1 5 for an overview of the geographic spread). In the following, we concentrate on the modeling of information transmission networks that Greif considers crucial for the regulation of cheating behavior by means of communication. We will interleave these aspects continuously with additional information relevant for modeling decisions.
Greif’s work relies on game theory and the interpretation of Nash equilibria as a model of institutions emphasizing a comparative-static analysis. This requires modelers to introduce strict assumptions and engenders an emphasis on contrasting stereotypes (defining fixed player preferences) while avoiding subtle, but potentially important, differences between societies (e.g., more fine-grained modeling of roles within societies). The use of Nash equilibria as a model of institutions further motivates a strong analytical abstraction from the reality (e.g., raising the question about the meaning of a discovered equilibrium), which bears the risk of assuming a strongly reductionist perspective.
Geography of the Mediterranean Basin during Fatimid reign.
Our work builds on Greif’s achievement, but we alter the method of analysis from a game-theoretical approach to that of individual-based multiagent simulation (J. Epstein & Axtell, 1996), which allows us to characterize individual entities, thereby leading to more nuanced perspectives on the historic developments. Inasmuch as game theory offers a method to develop strictly formalized utilitarian models, its outcome aims at determining evolving equilibria based on a unified formal representation of micro-level entities. This calls into question whether game theory is appropriate for the analysis of complex scenarios of social interrelations, since (classical) game-theoretic models ignore the potential dynamics that can unfold based on changing players, the player’s preferences, or even the “rules of the game” (e.g., mitigated by dynamically changing social network graphs, change in social structure, or negotiation of rules at runtime). 6 In contrast, multiagent-based simulation, which has been applied in a variety of scenarios from the area of social science (e.g., modeling of social behavior after disasters (Chandan et al., 2013), influence of extremists on cultural polarization (Shutters, 2013), massive economy models (Deissenberg, van der Hoog, & Dawid, 2008) allows us to paint a more complex and dynamic picture of social relationships which reduces the reliance on strict stereotypical behavior and the consideration of nonrational behavior in individuals.
Data relevant to pursue this analysis are collected from secondary literature, with Goitein (1967–2000), and more recently Goldberg (2012) for the Maghribi side. Both of those analyze the Geniza, a collection of letters and documents conducted in Hebrew and stored in the Ben Ezra synagogue, which survived the destruction of Fustat in 1168 by its ruler to prevent it from being captured by the crusaders. Today, its documents (containing around 1,500 business letters) are an important source for reconstructing information about all aspects of life of the Jewish community in Fustat and surroundings during the period between 990 and 1150 AD and beyond. The most comprehensive translation and compilation of Geniza documents is provided by Goitein (1967–2000), which made it accessible to the wider research community. Although his accounts on Jewish life are very comprehensive, his interpretation reflects a positive bias toward the society he revives from the Geniza (see Goldberg, 2011, for a discussion). Further studies, taking a more analytical perspective on selected aspects of the Geniza, include those of Goldberg (2012) and Ackerman-Lieberman (2007).
For the Genoese perspective, we source data from Greif’s (1994, 2006) studies on Giovanni Scriba’s cartulary containing 612 Genoese trade documents covering a period from 1155 to 1164. For further material on the Genoese side, we rely on van Doosselaere’s (2009) coding of notary documents, which provides an insight into the formal relationship between Genoese merchants and investors and which sheds light on the nature of those relationships.
Table 1 summarizes contrasting characteristics of both societies (including some aspects that reach beyond the context of this discussion).
Characteristics of Genoese and Maghribi Societies.
Informal Communication in the Genoese Society
For our initial experiment, we examine the Genoese society during the period around 1150, when contract enforcement—as far as is documented by Giovanni Scriba (van Doosselaere, 2009)—was based on formal institutions such as the commenda. From a network-centric perspective, institutions of that kind (e.g., commenda, sea loans, etc.) can be considered to have been centralized, based on the nature in which commercial law was enforced. However, as indicated earlier, this stereotypical perspective on Genoese practice can be challenged, since contract enforcement was not solely reliant on the formal mechanisms but also included private order enforcement on family members residing in Genoa, should a violator not return with his profits. In comparison to the information that can be obtained from notary records, informal enforcement is by nature hardly documented in a systematic manner and anecdotal at best.
A key aspect of Greif’s work is that Genoese investors were very competitive and generally maintained business secrets, as supported by S. A. Epstein (1994). Genoese investors rarely shared information with fellow investors, so as not to provide them with any competitive advantage based on any experiences derived at high cost, and they were even imprecise about their travel agendas in commercial agreements (Greif, 2006). This assumption of noncommunication might lead one to doubt the effectiveness of network analysis for this situation. But in fact, the strong assumption of the absence of informal communication among Genoese can be practically considered unlikely for our purposes, since investors could still have benefited from lying to other investors. Doing so would have allowed them to direct their competitors toward merchants with a high risk of fraud or less profitable markets and thus obscure the availability of other, potentially more truthful, merchants for their own use. There are further indicators that trust-based informal mechanisms to regulate economic behavior were desirable among Genoese, inasmuch as enforcement via formal channels was cumbersome, expensive, and bore risk with regard to its outcome (S. A. Epstein, 1994). These considerations give a clear indication that formal mechanisms served as a “safety net,” should informal means fail to address violations, rather than providing the assurance that violations should be met with equally formal means. The combined use of informal and formal sanctions to enforce formal agreements was thus the norm (Greif, 2006). A further aspect that could have motivated informal enforcement mechanisms is the considerable fraction of foreigners who entered as merchants in documented contracts, providing around 18.3% of the capital (Gonzalez De Lara, 2008; Greif, 1993, 1994, 2006). Handling cheating of those would have hardly been tractable by purely formal means, unless threatened by informal sanctions. Other pointers to informal rules within the Genoese society are provided by the existence of trust poles, represented by the complementary role of private order enforcement of agreements, and the ongoing rivalry among the competing Maneciano and Carmandino clans (Greif, 2006) for power over Genoa. The shift toward an Alberghi-based system 7 after the height of the commenda (van Doosselaere, 2009) suggests that Genoese were aware of benefits of informal communication, even if it was limited, prior to the formation of subgroups that constituted the Alberghi system.
This leaves us with the question of just how much communication would have been necessary to sustain cooperation by informal means. The available evidence suggests that the essence of Genoese long-distance trade relationships was not a family matter, 8 but involved outsiders and was stratified into two distinct role types: stationary investors and traveling merchants. Scriba’s cartulary (Gonzalez De Lara, 2008; Greif, 1993, 1994, 2006) further reveals that investors and merchants existed in a ratio of about 1.57 merchants per investor, which we use to establish an initial multiagent model representing informal interactions among Genoese traders.
Constructing a Base Model of Genoese Trader Interaction
Traders involved in contractual relationships had an interest in benefiting from the knowledge of other investors’ experiences. To achieve that, they might request information about other merchants from fellow investors. However, sharing this information would be in conflict with the selfish interest of maintaining information asymmetry on merchant compliance. It is in a competitive investor’s best interest to have a competitor experience this first hand instead of providing the benefit of social learning. Here, the investor could follow one of two strategies:
On one hand, one could simply not share any information about his market partners. This would leave any requester with the clear understanding that one is not willing to share and does not care about the requester’s fate. We can call this strategy the “individualistic” strategy.
9
On the other hand, the requestee could generally provide information to other traders truthfully in a benevolent manner, but also sometimes lie about one’s knowledge about other merchants, such as suggesting a cheater as noncheater, or noncheater as a cheater, respectively. In both of these cases, following this information would disadvantage the requester, who would then either be inclined to employ a cheater or exclude a potentially truthful merchant from employment. We call this strategy the “selfish” strategy, driven not only by the intent to maximize one’s own benefit but also to minimize the requester's information gain by leaving the requester in doubt about the truthfulness of a response, thus potentially increasing one’s benefit by misleading competing investors.
When receiving requests on information about candidate merchants, the probability of reporting truthfully is thus of central relevance and influences the overall dynamics of the model. We make truthfulness a function of trust in others, which is adjusted and is self-reinforcing based on experience.
The base model consists of investors and merchants that exist in a fixed ratio of 1.57, as determined by Gonzalez de Lara (2008) based on Greif’s (1993, 1994, 2006) sources. During each round, investors send trade requests to chosen merchants who accept those and realize a market transaction which, with a parameterized probability, results in gain or loss. For this model, we assume that investors can detect cheating correctly and thus develop a memory about individual merchants. Prior to employing merchants, investors can also request information about candidate merchants from fellow investors, about whose truthfulness they likewise maintain memory. Those investors familiar with the merchant can then either respond to the requests in a truthful manner or lie about a merchant’s compliance behavior. To increase the likelihood that investors get advice from fellow traders who had experience with a given merchant, investors can continue to enquire about a merchant until they receive information or hit a specified maximum number of requests. A further constraint of the model is the limitation of memory. To simulate the notion of forgetting, investors can only maintain memory about a fixed number of merchants and their advising fellow investors following first-in first-out semantics, thus overwriting the oldest entry with the most recent experience once the maximum number of memory entries is reached. Repeated experience with an already memorized merchant leads to an update of the previous value and treats the updated entry as if it was newly acquired information. The choice of merchants is random in principle, but is filtered based on the memory the potentially employing investor maintains about their past conduct (both from direct experience and from other investors’ advice). If an investor fails to pick a trading partner within a specified number of rounds (e.g., because all of them were cheaters), he chooses the merchant he had the best experience with (based on his memory).
In our model, merchants are of simpler nature than investors. They are initialized either as cheaters or noncheaters and accept and process any trade request.
Figure 2 visualizes this base scenario emphasizing the different communication relationships.
Trader scenario.
Dynamic Adjustment of Trust
The central dynamic aspect of the model revolves around the truthfulness of investors to each other. As indicated before, traders adjust the truthfulness of their behavior in a dynamic fashion. To develop a memory about the cheating behavior of their fellow investors, any advice given (and used to guide decisions) can be monitored in order to establish whether the given advice was truthful. The testing of advice occurs with an initial probability, and its results are associated with the advisor, who, in the negative case, is remembered in order to be avoided in future enquiries about merchants.
For example, if investors decide to endow capital to what proves to be a noncooperative merchant against the better advice from fellow investors, this is accounted in favor of the advisor; giving positive advice on a cheater, in contrast, will be measured against the advisor.
The outcome of testing a fellow investor’s advice not only reflects the subjective individual reputation but also affects the “trust” the investor has toward other investors in general, if we interpret trust as an individual’s subjective probability that other individuals perform an action in compliance with the subject’s expectations (as inspired by Gambetta’s, 1988, trust definition 10 ). Noncompliance of an advisor thus results in a reduction of trust; compliance reinforces the “belief” in the system of informal cooperation among investors. Note that, apart from trust developed between individual investors, the exchange of trust information about participating merchants effectively represents a reputation mechanism, in that it allows one to enquire about merchants’ compliance (see the reputation understanding of Conte and Paolucci, 2002) and, respectively, to ascertain the extent to which agents comply with socially desirable conduct.
Besides the obvious benefit from truthful reporting of cheaters, investors would derive an additional benefit: Testing advice is associated with a monitoring cost and thus it is in their best interests to expend as little effort as possible to monitor investor relationships. From a nonutilitarian perspective, we suggest that trust is a suitable representation in itself, since any motivation to test compliance would be considered irrelevant, if not even disrespectful, for a likewise trusting advisor. We thus suggest that the probability of testing advice should be negatively correlated with trust and thus the probability of truthful reporting (p truthful). So in our model every trust increment (δtrust) also reduces the probability for testing advice (p test) by the same value. Note that since the informal relationship networks leave space with regard to their assumed structure, we also investigate the performance of the model over network structures other than just random assignment. At this stage, we set the maximum number of relationships to 40, which optimistically assumes that investors knew nearly a quarter of all other investors (recall that 18.3% of all investors were foreigners).
Table 2 summarizes the parameters we use for the base model, along with an initial set of values. The pseudocode for trader interaction is shown in Algorithm 1, and the handling of advice requests is described in Algorithm 2.
Simulation Parameters.
Relationship Networks Among Genoese Investors
Although previous research covers the formal relationships between Genoese traders, that is, investors and merchants in Genoa, it does not address the relationships among Genoese investors themselves. To derive candidate network structures for a hypothetical Genoese investor communication network, it is helpful to know that the 612 contracts collected by Giovanni Scriba noted around 180 investors (Greif, 1993, 2006). Of those contracts' capital, 90% came from just 37 noble families (Greif, 1993, 2006). Given this concentration of investor roles, network topologies other than a random network seem to be realistic alternatives for modeling informal communication flow. Individual investors were not likely to have the same level of connectedness, and the social structure suggested the existence of social hubs that accounted for the majority of information spread. Involving central hubs in the information flow would greatly improve the speed with which information was spread. Though such a configuration would be representative for today’s Internet structure, the society could alternatively have been stratified into communities, thus having various clusters of more densely connected individuals who shared information more intensely but had few links to other communities. In the context of the Genoese model, that could have been supported by the importance of clan structures that were characteristic for the Genoese society (Greif, 2006). To test those possible approximations for social structures, we thus tested four different network topologies. In our arrangement, advisor relationships between investors can be based on
fixed randomly assigned relationships (FXD), which can be represented as directed random networks; fixed randomly assigned mutual relationships (FXDM), which are effectively undirected random networks; small-world networks (WS), which we model using the Watts–Strogatz algorithm (Watts & Strogatz, 1998); and finally, scale-free networks (BA), which are represented using the Barabasi–Albert (Barabasi & Albert, 1999) algorithm.
For the fixed random assignment (FXD), we generated an advisor network based on random association with a fixed number of relationships (maxRelationships). In the case of fixed random mutual relationships (FXDM), the random relationships are ensured to be of mutual nature; both parties have each other in their respective set of advisors. For both FXD and FXDM, each node in the network has a fixed number of relationships (maxRelationships). Given the lack of exact information about informal relationships, the WS is used to approximate the notion of communities of tightly connected individuals. Those communities are interlinked with significantly fewer interconnections, establishing numerous local social clusters. Necessary parameters for WS include the mean degree of the network (which we set to maxRelationships) and the probability to rewire initial randomly assigned links. Given the limited knowledge about the actual structure and after pretesting with values ranging from 0.1 to 0.9, we set this rewiring probability to .5, creating an optimistic tight linkage between different groups. BA are suitable for representing relationship networks that have a strong variation in connectivity. They generate power–law distributions by linking nodes based on preferential attachment, resulting in a wide degree distribution with a few strongly connected “social hubs,” a structure that has been shown to closely resemble link structures of the web (Clauset, Shalizi, & Newman, 2009). The maximum number of relationships for this network is also set to maxRelationships.
In the current model, the notion of reciprocity is implicit by keeping a memory about the experience with other investors; there is no explicit notion of “tit for tat” or a strategic choice of advisors. Investors decide individually whether they reply truthfully according to their situational value of p truthful, a probability that is adjusted based on previous interactions (see Dynamic Adjustment of Trust subsection).
In addition to those relationship networks, we analyze the society’s ability to ban cheaters in the absence of any direct communication among investors, which is equivalent to Greif’s individualistic strategy (Greif, 2006; which we denote as NoComm), in order to discriminate (a) performance benefits for cheater exclusion gained by informal communication from (b) the effect of individually memorizing cheaters.
Recalling the objective to identify under which conditions informal trust relationships would have been sufficient to control cheater behavior, the dependent variable is the fraction of reemployed potentially cheating merchants. Interesting characteristics of this investigation include the effectiveness of a given relationship network for isolating cheaters, thus the extent to which cheaters are excluded from the system. Beyond that we are interested in the efficiency of that process, thus identifying how fast this occurs; lower numbers of rounds imply higher efficiency.
Those two measures are operationalized as follows: We analyze throughout all execution rounds the fraction of employed potentially cheating merchants (q(r), where r represents the simulation round). We assume that the number of employed potential cheaters converges toward a relatively stable level without any further change, an assumption that has been established based on test runs. The effectivity describes the difference between the initial level of employed cheaters and the final level of employed cheaters. The final level of employed cheaters is a function of the trust level. The final cheater level is thus the mean of the value for all rounds that follow the establishment of a stable trust level. The trust level is considered stable if it is within a tolerance (of 0.001) of the maximum trust level. 11 We thus operationalize the effectivity as the fraction of potential cheaters in the first round (q(1)) reduced by the mean of all quota values for potentially cheating merchants following the establishment of a stable trust level, divided by q(1). Effectivity values thus lie between 0 and 1, where 1 indicates full banning of cheaters from business transactions and 0 indicates the failure to do so. The efficiency is directly derived from the number of execution rounds necessary to reach a stable trust level r stableTrust, which we interpret as being within a tolerance zone δtolerance of the maximum trust level (with a δtolerance set to 0.001), where a lower number of rounds indicates better performance. Efficiency is normalized with respect to the number of simulation execution rounds r max, with 1 representing highest efficiency (i.e., immediate convergence to final level) and 0 worst (i.e., no convergence to a final level at the end of the simulation execution rounds).
The necessary rounds to reach a stable trust level r
stableTrust can be defined as
The efficiency is then
Using r
stableTrust (the number of rounds until a stable trust level is established) and q(r) to indicate the fraction of employed potential cheaters for a given round r, we can express effectivity as:
For our initial set of simulations we parameterize the model as follows. We assume that 40% of all merchants potentially cheat when employed (cheaterQuota). As the evaluation of tipping points is of core interest when exploring trust in informal mechanisms or reliance on formal mechanisms, the initial value for p truthful for each individual is drawn from a random distribution with a mean of 0.5 and a standard deviation of 0.1. All further uses of p truthful in this work refer to the mean value of the random value assignment. p test is initialized with 0.5 without randomization; investors initially throw a dice in order to decide whether or not to test advice. As a result of pretests, values for maxRequests and memoryEntries have been set to 10 and 40, respectively. We analyze the results of 30 simulation runs of 10,000 execution rounds for each tested parameter combination. The number of executions is chosen in order to permit the simulation runs to converge to a stable trust level that allows comparative analysis of different configurations. The code for this simulation is available under Frantz (2013) or on request to the authors.
Simulation Results
The results for p truthful = 0.5 (Figure 3; Table 3) show very limited effectiveness for fixed random directed (FXD) and fixed random undirected networks (FXDM); only network types with wider degree distribution spread, such as WS and BA, show a meaningful reduction of cheaters. However, for WS, sufficiently high trust levels can only be established beyond 6,000 rounds. Effectivity for FXD and FXDM are limited to around 0.25; they thus only manage to prevent around one quarter of the cheaters from being reemployed. Comparing those results against the scenario without any communication (see results for NoComm in Table 3) supports the idea that informal communication across FXD and FXDM networks would only have marginal impact for an initial truthful reporting level of 0.5. Only nodes in WS and BA manage to collectively remove more than 30% of cheaters from the system, but, at least for the BA configuration, require many more rounds to do so—more than 6,000 rounds. p truthful = 0.5 represents the tipping point for the simulation outcome, and this is reflected in the fact that for this value of p truthful, BA displays a relatively large standard deviation for efficiency, capturing both cases in which cooperation was established, as well as cases in which investors did not converge toward cooperation. 12 The BA network was particularly sensitive to the randomized assignment, because the hub’s influence promotes the effective spread of the social hubs’ attitudes toward truthful reporting. Note that Figure 3, in contrast to the aggregate information in Table 3, only shows the output of a single representative simulation run. It further displays mean trust levels across the agents for the different networks, all of which originate from a mean around 0.5.
Single exemplary simulation run with p truthful drawn from a random number distribution with mean 0.5 (standard deviation 0.1).
Simulation Results for Informal Communication Among Genoese Traders.
Note. FXD = fixed random assignments; FXDM = fixed random mutual relationships; WS = small-world networks; BA = scale-free networks; M = Mean.
30 Runs per configuration; maxRequests = 10.
An initial trust level of 0.5 is clearly not sufficient to reliably protect against cheaters. So we adjusted the initial trust level to distributions with means of 0.51, 0.55, 0.6, 0.7, 0.8, and 0.9 to observe the sensitivity of the model. Note that higher trust levels appear unrealistic for the represented Genoese societies. Nevertheless, we include those for the sake of completeness of our analysis. The results sets are shown in Table 3.
As expected, higher levels of initial truthfulness improve the extent to which cheaters can be banned from future transactions. However, looking at the cheating levels, even for the optimistic assumption that 60% of Genoese investors would trust each other sufficiently, at best 45% of the cheaters would be removed from the system. Only if at least 70% of all Genoese investors “invest” in trust, will more than 50% of cheaters be collectively identified for most network types (except BA) and banned from future business transactions. On the whole, apart from the direct observation of the target variables, the results bear further interesting observations.
For network types with a wide distribution range (WS and BA), we observe a strong initial increase in effectivity, outperforming network types with fixed degrees. However, for higher trust levels we can observe that the “fixed” network types overtake WS and BA in their performance (with effectivity breaking even at a truthfulness level of 0.6), and eventually outperform WS and BA both with regard to effectivity and efficiency. At this stage, it is important to emphasize that efficiency values taken alone are hardly meaningful, as they only describe how quickly a stable level of truthfulness is reached, whatever its level. Only in combination with effectivity, the extent of truthfulness, do they gain meaning.
Looking at the different performances of network types, we see the reason for this behavior in the nature of the different network types. While higher degrees of connectedness of nodes initially facilitate a rapid spread within clusters of connected nodes, shaping “trust clusters,” they are not able to sufficiently penetrate to weakly connected nodes with cheater information. This limits the ability of the overall system to reliably bring cheaters to the investors’ “collective consciousness.” Figure 4 helps to analyze the effect of manipulating the initial trust probability p truthful. With increasing p truthful we can observe the relative changes in effectivity and efficiency for the different network types across the tested range of initial levels of truthfulness (solid lines indicate effectivity measures; dashed lines denote efficiency).
Effectivity and efficiency for different initial levels of truthfulness (p truthful) for maxRequests = 10 (mean values of 30 simulation runs per configuration).
For higher initial levels of truthfulness, we can generally observe that efficiency increases are more significant; investors require less time to share information about cheaters. This overshadows the continued, but more moderate, and in most cases near linear increase in effectivity. For higher initial levels of truthfulness (> 0.6), the core benefit for Genoese investors would have been the extent to which cheaters were identified and excluded from business activity, and not so much how quickly this occurs.
Apart from trust, our model has further parameters that allow exploration. Given our focus on the comparative view of different information transmission systems, at this stage we provide an overview of the effects caused by variation of maxRequests, that is, the maximum number of advice requests that can be raised by an investor. Increasing this parameter to the value 15 improves the effectivity from 0.05 to 0.1; increasing it to 20 requests enables an additional improvement by 0.05. Thus, higher numbers of requests do not bear significant benefit for the spreading of cheater information. Though effectivity increases with increasing numbers of advice requests, efficiency is not affected and remains stable. The relative performance among different network types is likewise unchanged.
At this stage, we have developed an understanding about how informal communication could have benefited Genoese investors to identify cheaters. However, we realize that banning the majority of cheating merchants required considerable initial trust.
Although the presented model has limitations with regard to a complete specification of parameters, it provides us with a template that allows one to incorporate the differences of the Maghribian information network into the base model to at least allow a comparative perspective. Thus, applying the Maghribi communication model to the Genoese society, as it has been laid out up to this stage, facilitates a better understanding of the dynamics of the different information transmission mechanisms involved in the respective societies. This forms the focus of the next section.
Applying the Maghribi Information Transmission Network to the Genoese Case
The Genoese model that we explored up to this stage can be described as the pull model of communication. It requires individuals to request information from fellow investors and operates in a reactive manner. The Maghribis, in contrast, operated in a proactive manner underlying a different motivation. We can describe their way of communicating as operating according to a push model. This makes it desirable to examine whether a communication model using the “push approach” could have enabled Genoese investors to operate successfully without commenda contracts.
Observant Maghribi traders proactively sent information to fellow traders with whom they shared ṣuḥba relationships (see Goldberg, 2012, and discussed further below), so as not only to report cheaters but also to inform their fellow traders about market prices and trade opportunities, honoring the importance of a lasting relationship. Preventing observed cheaters from being reemployed under any circumstances was thus a by-product of a nontransitive relationship for which internalized trust based on reciprocity was implicit and constitutive. The mere adoption of the institution “ṣuḥba” required participants to truly believe in it and carefully maintain each of their different ṣuḥba relationships. A single misstep in one of his relations could cause a partner to be expelled from economic participation, inasmuch as relationships could be dissolved unilaterally (and in absence of the partner). Furthermore, based on the strong interconnectedness, information would reach into the social environment and lead to immediate canceling of other relationships that the merchant enjoyed as a member of the aṣḥābunā (“our colleagues” or “our associates” [Goldberg, 2012]), a network of tightly interlinked ṣuḥba relationships. The central motivation of a Maghribi that had successfully entered the aṣḥābunā would thus not be centered on exclusion of cheaters but preserving one’s own belonging (and thus reputation) to the system of relationships the aṣḥābunā represented. Compliant behavior and proactive maintenance of relationships was the norm; proactive communication acted as an instrument to assure partners of continuous commitment to the relationship which entailed the sharing of information about others’ deviant behavior, and also bore the chance to increase one’s own reputation.
An important aspect that is not immediately reflected in our model of information transmission mechanisms but is reflected in the proactive and reciprocal nature of their general communication, is the fact that the Maghribi traders did not differentiate their roles into either provider of goods (similar to the Genoese investor) or selling agents (equivalent to Genoese merchants). They saw both functions integrated as part of their status as a full trader. However, following on from their apprenticeship and an increasing number of connections and affluence, traders incrementally shifted from handling and selling other traders’ goods to more stationary activities centered on the sending of goods to overseas associates who then realized the actual transactions on the sender’s behalf. Nevertheless, even if the role understanding changed over the period of a trader’s career, the obligations that bound one to the aṣḥābunā remained the same: Whenever associates sent goods, it was the obligation to accept and store those goods, independent of whether or not the receiver was supposed to sell those. Also, selling those goods would not be paid for as a service commission; instead, a selling trader gained a reciprocal favor against the associate to handle and sell one’s own goods in return. The extent of favors, or better, reciprocal obligations “earned” by handling goods, depended on the relative status difference of both trade partners (Goldberg, 2012). To warrant the different nature of communication behavior, the Maghribi model can thus be interpreted as a push model in which individuals proactively communicate their experience to partners with whom they share reciprocal relationships. This is in contrast to the more reactive pull approach modeled for the Genoese case.
At this stage of our discussion, we need to pause and discuss the classification of the ṣuḥba with regard to its formality and justify the use of its characteristics in our model: Was the ṣuḥba relationship formal or informal? In his work, Greif discusses institutions Maghribi traders could rely on for the enforcement of compliance (Greif, 1989). For his game-theoretical scenario, however, Greif is not explicit about the assumed institution, but in fact implies an inherently informal communication and reciprocal fulfillment based on cultural and contextual traits (Maghribis as Jews in a Muslim environment). Another more recent attempt to analyze the Geniza by Goldberg (2012) suggests, that for the analyzed sample, the majority of relationships (67%) represented ṣuḥba relationships, while only very few (3.5%) were explicitly informal mu’āmala relationships. However, if those were the relationships Greif’s analysis relied on, they would hardly be represented in the Geniza letters Greif bases his own analysis on, which would challenge his suggestion of a purely informal institution. Goldberg (2012) supports the suggestion that the ṣuḥba relationship has historically been misinterpreted as “informal” by central figures such as Goitein and Udovitch, despite its ritual characteristics but generally on account of its lack of explicit codification. An indicator for this presumed misinterpretation is Greif’s effort to differentiate Maghribian institutions sharply from the Italian-style commenda (Maghribi equivalence: “eseq” in Jewish law; “qirad algoyim” in Arab [Greif, 1989]) and by implication Genoese business practices, which inherently relied on codified contracts, thereby bearing a very sharp and unambiguous characterization as “formal.” For the modeling metaphor applied in our work here, we see the ṣuḥba relationship as prototypical for the Maghribi relationship characteristics Greif describes in his work (unwritten reciprocal relationship). Whether one concludes that the ṣuḥba was formal or informal, its reinterpretation does not affect the simulation model per se.
To accommodate the push mechanism in our model for the Genoese trading community (in order to see how Genoese could have addressed the problem of cheating if using Maghribi transmission principles), we modify the base model and make investors share their respective trader experiences with related investors immediately. However, advice is not blindly accepted. Instead, receivers make the incorporation of shared information dependent on whether they know the sending investor, and if so, if he is currently classified as a good advisor. Maghribis would not share this information with traders they did not enjoy a reciprocal relationship with. For our model, this has two consequences. For one thing, if the sending investor is unknown, advice is simply ignored. As a second consideration associated with the reciprocal nature of communication networks, the fixed directed network (FXD) represents an unrealistic communication structure, since an investor could have an advisor relationship with another investor who, in turn, would not necessarily recognize the former as related. We will thus exclude the network type FXD from further discussion when applying Maghribi-style push communication to the Genoese model.
Applying this alternative communication strategy, we can observe a strong reaction for an initial trust value of 0.5 (p truthful; see Figure 5 and Table 4).
Single exemplary simulation run using Maghribi-style communication for p truthful = 0.5.
Simulation Results for Informal Communication for Genoese Traders With Proactive Communication.
Note. FXD = fixed random assignments; FXDM = fixed random mutual relationships; WS = small-world networks; BA = scale-free networks; M = Mean.
30 Runs per configuration; maxRequests = 10.
For values above the trust tipping level of 0.5 (see Table 4), the spreading of trust information is much more rapid (near immediate) among connected investors. Values below 0.5 likewise cause the collapse of cooperation by rapid loss of overall trust (in this context note the higher standard deviation for p truthful = 0.5, which is caused by simulation runs that result in noncooperation, i.e., collapse of trust). If establishing cooperation, even for p truthful = 0.5, between 55% and 68% of cheaters are removed from the system. Efficiency values are likewise high, consistently lying above 0.7, with the exception of the BA, which is caused by the strongly varying connectedness in different simulation runs and thus stronger variation in trust levels (effectivity) and efficiency (see standard deviation). For p truthful > 0.5, nearly all network types consistently remove 70 or more percentage of cheaters from future transaction. Thus, proactive communication clearly outperforms a reactive variant model at the price of a strong belief in a mutual relationship.
The ability to exclude cheaters is constrained by two aspects, first by the connectivity in different networks. We can observe for WS and BA that those forms of network connectivity are (though potentially only slightly) more limited with regard to information spreading, reaching their maximum at around 78% (small-world) and, even less successfully in the case of the unequally connected BA, and so fail to sufficiently supply remote investors with needed information. Note that we parameterized the small-world Watts–Strogatz algorithm with an optimistic rewiring probability of .5, leading to a strong interconnection among different communities. Possibly more realistic lower levels of connectivity among different communities (which allows better alignment with the strongly clustered clan structure of Genoa) could be considered by reducing the rewiring probability. For a rewiring probability of .3, the maximum effectivity level lies at around .75.
In general, in order to investigate network structures for the different societies, we must rely primarily on historical narrative accounts. As mentioned previously, 37% of Genoese investors lent around 90% of the capital, which supports the idea of a wide network distribution. This is further supported by van Doosselaere’s (2009) analysis of the commenda network structure, which suggests a power law distribution for investor–merchant relationships. A BA can thus be a possible representation for the informal relationships in that society. In this model, we are looking at relationships among investors, not investors and merchants. Moreover, those relationships are considered to be of informal nature; whereas van Doosselaere (2009) only analyzes formal, that is, documented, relationships.
For the Maghribian side of things, we can refer to Goldberg’s statement concerning the relationship of two prominent Maghribis, Ibn ’Awkal, who enjoyed around 150 relationships, as well as Nahray ibn Nissīm, a participant in over 400 relationships. Although accounts are incomplete and bear a bias toward traders residing in Fustat (85% of letters have Fustat as their destination), which was the capital of the Fatimid empire as well as the location of the Ben Ezra Synagogue (which was the storage location of the Geniza), Goldberg states:
The network of aṣḥābunā in the Ibn ‘Awkal and Nahray groups should thus not be understood as perfectly connected, monolithic, or composed of equally strong ties. [ … ] [The network had] enough connections to be mentioned as carrying out commercial services or having a partnership with more than one other merchant in the network. The network was always in flux, both through addition, retirement, or death of individual merchants [ … ]. (Goldberg, 2012)
Based on literature accounts such as Goldberg (2012), relationship networks of young traders had a strong local focus and only expanded over some time, as they established relationships with more remote traders. Although challenging to confirm, it is likely that young Maghribian merchants were interconnected by a WS covering their trade locations, which only gradually expanded over time. By filtering unskilled trade apprentices, full traders enjoyed lasting remote connections with particularly successful ones (such as Ibn ’Awkal and Nahray ibn Nissīm) at their center. Taking the assumption of the carefully groomed trader community of the Maghribis, a small-world-type relationship network might have been more likely in that case. For the Genoese, however, given the concentration on some 90% of capital on 37 noble families (of around 180), the BA representation is a good candidate. For the Maghribian side, a secret of their success in banning cheaters collectively, beyond their belief-based proactive communication behavior, may also be partly attributed to the structure of their relationship network. The results for WS in our simulation support that possibility.
Synthesis
In this article, we have looked at the 12th-century Genoa trading sphere that relied on formal institutional mechanisms to enforce the cooperative behavior which Avner Greif contrasted to the Maghribi traders. Indeed, in contrast to the Genoese traders, the North African Maghribi traders group that operated during a similar time frame hardly used formal mechanisms to govern economic behavior but solely relied on informal reciprocal trust relationships. The nature of this problem is ultimately an instance of the investigation into the question concerning the interrelation between culture (and trust as culture-dependent characteristic) and economic performance (e.g., Tabellini, 2005; Knack, 2001). Greif’s game-theoretical approach to this problem relied on strong assumptions contrasting the information-sharing Maghribis against our nonsharing Genoese.
In our work, we have weakened some of those strict assumptions (which we justify by appealing to more detailed historical analyses) and translated the information transmission mechanism into a multiagent simulation model which made this scenario model accessible beyond the more confined comparative-static analysis. We tested the ability to communicate information identifying potential cheaters in a peer-to-peer fashion across different network types, which served as potential candidates to represent communication relationships within the Genoese community. On the basis of our simulation studies, we were able to support Greif’s assumption that the limited trust among Genoese was insufficient to address the cheating problem in a purely informal manner. As a second step, we incorporated Maghribi-style communication based on reciprocal relationships to test its impact on an otherwise unchanged network structure. This yielded significantly better performance and confirmed Greif’s finding that Maghribis could sustain cooperation based on informal means. As a consequence, the contribution of our work lies in the development of a more realistic and detailed model based on literature accounts, including both quantified information and anecdotal evidence that goes beyond the dichotomous metaphor of communication versus noncommunication to a more differentiated view on possible communication patterns. We see this modeling approach in line with Edmonds and Moss’ (2005) demand for a more descriptive modeling approach incorporating a wider range of potentially weaker sources.
With the departure from the rigid game-theoretical comparative approach, our more differentiated perspective on both trader communities opens up further research directions. The information transmission mechanisms analyzed in this work depend on culture-specific institutions, such as the ṣuḥba, and the beliefs that accompany its success. We suggest that a key difference lies within the sophisticated Maghribian apprenticeship system, which required a strong level of social investment (both affording time and building reputation). For future research, we thus intend to establish an endogenous perspective of the respective societies, incorporating more of the peculiarities both societies exhibited (such as group internal power structures, incomplete information about cheaters [see, e.g., Aydinonat, 2006]), and thus moving beyond the comparative perspective assumed here.
Concluding Remarks
To conclude, we would like to provide some future directions in which this kind of agent-based institutional modeling can be taken. This work investigates institutions with respect to their central role in the development and prosperity of societies. In this respect, it follows the main theme of New Institutional Economics (NIE) (Eggertsson, 2013; North, 1990). Central to this approach is the exploration of the characteristics of informal and formal institutions, and more so, the identification of drivers that led to the choice of whichever institution type and the path dependence of this decision. Our approach explores this by using the Maghribi and Genoese traders as an example, an example that is particularly appealing because of the comparative perspective that those two societies offer. In this respect, a key characteristic ascribed to informal institutions is the initially cheap distributed enforcement, facilitated by small group size and relatively closed groups (e.g., high exit cost for interdependent individuals). However, NIE researchers (Acemoglu & Robinson, 2012; North, 2005) have associated growth and economic development with comprehensive and explicit institutional frameworks that centralize enforcement of cooperation by delegating it to enforcement specialists. It is argued that this shift to formal institutions enables open societies to flourish and to overcome the scalability limitations of social enforcement mechanisms, and in particular, the limited control over second-order violators, that is, individuals who do not participate in punishing violators, thus not bearing their part of the enforcement cost and reducing the effectivity of the institution. These notions lead to the question as to why individuals cooperate in the first place. Retracing the onset of the formation of social institutions, Bowles and Gintis (2004) take an evolutionary approach and postulate the social structure of Pleistoscene society. They suggest the existence of strong reciprocators that have limited concern for their selfish benefit but assure punishment of noncompliant behavior, and show that their coercive force increases group survival while bearing interesting effects on limiting the average group size. A different take to explain the existence of seemingly altruistic enforcers has been taken by de Quervain et al. (2005). Their experiments challenge the rational account on human behavior by suggesting that punishment of economic cheaters can elicit positive rewards (e.g., satisfaction) that overrule the cost associated with enforcement, which is otherwise commonly seen as the central deterrent to communal institutional enforcement in cooperation games, and a central problem when involving public goods (Ostrom, 2005).
The work presented here likewise concentrates on the role of informal institutions in maintaining cooperation and presents an example of introducing indirect reciprocity (and thus a shift from mere personal to communal enforcement [Greif, 1992]) into a model of an individualistic Genoese trader society. But in the absence of impartial norm enforcers (such as Bowles and Gintis’ strong reciprocators), the competitive nature and thus limited trust among Genoese traders would have prevented the effectiveness of this informal institution, leaving the effectiveness of institutions only reliant on trust.
Measuring the penetration of cooperative behavior throughout the Genoese relationship network opens avenues for further exploration. At the current stage, we assume static relationships. Once established at the onset of the simulation, relationships do not change. Insofar as we know little about the Genoese informal relationship network, we are limited in the degree to which we can characterize its dynamics. Related efforts by Bravo, Squazzoni, and Boero (2012) and Skyrms and Pemantle (2000) explore experimentally the importance of trust-based partner selection by introducing a dynamic adaptation of links which proves to be effective in constraining freeriding, an aspect that could well be considered realistic for the Genoese society. A different approach is chosen by Zschache (2012), who introduces qualities of relationships and suggests the endogenous change of network structures based on social comparisons by observing peers’ performance in public goods games.
Potential refinement could also include stronger consideration of asymmetric impact from experienced compliant and violating behavior, that is, stronger reaction to cheating behavior in expectation of otherwise compliant behavior. However, at this stage, this would introduce further assumptions about the Genoese investor society. To that extent, could we assume an expectation of truthful reporting of cheaters in a society that is known for its secrecy, or the opposite? Howsoever, on an individual level, one can probably assume that previous experience with another peer would affect the impact of future interactions.
Looking at the institution itself, a practical challenge considered crucial when representing social institutions is the modeling of second-order enforcement, that is, assuring the continuous sanctioning of cheaters by norm participants to sustain the norm’s salience. We think that approaching this aspect by rebuilding institutions in an endogenous fashion bears the strongest potential to explore and to understand why the Maghribian society was able to rely on informal institutions “from the inside.” Important differing characteristics include the exploration of the social stratification between investor and merchant among Genoese and the unified role understanding among the Maghribis, but agent-based modeling can enable us to dive deeper and include further details that Goitein, Greif, Goldberg, and others have carefully extracted from historic accounts to retrace the institutional paths of both societies:
The Maghribis’ operations were constrained by the rulers of the Fatimid Empire, such as the limited goods they were allowed to transport, 13 but as long as they were operating within its realm, they enjoyed considerable autonomy in building maintaining mutual trade relationships within their own community, backed by an otherwise culturally and institutionally homogeneous environment. Their fate can only be traced into the 13th century, when the Egyptian rulers forced them to cede their trade (Greif, 1993), upon which they dispersed into their own Jewish communities.
For the Genoese, in contrast, their open society (the city grew from 30,000 to 100,000 between 1200 and 1300 AD [Greif, 1994]) was not embedded in a rule system that had significant reach outside its city boundaries. Given the constant influx of investors and opportunists, the Genoese had no choice but to opt for formal institutions in order to maintain cooperation in early historic trading—and establish the prototype of an institutional environment now thought to be constitutive for the prosperity of modern societies.
Footnotes
Acknowledgments
We wish to acknowledge the input from various parties that greatly improved the article. Important to mention in this context are the three anonymous reviewers who offered constructive advice, both including technical and narrative aspects which enhanced the final version of the article. We would like to thank Keith Rogers for introducing us to the medieval trading scenario. The authors would further like to thank Bastin Tony Roy Savarimuthu and Marzieh Jahanbazi for their feedback on the initial draft of the article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.