INFO709 Final Paper
Prediction Markets as Complex Adaptive Systems
Aarathi Parameswaran
Fall Semester 2025

1 Prediction Markets

Prediction markets are platforms or institutional mechanisms in which participants buy and sell contracts tied to the outcomes of future events. These markets are designed to elicit and aggregate dispersed information from participants, thereby producing collective forecasts. Also referred to as information markets or idea futures [21], prediction markets operate by allowing traders to express beliefs through prices. Contracts typically pay a fixed amount if a specified outcome occurs. The market price of a contract is interpreted as the probability of the corresponding event, reflecting the aggregate belief of market participants. For example, if a contract paying 1 USD conditional on an event trades at 0.60 USD, the market-implied probability of that event is 60%. In this sense, prediction market prices function as market-aggregated probabilistic forecasts.

Although prediction markets have existed and been studied for several decades (like the Iowa Electronic Markets or IEM), their prominence has increased markedly over the past five years. The growth of online market infrastructures, digital communities, and social media has enabled the growth of large-scale platforms such as Polymarket, Kalshi, Manifold and PredictIt, renewing interest in prediction markets as forecasting tools [20]. These platforms frequently focus on highly salient and polarizing topics, including elections, geopolitical conflicts, sports, and cryptocurrency, thereby attracting substantial public attention [7]. During the 2024 U.S. election cycle, for instance, nearly USD 2.4 billion was wagered across online prediction markets between September 1 and Election Day [6].

Because these markets operate online, they generate rich, high-frequency data on how individuals perceive uncertain events. Empirical studies typically evaluate their performance by comparing market-implied probabilities with realized outcomes, measuring forecasting accuracy and efficiency. While earlier prediction markets were largely confined to controlled academic experiments, contemporary platforms increasingly shape public expectations, media narratives, and political discourse.

1.0.1 How predictive are these markets?

Prediction markets have long been promoted as superior forecasting tools, often claimed to outperform traditional information aggregation methods such as polling [5]. Contemporary online platforms make particularly strong claims about their predictive power, frequently citing anecdotal evidence from the 2024 U.S. elections. During this period, market-implied probabilities on platforms such as Polymarket and Kalshi diverged substantially from polling estimates. Online prediction markets are also designed to maintain anonymity since they use blockchain technology, as opposed to other conventional forms [5]. In response, these platforms have increasingly framed themselves as “truth-telling” oracles that reveal underlying political realities more accurately than conventional polling methods.

From Polymarket’s documentation [16]:
Polymarket is the world’s largest prediction market, allowing you to stay informed and profit from your knowledge by betting on future events across various topics. Studies show prediction markets are often more accurate than pundits because they combine news, polls, and expert opinions into a single value that represents the market’s view of an event’s odds. Our markets reflect accurate, unbiased, and real-time probabilities for the events that matter most to you. Markets seek truth.

As a result, prediction market prices are often treated as credible reflections of public expectations. However, this perception persists despite evidence of substantial volatility in market prices, at times driven by the actions of a small number of large traders. Such dynamics raise questions about whether observed prices reflect broadly aggregated beliefs or instead the influence of concentrated capital.

The purported truth-telling capacity of prediction markets is commonly justified through the logic of the “wisdom of crowds”, under the assumption that participants possess diverse information and are incentivized to act on it, market prices should efficiently aggregate dispersed knowledge. This leads to a central empirical question: when large financial stakes and heightened public attention are present, do prediction markets in fact aggregate political information accurately and efficiently?

Empirical evidence reveals substantial variation across these different online platforms. A recent study by Clinton and Huang [6] analyzed nearly 2,500 political prediction markets across four platforms during the final five weeks of the 2024 U.S. presidential campaign, finding that while 93% of PredictIt markets correctly predicted outcomes better than chance, accuracy fell to 78% on Kalshi and 67% on Polymarket. Even the supposedly most accurate markets showed little evidence of efficiency, as prices for identical contracts diverged across exchanges, daily price changes were weakly or negatively correlated, and arbitrage opportunities also persisted. This variation across platforms operating under similar trading mechanisms, during the same information-abundant period suggests that accuracy is not guaranteed by market design alone but depends on other institutional, structural, and behavioral factors.

1.1 Economic Models of Prediction Markets

Economic studies of prediction markets have primarily operated within the neoclassical framework, grounding analysis in rational expectations equilibrium and the Efficient Market Hypothesis. Wolfers and Zitzewitz’s [21] foundational survey exemplifies this approach, arguing that prediction markets efficiently aggregate dispersed information through trading, with contract prices revealing market-aggregated forecasts of future events. They state: “If a prediction market is efficient, then the prices of these contracts perfectly aggregate dispersed information about the probability of each candidate being elected.” Under this logic, prices should reflect all available information, and agents should not be able to systematically profit from public information which is the hallmark of weak-form market efficiency.

The theoretical foundations rely on standard equilibrium analysis with homogeneous agent assumptions. Grossman [8] demonstrates that under CARA (Constant Absolute Risk Aversion) utility functions, where risk aversion remains constant regardless of wealth, and normally distributed private signals which traders draw from, equilibrium futures prices perfectly summarize all trader information. Wolfers and Zitzewitz extend this result, showing that under logarithmic utility, prediction market prices equal the mean belief among traders (or a wealth-weighted average if beliefs and wealth are correlated). While these models allow for heterogeneity in beliefs and wealth, they treat such variation as noise around a rational core and maintain core neoclassical assumptions: fixed preferences, rational expectations, and convergence to equilibrium.

However, neoclassical theory faces fundamental challenges in explaining prediction market behavior. The Milgrom-Stokey “no-trade theorem” [12] proves that under rational expectations and common priors, no trade should occur as each trader should infer that anyone willing to trade possesses superior information and therefore refuse to participate. Wolfers and Zitzewitz acknowledge this paradox explicitly: “Explaining why there is any trade in prediction markets remains an important open theoretical question.” Their proposed resolution attributes this to uninformed outsiders with hedging or entertainment motives, or to manipulators attempting to influence prices, essentially admitting that the core rational-agent framework cannot explain the market activity it seeks to model.

A related theoretical difficulty emerges from Grossman and Stiglitz [9], who demonstrate that prices cannot be perfectly efficient, challenging the Efficient Market Hypothesis within the neoclassical framework. If prices fully reflected all information, no trader would incur the cost of gathering information, undermining the very mechanism by which information enters prices. They conclude that equilibrium must feature sufficient pricing inefficiency to reward information discovery which contradicts the strong form of the Efficient Market Hypothesis while remaining necessary for market functioning.

Empirically, Wolfers and Zitzewitz document that prediction markets often achieve impressive accuracy, outperforming expert forecasts and traditional polls in domains ranging from elections to economic indicators. Their analysis of the Iowa Electronic Markets shows average forecasting errors of 1.6 percentage points for presidential elections compared to 1.9 for Gallup polls, while market-based macroeconomic forecasts outperform survey-based expert consensus. Markets also appear to satisfy weak-form efficiency, with prices following random walks and responding rapidly to new information. Yet the same survey documents systematic anomalies that the theoretical framework struggles to explain, like the favorite-longshot bias (where low-probability events are systematically overpriced), laboratory experiments showing information aggregation failures and bubbles, and context-dependent manipulation resistance. Wolfers and Zitzewitz note these anomalies but treat them as scattered exceptions rather than as systematic patterns requiring theoretical explanation.

The neoclassical framework thus establishes prediction markets as information aggregation mechanisms under ideal conditions but cannot explain when or why these conditions break down. The no-trade theorem remains unresolved, the tension between efficiency and information discovery persists, and the variation in accuracy across platforms and contexts lacks a coherent theoretical account within the equilibrium paradigm. These gaps motivate an alternative approach that treats prediction markets not as equilibrium-solving mechanisms but as evolving complex adaptive systems.

1.2 Prediction Markets as Complex Adaptive Systems

Prediction markets exhibit all defining properties of complex adaptive systems, as described by Holland and Miller [10]:

Many economic systems can be classified as complex adaptive systems. Such a system is complex in a special sense: (i) it consists of a network of interacting agents (processes, elements); (ii) it exhibits a dynamic, aggregate behavior that emerges from the individual activities of the agents; and (iii) its aggregate behavior can be described without detailed knowledge of the behavior of the individual agents. An agent in such a system is adaptive if it satisfies an additional pair of criteria: the actions of the agent in its environment can be assigned a value (performance, utility, payoff, fitness or the like); and the agent behaves so as to increase this value over time. A complex adaptive system, then, is a complex system containing adaptive agents, networked so that the environment of each adaptive agent includes other agents in the system.

Prediction markets satisfy all five criteria as described by Holland and Miller: (i) there are networked agents as traders interact through price signals, ideology and identity, and social media and buy and sell contracts amongst each other (ii) there is emergent aggregate behavior as market prices arise from trading between agents rather than some central price calculation (iii) the macrobehavior is describable without microlevel details, as the system emerges into accuracy regimes without having to track every individual trade (iv) there are value-assigned actions as traders earn profits/losses based on forecast accuracy, which would be their payoff (v) adaptive behavior exists as agents adjust strategies based on past performance. Critically, Holland and Miller emphasize that adaptive agents exist in environments created by other adaptive agents, generating feedback loops absent in equilibrium systems. In prediction markets, this manifests too with different strategies, with better strategies accumulating more information and capital.

Participants in such markets involve heterogeneous agents with differing information, incentives, and cognitive limitations. Individual beliefs are subjective and conditional on private knowledge and publicly observable signals, including prices and other’ trades. As a result, prices reflect not objective probabilities, but the outcome of strategic interaction among adaptive agents, which captures the essence of many of the complex systems that have been discussed in class readings so far.

Recent work incorporating network and agent-based approaches represents some movement towards complexity thinking. For instance, Restocchi et al. [17] use agent-based modeling to model traders embedded in a social network who hold opinions about an event’s probability and place bets accordingly. Agents update their beliefs through local interactions following the Deffuant bounded-confidence model of opinion dynamics. Their results show prediction accuracy is maximized at intermediate levels of opinion variance, highlighting tradeoffs between diversity and coordination. However, their framework treats belief updating as consensus-seeking rather than as competition among heterogeneous forecasting strategies.

A recent paper that addresses this concern of the potential inaccuracy of information aggregation by prediction markets is that by Clinton and Huang [6]. Their comprehensive empirical analysis documents regime-dependent behavior that neoclassical models cannot explain but complexity frameworks anticipate. Analyzing nearly 2,500 political prediction markets across four exchanges (IEM, PredictIt, Kalshi, and Polymarket) during the final five weeks of the 2024 U.S. presidential campaign, they find that while 93% of PredictIt markets correctly anticipated outcomes better than chance, accuracy falls to 78% on Kalshi and 67% on Polymarket, despite Polymarket handling roughly 2.1 billion USD in volume compared to only 5.9 million USD on PredictIt. Accuracy also varies systematically by contract type, since national presidential outcome markets are relatively accurate, whereas more speculative markets like the content of speech or margin of win perform little better than or even worse than a coin flip. Yet even the most accurate markets rarely behave as if they are efficiently aggregating information. Prices for nearly identical contracts diverge sharply across exchanges and daily price changes are often weakly correlated or negatively autocorrelated, indicating mean reversion rather than a random walk. Opportunities for cross-exchange arbitrage persist on most days, peaking in the final two weeks before Election Day when information is most abundant. They further show that price movements in contracts tied to the same underlying event frequently move in unexpected ways as mutually exclusive outcomes sometimes rise or fall together and that large swings in prices often occur on days with no identifiable political events, suggesting that internal market dynamics and trader behavior, rather than new information, drive much of the observed volatility. Together, these findings challenge the view that prediction markets inherently aggregate dispersed information into a single efficient forecast and instead point to platform-specific, contract-specific, and temporally varying regimes of accuracy and inefficiency that depend on trader composition, institutional design, and strategic interaction which are all central concerns of a complexity perspective.

The following sections attempt to describe a potential complexity science framework that could explain the mechanisms generating the empirical patterns Clinton and Huang [6] paper. Rather than treating outcomes as scattered anomalies, this approach can potentially show whether these emerge systematically from interactions among different mechanisms drawing from a part of the corpus from Volumes 3 and 4 of the Foundational Papers [11], to address the empirical patterns and attempt to resolve theoretical inconsistencies that neoclassical models fail to explain. These interacting mechanisms could determine whether prediction markets aggregate information or amplify noise, and when this occurs.

1.3 Mechanisms of Information Aggregation in Complex Adaptive Markets

1.3.1 Bounded Rationality and Adaptive Ecologies

Most models assume traders in prediction markets to be rational, owing to the financial incentives and information aggregation under the efficient market hypothesis. However, given that the traders are not rational forecasters by virtue of being human, their behavior also reflects biases that can be documented. Traders may act on their own beliefs or overreact to recent information, as classic models of investor psychology predict. They may also allow for identity or ideology driven trading rooted in partisan or candidate psychological attachments rather than profit maximization, like economic models presume [6].

The Milgrom-Stokey no-trade theorem claims that under rational expectations and common priors, no trade should occur as each trader should infer that anyone willing to trade possesses an information advantage and refuse to participate [12]. Yet prediction markets exhibit substantial trading volume. Wolfers and Zitzewitz acknowledge this as an “important open theoretical question” and propose uninformed outsiders or manipulators as explanations for this behaviour, effectively conceding the rational-agent framework cannot explain observed behaviour [21].

Arthur’s inductive reasoning framework resolves this paradox by reconceptualizing how traders form expectations [2]. Rather than computing rational expectations via Bayesian updating, boundedly rational agents satisfy using competing forecasting heuristics. Herbert Simon introduced bounded rationality and established the cognitive foundations, describing how agents face computational and informational limits that make optimal calculation infeasible, forcing reliance on good enough rules of thumb [19].

Arthur’s El Farol Bar model formalizes this process [2]. Agents must predict bar attendance to decide whether to attend (they prefer going only if fewer than 60 people attend). No deductive solution exists because if everyone uses the same correct predictor, it becomes self-negating, which is essentially what the no-trade theorem anticipates. Instead, each agent maintains their own internal forecasting models or portfolio of predictors which could be recent average attendance, and assign weights to competing predictors based on their information about recent performance. The system does not necessarily converge to equilibrium, as agents constantly maintain and update their multiple internal forecasting models. Rather, successful predictors attract usage, become overcrowded, and perform poorly, triggering shifts to alternatives.

This generates what Arthur calls an “ecology of predictors”, which is a constantly co-evolving distribution of active forecasting strategies. In prediction markets, traders similarly employ different strategies based on their active predictors. Price at any moment reflects the weighted average of active predictors, not a consensus probability, and heterogeneity of these internal models is what allows it to exhibit price discovery.

Clinton and Huang provide direct empirical validation [6]. Polymarket exhibited negative serial correlation in daily price changes, which was ρ̄ = −0.324 for presidential markets, where price increases one day predicted decreases the next. This can be inconsistent with rational Bayesian updating which should produce random walks, as most theoretical models had assumed, but fits better with Arthur’s dynamic ecology of predictors. For instance, momentum strategies attract imitators, which can lead to overcrowding, which in turn triggers contrarian corrections and causes prices to overshoot in opposite direction.

The large swings identified without any external specific political event suggests that price movements reflect trader strategies changing rather than prices responding to new information [6]. In contrast, PredictIt showed weaker serial correlation, consistent with greater predictor diversity maintained by position caps preventing single strategies from dominating. PredictIt’s position caps set an 850 USD maximum per contract and a 5,000 trader limit per market.

This framework differs from Restocchi et al.’s [17] Deffuant bounded-confidence model, where agents converge toward shared beliefs through averaging. Arthur’s traders instead don’t seek consensus, they maintain competing hypotheses in ongoing competition making price formation the outcome of an ecology, not an average. The negative serial correlation could possibly support Arthur’s non-converging framework.

1.3.2 Information Cascades and Path Dependence

Reynolds’ flocking BOID model demonstrates how complex collective patterns emerge from simple local interaction rules such as separation to avoid crowding neighbors, alignment to move toward average heading of neighbors, and cohesion to move toward average position of neighbors [18]. Translated to prediction markets, these become: separation as contrarian betting when overcrowded, alignment by following momentum/whale traders, and cohesion by gravitating towards consensus prices.

The critical insight is that global coordination arises without central control, as each trader responds to local signals, generating market-level cascades. When alignment dominates, herding occurs; when separation dominates, fragmentation emerges. The balance depends on the network topology of the agents or how traders observe and influence each other.

In prediction markets, blockchain transparency while allowing for anonymity also creates connectivity. Polymarket’s public leaderboards allow traders to identify and follow high-performing whales and create network structures where a few nodes have disproportionate influence. Clinton and Huang document this by taking the case of the “French Whale”, which saw a single large trader’s positions “substantially alter prices and dominate press coverage,” demonstrating how concentrated capital combined with visibility creates self-reinforcing feedback [6]. Followers imitate whale trades, amplifying initial movements regardless of informational content.

Arthur using the idea of path-dependence demonstrates that under increasing returns and positive feedback, markets can lock into suboptimal equilibria based on early conditions depending on random historical events [1]. He mentions that two properties that arise with increasing returns are non-predictability and potential inefficiency, where ex-ante knowledge of adopters’ preferences and the technologies’ possibilities may be insufficient to predict the market outcome. Increasing returns could cause a lock-in to an inferior technology, or in the case of prediction markets, early traders could move prices in a particular direction, subsequent traders interpret the price movement as reflecting information and trade in the same direction, creating a self-reinforcing feedback that locks the market into an inaccurate forecast despite contrary evidence.

Arthur also distinguishes between diminishing returns and increasing returns. Critically, he shows that different parts of the same economic system can operate under different returns regimes simultaneously, with the regime determined by institutional structure rather than inherent to the technology. Prediction markets may also exhibit this regime-dependence. Under governance structures that impose caps and arbitrage constraints, markets may operate under diminishing returns. Berg and Rietz [3] document that IEM has account limits that may create diminishing returns to any individual trading strategy. A whale attempting to accumulate a large position hits the cap and cannot grow further, preventing the self-reinforcing advantage that increasing returns would provide. Moreover, IEM’s unit portfolio structure ensures that larger price deviations from fundamental value become more profitable to arbitrage, strengthening corrective forces as distortions increase. When attempted manipulation, their orders were overwhelmed by ordinary traders—the manipulation strategy faced increasing costs and diminishing effectiveness as it scaled up [3].

In contrast, platforms without governance constraints operate under increasing returns. Polymarket’s unlimited cryptocurrency stakes create increasing returns to whale size as larger positions generate more visibility via blockchain transparency, attracting more followers whose imitation amplifies the whale’s market impact, making the whale appear even more successful which again attracts still more followers [6]. This is Arthur’s increasing returns loop where an early advantage becomes self-reinforcing through network effects, learning effects, and adaptive expectations [1].

Critically, Arthur emphasizes that returns regimes are institutional choices, not technological necessities. Polymarket could adopt caps and impose diminishing returns and PredictIt could remove caps and create increasing returns. The lock in occurs because each platform’s trader ecology adapted to its returns regime. Polymarket’s whales and momentum traders prosper under increasing returns and would probably leave if caps were imposed and PredictIt’s arbitrageurs and accuracy-focused traders prosper under diminishing returns and couldn’t compete against whales. The returns regime constructs the trader population, which then sustains that regime which is what leads to institutional path-dependence at the platform level.

1.3.3 Institutional Governance and Design Principles

Given that Berg and Rietz [3] show that manipulation can be avoided and discouraged in prediction markets owing to market design, which allows them to remain reliable forecasting tools, along with Arthur’s lock-in regimes [1] being institutional choices, it would be interesting to look at how prediction markets function as institutions that can be governed.

Elinor Ostrom identifies design principles that enable common-pool resource systems to sustain cooperation without top-down enforcement [15]. If one were to study prediction markets through this lens, they would constitute information commons, where traders contribute private knowledge whose aggregation benefits all. Free-riding would be possibly trading on others’ information without contributing, manipulation would be perturbing the information pool, and noise trading like adding misinformation are all ways that threaten the system’s viability.

Berg and Rietz [3] demonstrate how IEM operationalize Ostrom’s principles through specific institutional mechanisms. They find that two features prove critical for manipulation resistance:

IEM restricts individual accounts to 500 USD investment per election cycle, limiting trader size relative to the market. In 2000, with hundreds of traders, 500,000+ contracts, and 210,633 USD total investment, a 500 USD account represented less than 0.25% of liquidity, which was too small to sustain price manipulation against hundreds of other traders with hundreds of thousands of dollars. When some researchers attempted manipulation using large orders, Berg and Rietz note dozens of other traders routinely submitted larger orders, so the manipulator’s trades were absorbed as noise.

This operationalizes Ostrom’s Principle 1, clear boundaries defining who accesses the resource and Principle 2, the congruence between rules and local conditions [15]. With Principle 1, participants or traders know who are part of the defined set of relationships and who they can cooperate with, and with Principle 2, the amount of the resource is restricted. By capping individual stakes, IEM prevents whales from dominating, maintaining distributed participation where no single actor can coordinate market-level behavior.

Polymarket lacks this constraint. Clinton and Huang document the consequences where it was seen that the “French Whale” accumulated positions large enough to move markets and shape media narratives [6]. Blockchain transparency allowed other traders to identify and follow the whale, amplifying influence through imitation cascades. The platform’s unlimited crypto stakes created a niche where whale-dominated dynamics were viable, so early whales attracted attention, attention attracted followers and the follower profits further reinforced whale strategies.

The second feature, as Berg and Rietz state is that “Unit portfolios do not make manipulation impossible, just difficult,” noting that manipulators creating arbitrage violations face an onslaught of arbitrageurs. This operationalizes Ostrom’s Principle 4, by the system selecting its own monitors through market design rather than external authority, where the institution recruits profit-motivated traders to enforce price discipline, by observing users behaviors.

Clinton and Huang show this mechanism’s absence on Polymarket. Contracts trade separately across platforms with no bundled arbitrage enforcement, allowing arbitrage opportunities to persist on 95% of days during the 2024 election’s final weeks. Without self-enforcing constraints, manipulation attempts face weaker counterpressures.

What also lacks in these online prediction markets is the sixth design principle [15], since the decision about the actual outcome of an event is determined differently on each platform and usually by an external non-user of the platform or an official who determines the payout, which has often led to alleged unfair losses [13].

Table 1 summarizes how different governance structures map to Ostrom’s principles and generate divergent outcomes:

1.3.4 Endogenous Preferences and Cultural Selection

Neoclassical economics also treats preferences as exogenous, where agents arrive at markets with fixed utility functions based on previously established preferences. Bowles (1998) demonstrates that economic institutions endogenously shape preferences through three channels: socialization, where repeated participation cultivates values, framing effects which is how the way choices are presented affects evaluation by participants, and social learning by observing others’ behavior updates beliefs about what is deemed as appropriate conduct [4].

In prediction markets, platform design can select for and cultivate distinct trader types. Clinton and Huang also provide evidence of possible preference-based selection [6]. Polymarket’s crypto-denominated, high-stakes environment with visible leaderboards and blockchain transparency attracted traders motivated by momentum trading, expressive/identity-driven participation, and short-term speculation rather than accuracy. In contrast, PredictIt’s academic origins which might have potentially influenced its participants, smaller stakes, and higher transaction costs (with 10% profit fee and a 5% withdrawal fee) may have selected for more academically oriented traders which cultivated a culture that prioritized accuracy.

These differences persisted despite 2.1B USD flowing through Polymarket versus 5.9M USD through PredictIt which indicated that trader type and culture generated may have mattered more than aggregate liquidity. Markets construct the preferences of participants who then reconstruct the market environment, creating self-reinforcing cultural lock-in.

This behavior and coevolution of traders can also be explained by the construction of the niches they create. Odling-Smee et al. introduced this niche construction theory, by which organisms don’t just adapt to environments but they modify environments in ways that feed back on selection pressures [14]. Applied to prediction markets, platforms and traders could be understood to co-evolve through niche construction in a three-fold manner:

Informational niche: Polymarket’s media coverage of whale trades and price volatility created an attention economy where volatility itself became newsworthy, attracting speculators seeking publicity rather than forecasters seeking accuracy [6]. PredictIt’s academic framing attracted researchers seeking data and validation.

Institutional niche: Polymarket’s U.S. restrictions led to VPN-enabled trading, creating a technically sophisticated, crypto-native user base that then demanded more crypto-integrated features [6]. PredictIt’s regulatory exemption attracted users comfortable with smaller stakes and tighter constraints.

Preference niche: Early success of momentum traders on Polymarket’s unrestricted platform attracted more momentum-seeking participants, while PredictIt’s caps repelled them. Platforms inherited different regulatory niches, which selected different populations, who constructed different informational/preference environments, making it difficult for either platform to converge on the other’s model.

Traders weren’t necessarily just responding to a shared informational environment but to locally constructed niches specific to each platform’s evolutionary trajectory.

2 Interacting Mechanisms and Regime Dynamics

The four mechanisms do not operate independently but rather can be seen to interact to produce qualitatively distinct regimes where markets either aggregate information or amplify noise. These regimes may represent different stable states with different dynamics:

2.1 Information-Aggregating Regime

Governance constraints impose prevent any single strategy or trader from dominating through self-reinforcement [3]. This maintains predictor diversity in Arthur’s ecology, and different strategies coexist because no one strategy can grow large enough to crowd out alternatives [2]. The institutional structure selects for accuracy-motivated traders which could increase accuracy and reduce discrepancies across platforms [4]. Ostrom’s design principles and self-enforcing monitoring can help maintain this information commons [15].

2.2 Noise-Amplifying Regime

Absence of governance creates can allow for whale size, momentum strategies, and reputation to amplify through positive feedback. Arthur’s ecology collapses toward a low predictor diversity and whale following and momentum trading dominate because they can succeed through self-reinforcement and crowding out [2]. The institutional environment selects for speculation-motivated traders who construct an informational niche where volatility is valued [4][14]. Without Ostrom’s boundaries or monitoring, the information commons degrades through the noise [15].

2.3 Conclusion

Holland and Miller’s observation that “complex adaptive systems usually operate far from a global optimum” because “niches are continually created by new adaptations” explains regime persistence [10]. Early governance choices (IEM’s 1988 academic exemption creating caps vs. Polymarket’s 2020 crypto basis) determined initial returns regimes, which selected initial trader types, whose strategies constructed informational niches, which attracted similar traders, reinforcing the initial regime through Arthur’s increasing returns at the platform level [1].

Switching regimes requires coordinated transitions across all four mechanisms. This is Arthur’s potential inefficiency, where both platforms and participants recognize suboptimal outcomes but face coordination failures preventing regime shifts [1].

The complexity framework thus explains Clinton and Huang’s empirical patterns not as deviations from equilibrium but as emergent properties of interacting adaptive mechanisms operating in different institutional environments. Accuracy is not an inherent property of prediction market mechanisms, contrary to claims by platforms like Polymarket that “markets seek truth” [16], but an achieved outcome dependent on governance structures that control returns regimes, maintain predictor diversity, shape network topology, and select trader populations whose preferences align with information aggregation rather than speculation.

Prediction markets are best understood not as probability oracles but as evolving complex systems whose forecasting performance depends on institutional design, network structure, and historical trajectory. Recognizing this can reframe both their promise and their limitations, shifting attention from optimal mechanisms to adaptive governance and management of these platforms effectively. If prediction markets are to serve as reliable forecasting tools rather than speculative vehicles, governance design is important.

2.3.1 Limitations and Future

This analysis relies primarily on Clinton and Huang’s 2024 election data and Berg and Rietz’s IEM evidence [6][3]. Prediction markets exemplify a broader class of information aggregation systems like social media, recommendation algorithms and crowdsourcing platforms where accuracy depends on institutional management. The complexity framework provides tools for understanding when and why aggregation succeeds or fails, moving beyond idealized equilibrium assumptions toward empirically grounded analysis of institutional functioning in adaptive systems. To better model decision-making in these agents, it would be useful to incorporate neural networks in the agent-based model to replicate predictor dynamics and model heurestics.

References

[1] W. Brian Arthur. “Competing Technologies, Increasing Returns, and Lock-In by Historical Events”. In: The Economic Journal 99.394 (Mar. 1989), pp. 116–131. doi: 10.2307/2234208. url: https://www.jstor.org/stable/2234208.

[2] W. Brian Arthur. “Inductive Reasoning and Bounded Rationality”. In: American Economic Review 84.2 (May 1994), pp. 406–411.

[3] Joyce E. Berg and Thomas A. Rietz. “Market Design, Manipulation, and Accuracy in Political Prediction Markets: Lessons from the Iowa Electronic Markets”. In: PS: Political Science & Politics 47.2 (Apr. 2014), pp. 271–276. doi: 10.1017/S1049096514000043.

[4] Samuel Bowles. “Endogenous Preferences: The Cultural Consequences of Markets and Other Economic Institutions”. In: Journal of Economic Literature 36.1 (Mar. 1998), pp. 75–111.

[5] Patrick Buckley. “Blockchain Based Prediction Markets”. In: Journal of Prediction Markets 18.3 (2024), pp. 73–86. doi: 10.5750/jpm.v18i3.2182. url: https://ubplj.org/index.php/jpm/article/view/2182.

[6] Joshua D. Clinton and Tzu-Feng Huang. Prediction Markets? The Accuracy and Efficiency of $2.4 Billion in the 2024 Presidential Election. Preprint. Version 1. SocArXiv, Dec. 2025. doi: 10.31235/osf.io/d5yx2. url: https://osf.io/preprints/socarx/d5yx2.

[7] Michael B. G. Froman. The Rise of Geopolitical Prediction Markets. Council on Foreign Relations. 2025. url: https://www.cfr.org/article/rise-geopolitical-prediction-markets.

[8] Sanford J. Grossman. “On the Efficiency of Competitive Stock Markets Where Traders Have Diverse Information”. In: The Journal of Finance 31.2 (May 1976), pp. 573–585. doi: 10.2307/2326627.

[9] Sanford J. Grossman and Joseph E. Stiglitz. “Information and Competitive Price Systems”. In: The American Economic Review 66.2 (May 1976), pp. 246–253. url: https://www.jstor.org/stable/1817218.

[10] John H. Holland and John H. Miller. “Artificial Adaptive Agents in Economic Theory”. In: The American Economic Review 81.2 (1991), Papers and Proceedings of the Hundred and Third Annual Meeting of the American Economic Association, pp. 365–370. url: https://www.yildizoglu.fr/moddyn2/articles/hollandmiller_artificial%20adaptive%20agents.pdf.

[11] David C. Krakauer, ed. Foundational Papers in Complexity Science. Vol. 3–4. Scholars Series. SFI Press, 2024.

[12] Paul Milgrom and Nancy Stokey. “Information, Trade and Common Knowledge”. In: Journal of Economic Theory 26.1 (Feb. 1982), pp. 17–27. doi: 10.1016/0022-0531(82)90046-1.

[13] Parker Molloy. Kalshi Users Are in an Uproar Over Their Time ‘Person of the Year’ Bets. Fast Company. Dec. 2025. url: https://www.fastcompany.com/91459584/kalshi-polymarket-time-magazine-artificial-intelligence-sam-altman.

[14] F. John Odling-Smee, Kevin N. Laland, and Marcus W. Feldman. “Niche Construction”. In: The American Naturalist 147.4 (Apr. 1996), pp. 641–648. doi: 10.1086/285870. url: https://doi.org/10.1086/285870.

[15] Elinor Ostrom. “Collective Action and the Evolution of Social Norms”. In: Journal of Economic Perspectives 14.3 (Summer 2000), pp. 137–158. doi: 10.1257/jep.14.3.137.

[16] Polymarket. What is Polymarket? Polymarket Documentation. 2024. url: https://docs.polymarket.com/polymarket-learn/get-started/what-is-polymarket.

[17] Valerio Restocchi et al. “Opinion Dynamics Explain Price Formation in Prediction Markets”. In: Entropy 25.8 (2023), p. 1152. issn: 1099-4300. doi: 10.3390/e25081152. url: https://www.mdpi.com/1099-4300/25/8/1152.

[18] Craig W. Reynolds. “Flocks, Herds, and Schools: A Distributed Behavioral Model”. In: Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques. SIGGRAPH ’87. New York, NY, USA: ACM, 1987, pp. 25–34. doi: 10.1145/37401.37406.

[19] Herbert A. Simon. “Bounded Rationality and Organizational Learning”. In: Organization Science 2.1 (Feb. 1991), pp. 125–134. doi: 10.1287/orsc.2.1.125.

[20] Jason Wingard. The Polymarket Effect: How Prediction Markets Are Beating The Experts. Forbes. Nov. 2025. url: https://www.forbes.com/sites/jasonwingard/2025/11/19/the-polymarket-effect-how-prediction-markets-are-beating-the-experts/.

[21] Justin Wolfers and Eric Zitzewitz. “Prediction Markets in Theory and Practice”. In: NBER Working Paper 12083 (Mar. 2006). doi: 10.3386/w12083. url: https://www.nber.org/papers/w12083.