Complex Systems in Economics and Social Sciences

The application of ideas from complex systems theory to economics and the social sciences forms a new paradigm—“complexity economics”—as opposed to traditional equilibrium models. This paradigm better explains economic crises, inequality, and leaps in innovation.

Complexity Economics

Traditional neoclassical economics: agents are rational and homogeneous, markets strive for equilibrium, forecasting describes deviations around equilibrium. This picture is convenient mathematically, but poorly explains crises, inequality, and innovation.

Complexity economics (W. Brian Arthur, SFI, 1994–2020): agents have bounded rationality (Herbert Simon). Heterogeneous strategies: fundamentalists, technicians, trend-followers. Adaptive (not rational) expectations. The economy = a constantly evolving ecosystem, not an equilibrium mechanism.

SFI (Santa Fe) experiment: ABM stock market (Palmer et al., 1994): 100 trader-agents with varying strategies trade stocks. Strategies evolve via genetic algorithm (the best strategies survive). Results: technical traders (chartists) coexist with fundamentalists; volatility “clusters”; bubbles arise spontaneously. Reproduces the real market better than DSGE models.

Financial Markets as Complex Systems

Stylized facts of markets (Cont, 2001): 11 persistent statistical properties not explained by the efficient market hypothesis:

• Fat tails of returns: $P(|r| > x) \sim x^{−\alpha}$, $\alpha \approx 3$ (“cubic tail law”)
• Volatility clustering (GARCH effect): large fluctuations follow large ones
• No autocorrelation of returns, but high autocorrelation of $|r_t|$ and $r_t^2$
• Long memory of volatility: ACF($|r_t|$) declines as $t^{−\beta}$, $\beta \approx 0.2$

These facts are reproduced by ABM models, but not standard DSGE.

Flash crashes: May 6, 2010—the Dow Jones index fell by 1000 points in 36 minutes and almost fully recovered. Cause: a cascading failure in the algorithmic trading system—an effect of a complex system. Algorithms reacting to price movements set off a self-reinforcing feedback loop.

Systemic risk and “too connected to fail”: The interbank network is scale-free. The bankruptcy of a hub (Lehman Brothers, 2008) → cascading defaults. The myth of diversification: risk increased with increased interconnectedness (May’s paradox for finance).

Network Economics and Market Platforms

Network effects: the value of a network to a user grows with the number of users. Metcalfe: $V \sim n^2$. Facebook, Uber, Amazon extract value from network effects.

Critical mass: below the user threshold → the network is not viable. Above → exponential growth. Explains “winner-takes-all”: a network with a slightly larger market share gets even more users → monopoly.

Two-sided markets: a platform connects two types of agents (buyers + sellers on Amazon, developers + users on iOS/Android). Pricing is complex: often one side is subsidized (WhatsApp is free, monetization via advertising to the other “side”).

Urban Complexity and Scaling Laws

Scaling laws of cities (Bettencourt, West, 2007): for $N > 100$ thousand cities worldwide, $W \sim N^\beta$:

• $\beta > 1$ (superlinear): salaries ($\beta=1.15$), patents/innovation ($\beta=1.27$), GDP ($\beta=1.13$), crime ($\beta=1.16$), disease ($\beta=1.23$)
• $\beta < 1$ (subintensive): road length ($\beta=0.83$), number of gas stations ($\beta=0.77$), electricity consumption ($\beta=0.87$)

Interpretation: large cities save on infrastructure ($\beta < 1$—economies of scale) and generate disproportionately more innovation and problems ($\beta > 1$—“pace of life” acceleration). Mechanism: density of interactions grows superlinearly with the number of inhabitants.

Pace of life in cities: pedestrian walking speed (Bornstein & Bornstein, 1976): increases proportionally to $\log(N)$. The “pulse” rate of a metropolis quickens with its size.

Numerical Example: Scaling Law of Russian Cities

Data on the 15 largest cities in Russia (2023): Moscow (12.6M), Saint Petersburg (5.6M), ... Novosibirsk (1.6M). GRP per capita vs population: OLS in log-log → $\beta \approx 1.12$ ($R^2 = 0.82$). Moscow produces $1.12^{1.93} \approx 1.24$ times more GRP per capita than would be expected linearly from the difference with St. Petersburg.

Assignment: Collect data on the 20 largest cities of your country (population, GDP, number of patents, road length). (1) Build log-log scatter plots for each indicator vs population. (2) Estimate $\beta$ via OLS. (3) Test the superlinearity hypothesis (t-test: $\beta=1$ vs $\beta>1$). (4) Use NetworkX: model an intercity trade network (edges = turnover of goods). Calculate betweenness centrality—which city is the most “central”?