96 experiments across 16 agent-based models exploring Beinhocker's Origin of Wealth
Last updated March 27, 2026 · Seed 42 · Full reproducibility
Eric Beinhocker argues that the economy is not an equilibrium machine but a complex adaptive system — an evolving ecology of strategies, technologies, and institutions that perpetually generates novelty, never settles into rest, and produces its own crises as naturally as it produces growth. This research program tests that thesis through computational simulation: 16 models, 96 experiments, and thousands of agent interactions producing measurable emergence, non-linearity, power laws, and evolutionary dynamics.
Launch Simulations →Sixteen models, ninety-six experiments. Each section presents key findings, metrics, and emergent behaviors.
Heterogeneous adaptive traders with evolving forecasting rules on a call market. Based on Arthur, Holland, LeBaron, Palmer & Tayler (1997).
Learning agents produce fat-tailed returns with a Hill tail index of 2.55 — strikingly close to the empirical cubic power law (~3) found in real equity markets — while the rational baseline produces near-Gaussian tails. Fat tails, volatility clustering, and excess volume are the natural signature of adaptive agents co-evolving in a market.
As soon as heterogeneity and learning are introduced, things get much richer and more complex. … All this price movement is driven by the dynamic interactions of various rules in the population and has little or nothing to do with changes in the underlying economic value of the stock. Nor are the complex patterns due merely to random noise. Instead, there is a complex battle of beliefs going on within the heads of agents and among the agents, which leads to volatility and complex patterns in the market.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 6, p. 138| Experiment | Key Parameter | Kurtosis | Tail Index | Vol. Cluster | Insight |
|---|---|---|---|---|---|
| 1a. Learning (baseline) | Default (25 agents) | 45.3 | 2.55 | 0.489 | Realistic market statistics emerge from learning |
| 1b. Rational (no learning) | Fixed rules | 4.8 | 46.38 | -0.055 | Equilibrium fails: highest volatility, near-zero volume |
| 2. High Mutation | Rate 0.10 (3x) | 51.7 | 2.50 | 0.555 | More mutation sustains edge-of-chaos dynamics |
| 3. Large Population | 100 agents (4x) | 18.0 | 3.85 | 0.232 | Diversity stabilizes; fat tails persist |
| 4. Low Risk Aversion | Lambda 0.1 (5x lower) | 69.5 | 2.75 | 0.410 | Proto-bubbles; skewness 2.78 |
| 5. Fast Evolution | GA interval 50 (5x faster) | 140.9 | 2.79 | 0.449 | Red Queen dynamics triple kurtosis |
The rational-expectations baseline — the gold standard of neoclassical finance — produces the worst outcomes: highest volatility, largest mispricing, near-zero liquidity. The perpetually out-of-equilibrium learning market tracks fundamentals better, trades more actively, and distributes wealth more equally. Equilibrium is not a useful approximation; it is the opposite of what happens. Adaptation and instability are two sides of the same evolutionary coin.
Four-echelon supply chain with Sterman's anchor-and-adjust ordering heuristic. Based on Forrester (1961) and Sterman (1989).
Shipping delay dominates all other factors with super-exponential cost scaling: delay 2 costs $3,236; delay 4 costs $35,908 (11x); delay 5 costs $142,427 (44x). A single, tiny, one-time demand perturbation generates 30–50 ticks of wild endogenous oscillation — the business cycle is purely structural, not driven by external shocks.
The Beer Game is not a mere wiggling-jelly propagation mechanism. The game receives a single exogenous shock — the increase in orders from four to eight. But unlike a jelly given a single tap, once the oscillations start in the Beer Game, the system never returns to equilibrium. The ultimate source of the oscillations in the Beer Game is not the exogenous shock itself (it just gets things started), but the behavior of the participants and the feedback structure of the system. The system is not propagating exogenous dynamics; it is endogenously creating them.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 8, pp. 171–172| Experiment | Key Parameter | Total Cost | Max Bullwhip | Insight |
|---|---|---|---|---|
| 1. Baseline (behavioral) | Delay=2, step demand | $3,236 | 7.89x | Severe bullwhip from simple heuristics |
| 2. Rational comparison | Optimal ordering | $1,984 | 2.36x | 63% bounded-rationality surcharge |
| 3. Longer delay (4) | Delay doubled | $35,908 | 29.18x | 11x cost; super-linear scaling |
| 4. Sine demand | Continuously varying | $40,211 | 7.59x | Never converges; perpetual oscillation |
| 5. Information sharing | Demand transparency | $1,844 | 5.05x | 43% savings; nearly matches rational |
| 6. Extreme delay (5) | Delay=5, 80 ticks | $142,427 | 60.40x | Signal destruction; 2,137 cases idle |
Business cycles may be endogenous rather than exogenous. The architecture of the system — its delays, topology, information structure — dominates agent sophistication by an order of magnitude. The structure tax (44x from delay) dwarfs the human tax (1.6x from bounded rationality). Economies are not machines to be fine-tuned but weather systems that generate their own dynamics.
Kauffman random Boolean networks exploring the phase transition between order and chaos as a function of connectivity. Based on Kauffman (1969, 1993).
At K=2, the network sits at the edge of chaos with measured Derrida parameter 1.010 (theoretical: 1.000). This is the sweet spot where perturbations propagate far enough to enable adaptation but not so far as to destroy coherence. At K=5, cascades engulf 93.4% of the network and no stable states can be found — the complexity catastrophe.
This kind of interdependency in a network creates what Kauffman calls a complexity catastrophe. The effect occurs because as the network grows, and the number of interdependencies grows, the probability that a positive change in one part of the network will lead to a cascade resulting in a negative change somewhere else grows exponentially with the number of nodes. This in turn means that densely connected networks become less adaptable as they grow.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 7, p. 152| Experiment | Regime | Derrida | Cascade % | Attractors | Insight |
|---|---|---|---|---|---|
| 1. Phase Diagram | Sweep K=1..7 | varies | varies | varies | Sharp boundary follows 2Kp(1-p)=1 |
| 2. Ordered (K=1) | Ordered | 0.445 | 5.8% | 30/30 | Perturbations absorbed; stable but rigid |
| 3. Critical (K=2) | Critical | 1.010 | 11.8% | 30/30 | Edge of chaos; max cascade 24 (48%) |
| 4. Chaotic (K=5) | Chaotic | 2.265 | 93.4% | 0/30 | Total chaos; no attractors found |
| 5. Hierarchy vs Random | Chaotic | ~1.50 | 60-62% | 30/30 | Hierarchy's value emerges at larger N |
| 6. Scale (N=20..200) | Critical | ~1.00 | 7-23% | 30/30 | Critical networks scale gracefully |
The ~7-person working group limit is not arbitrary cognitive psychology but an emergent constraint from the mathematics of interdependent decisions. Hierarchy is computational infrastructure for managing complexity, not merely a power structure. Organizations should target K=2 effective connectivity — the edge of chaos where complex adaptive systems are most productive.
Hybrid Jain-Krishna / Bak-Sneppen model of ecosystem evolution with directed weighted interaction graph. Self-organized criticality and power-law cascades.
Cascade sizes follow a power law with exponent ~2.0, consistent with empirical extinction data. The system drives itself to criticality without tuning. Tripling connection density transforms the ecosystem from occasional small cascades (mean 2.5) to perpetual catastrophe (mean 56.9) — the connectivity-fragility paradox.
Jain and Krishna noted three distinct phases to the punctuated equilibrium pattern. First, in a random phase, the network percolates along without much structure, and random changes occur without much effect. Then, an innovation sends the network suddenly into a growth phase. … The network continues to bubble along for a while in the organized phase, but then an innovation or a random change hits a keystone species. Changes influencing the keystone species radiate into the structure, and the network crashes in a wave of extinction. The process then begins again with a new random phase.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 8, p. 174| Experiment | Key Parameter | Mean Cascade | Max Cascade | Insight |
|---|---|---|---|---|
| 1. Baseline (100 species) | Conn. prob. 0.05 | 2.5 | 21 | Power-law cascades; exponent ~2.0 |
| 2. Small ecosystem (20) | N=20 | 2.1 | 12 (60%) | Volatile; 10 punctuation events |
| 3. Large ecosystem (200) | N=200 | 23.4 | 120 (60%) | Rare but catastrophic tail events |
| 4. Dense (0.15) | Conn. prob. 3x | 56.9 | 89 | Perpetual catastrophe; bimodal cascades |
| 5. Sparse (0.02) | Conn. prob. 0.4x | 2.4 | 11 | Achieved ORGANIZED phase; low diversity |
| 6. Cascade test | Systematic removal | 2.2 | 9 | Keystone heuristic misses structural nodes |
Creative destruction is not an occasional disruption of an otherwise stable system — it is the system's natural mode of operation. The power-law cascade distribution explains why most innovations cause minor adjustments while a handful — the steam engine, the internet — trigger waves of creative destruction reshaping entire industries. Disequilibrium is the norm.
Harrington's organizational adaptation simulation with tournament-based promotion in changing environments. Based on Harrington (1999) and March (1991).
The promotion tournament is a rigidity ratchet: it ruthlessly selects rigids upward during stable periods, then delivers catastrophic performance collapse (+0.87 drop) when the environment shifts. At stability=500 (the "Thatcher trap"), the organization is 97.5% rigid with peak performance of 1.44 and zero resilience. At stability=20, the hierarchy inverts: the top is 100% flexible, and transition costs vanish.
Evolutionary systems work best when their sensitivity to change is in a medium, in-between range. If an evolutionary system is too insensitive to change, then the system will not be able to keep up with the pace of change in its environment. However, if a system is overly sensitive to change, then small changes can have large consequences. This oversensitivity is a problem because if a system has been successful in the past, then few major changes are likely to improve it. Rather, the odds are that the vast majority of possible major changes will harm it.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 7, pp. 157–158| Experiment | Stability | Overall Rigid% | Avg Perf. | Transition Cost | Insight |
|---|---|---|---|---|---|
| 1. Baseline | 100 | 75.0% | 1.194 | +0.337 | Purge-crash cycle |
| 2. High Stability | 500 | 97.5% | 1.439 | n/a | Thatcher trap: peak performance, zero resilience |
| 3. High Volatility | 20 | 60.0% | 1.044 | +0.041 | Top 100% flexible; hierarchy inverts |
| 4. Random Mode | 100 (rand) | 90.0% | 1.256 | +0.224 | Pattern matters as much as frequency |
| 5. Deep Hierarchy (6 levels) | 100 | 71.4% | 1.192 | +0.049 | Every level above L0 is 100% rigid by tick 50 |
| 6. Shallow Hierarchy (2 levels) | 100 | 75.0% | 1.096 | +0.138 | Less amplification, more stochastic noise |
| 7. High Experience Weight | 100 | 52.5% | 3.116 | +0.011 | Seniority overwhelms strategy; 0 tick recovery |
The key to long-term survival is not being the best adapted to the current environment, but being the most adaptable to environments that have not arrived yet. Pure exploitation yields peak 1.44 performance with zero resilience. Pure exploration yields steady 1.04 with perfect resilience. The 25% gap is the cost of insurance — and you cannot know if the premium is worth paying until the disruption arrives.
Iterated Prisoner's Dilemma with evolutionary dynamics on a spatial grid. Seven strategies compete: AllC, AllD, TFT, GTFT, Pavlov, Random, Grudger. Based on Axelrod (1984).
The same strategies, payoff matrix, and initial conditions produce diametrically opposite outcomes depending on population structure: spatial yields 100% cooperation (Grudger dominant at 46%), while the well-mixed tournament yields 0% cooperation (AllD dominance at 100%). Population structure is the foundation of cooperation.
That we are conditional cooperators and altruistic punishers should not be surprising. Our hominid ancestors spent about 2 million years of their existence living in small bands for which cooperative behavior and survival were highly correlated. Today, people still inhabit networks of social interactions in which reciprocity — I'll scratch your back if you scratch mine — is important. We are all better off if we help each other out, but this creates the potential for abuse by those who take benefits without giving back.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 6, p. 121| Experiment | Condition | Coop. Rate | Winner | Insight |
|---|---|---|---|---|
| 1. Spatial Baseline | 50x50 grid, noise=0 | 100% | Grudger (46%) | Spatial clustering enables cooperation |
| 2. Tournament Baseline | Well-mixed, 100 agents | 0% | AllD (100%) | Without spatial protection, defection wins |
| 3. High Noise (10%) | Spatial, noise=0.10 | 75.3% | Pavlov (97.6%) | Win-Stay/Lose-Shift corrects errors best |
| 4. Zero Noise | Spatial, noise=0 | 100% | Grudger (46%) | Permanent retaliation is optimal when deterministic |
| 5. High Mutation (5%) | Spatial, mut=0.05 | 97.3% | Grudger (71%) | All 7 strategies persist; cooperation robust |
| 6. High Temptation (T=10) | Spatial, T=10 | 99.9% | Grudger (94%) | Higher temptation strengthens cooperation (paradox) |
Cooperation is an emergent property of spatial evolutionary dynamics, not a pre-programmed outcome. Structure matters — the same agents with the same rules produce radically different macro-level outcomes depending on population structure. Markets, legal systems, and social norms evolve bottom-up from repeated strategic interactions.
Heterogeneous agents on a 2D resource landscape with vision, metabolism, movement, and optional reproduction. Based on Epstein & Axtell (1996).
Moderate inequality (Gini 0.26–0.40) emerges from identical behavioral rules and random initial conditions. Metabolism matters more than vision for survival (r = -0.55 to -0.83 vs r = 0.22 to 0.39). With reproduction enabled, traits evolve cumulatively toward near-optimal values: 97.3% of agents converge to metabolism=1 after 300 ticks.
The answer is, in essence, “everything.” The skewed distribution is an emergent property of the system. It is a macro behavior that emerges out of the collective micro behavior of the population of agents. The combination of the shape of the physical landscape, the genetic endowments of the agents, where they were born, the rules that they follow, the dynamics of their interactions with each other and with their environment, and, above all, luck all conspire to give the emergent result of a skewed wealth distribution.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 4, p. 86| Experiment | Key Parameter | Survival | Gini | Mean Wealth | Insight |
|---|---|---|---|---|---|
| 1. Baseline | 400 agents, 50x50 | 43% | 0.375 | 211.6 | Selection on metabolism > vision |
| 2. High Density (800) | Double population | 36% | 0.395 | 153.2 | Competition suppresses wealth |
| 3. Low Regrowth (0.25) | Quarter regrowth | 28% | 0.263 | 145.3 | Scarcity equalizes: not enough surplus |
| 4. Reproduction | Threshold=50, 300 ticks | 326% | 0.148 | 33.6 | Trait evolution; Malthusian equilibrium |
| 5. Large Grid (100x100) | 800 agents, big grid | 16% | 0.376 | 225.9 | Geography dominates genetics |
| 6. Extreme Ranges | Vision 1-10, met 1-6 | 33% | 0.308 | 247.8 | Brutal selection; dead invisible to Gini |
Wealth inequality is not imposed externally but emerges endogenously from agent heterogeneity interacting with resource geography. The reproduction experiment demonstrates how differential survival and heredity with variation produce directional evolutionary change — a purely economic analog of biological natural selection. The system never reaches static equilibrium.
Firms search a 12-dimensional binary fitness landscape with tunable epistatic interdependencies (K), using local hill-climbing, long-jump adaptation, and recombination. Based on Kauffman's NK model.
Landscape ruggedness (K) drives market turbulence: K=8 produces 40 firm destructions and 0.483 technology diversity, while K=1 produces only 17 destructions and 0.222 diversity. The mixed-strategy portfolio (local + long-jump + recombination) outperforms any single search strategy, confirming that diversity of approaches is the source of economic fitness.
For landscapes that are in between, are rough-correlated, and have complex features such as plateaus, holes, and portals, evolution is hard to beat. And when the landscape is constantly changing, when the search problem is a dynamic one, when one must balance the tension between exploring and exploiting — evolution truly is the grand champion.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 9, p. 213| Experiment | K | Strategy | Best Fitness | Destructions | Insight |
|---|---|---|---|---|---|
| 1. Baseline | 4 | Mixed | 0.800 | 31 | Local climbers dominate final population |
| 2. Smooth (K=1) | 1 | Mixed | 0.656 | 17 | Quick convergence, then stagnation |
| 3. Rugged (K=8) | 8 | Mixed | 0.767 | 40 | Highest innovation AND destruction |
| 4. High Mutation | 4 | Mixed, mut 3x | 0.800 | 32 | Noise is not a substitute for strategy diversity |
| 5. All Local Climbers | 4 | 100% local | 0.800 | 24 | Fast initial gains, then trapped on local optima |
| 6. All Long Jumpers | 4 | 100% long-jump | 0.800 | 23 | Slower convergence, sustained exploration |
Technology is fundamentally combinatorial search. No single search strategy dominates — economies benefit from a portfolio of approaches: some firms doing incremental improvement, others making radical leaps, others recombining. Creative destruction (Schumpeter's gale) is reproduced, with intensity depending on landscape ruggedness (industry complexity).
Four interacting sub-models — supply chain, stock market, ecosystem dynamics, and organizational adaptation — linked through composite firm health, dependency networks, and cascade failures.
Scale creates fragility, not just bigger failures: at 40 firms, the market crash kills 28 (70%); at 100 firms, the same supply shock triggers total wipeout (100%) before the panic even arrives. The cascade arrives 47 ticks earlier in the larger economy — a phase transition in network behavior. Shock sequencing matters more than magnitude: a 2.5x supply shock alone causes zero failures, but combined with a panic, it is lethal.
Oscillations that do not settle down, punctuated equilibrium, and power laws — these are all signature behaviors of a complex adaptive economy at work. The real-world economy is a far more interesting place than the equilibrium world imagined by Traditional Economics. Complexity Economics does not have all the answers to the puzzle of economic patterns, but it provides us with new tools to begin to understand how these various factors combine to result in the behaviors we observe.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 8, p. 185| Experiment | Scenario | Mean GDP | Max Unemploy. | Cascades | Insight |
|---|---|---|---|---|---|
| 1. Normal Operations | No shocks, 200 ticks | 11.20 | 0% | 0 | Peacetime rigidity drift (flex 0.40 to 0.20) |
| 2. Supply Shock | 2.5x demand at t=100 | 15.42 | 0% | 0 | Bullwhip visible; GDP overshoots 3x |
| 3. Tech Disruption | Forced at t=150 | 11.20 | 0% | 0 | Zero GDP impact; flex jumps 0.20 to 0.64 |
| 4. Market Crash (40) | Triple shock sequence | 18.19 | 67.5% | 28, 3 | Panic is lethal; 10-tick delay before cascade |
| 5. Stress Test | Repeated shocks, 300t | 12.75 | 0% | 0 | Flexibility ratchets to 0.80; full survival |
| 6. Large Economy (100) | Market crash, 100 firms | 18.26 | 97% | 100, 3 | Total wipeout; phase transition in fragility |
Repeated, spaced shocks make the economy more resilient by driving organizational flexibility upward — validating Beinhocker's argument that economies need periodic disruption to maintain adaptive capacity. The integrated model shows how supply chain stress, market sentiment, technology shifts, and organizational rigidity interact to produce emergent macroeconomic dynamics that no single sub-model could generate alone.
Arthur's inductive reasoning model: N agents with competing prediction strategies decide weekly whether to attend a bar with limited capacity. No deductive equilibrium exists.
The system self-organizes with mean attendance near the threshold (57.3 vs. 60) despite no coordination mechanism. Mean prediction accuracy hovers at 45–47% — worse than coin-flipping — directly demonstrating the self-referential paradox: when your prediction affects the outcome, no strategy can consistently win.
Arthur's bar problem raises the intriguing possibility that much of the volatility we see in the real-world economy may be generated by the dynamics of people's decision rules, rather than by exogenous, random shocks.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 6, p. 125| Experiment | Key Parameter | Mean Att. | Crossing Rate | AC(1) | Insight |
|---|---|---|---|---|---|
| 1. Baseline | N=100, T=60, k=10 | 57.3 | 0.543 | -0.206 | Self-organization around threshold |
| 2. Small Pop (N=50) | No scarcity | 46.9 | 0.000 | 0.965 | Trivial: all attend, no paradox |
| 3. Large Pop (N=200) | Severe scarcity | 66.9 | 0.513 | -0.066 | Doubled volatility; nearly random |
| 4. Low Threshold (T=30) | 30% capacity | 32.8 | 0.538 | -0.097 | Periodic strategies dominate |
| 5. Few Strategies (k=3) | Limited repertoire | 58.1 | 0.578 | -0.328 | More predictable oscillation |
| 6. Many Strategies (k=20) | Rich ecology | 59.0 | 0.583 | -0.138 | Better self-organization; randomizers win |
The El Farol problem is a direct challenge to rational expectations. There is no way to form a "rational expectation" about attendance because your expectation changes your behavior which changes the outcome. Beinhocker argues this self-referentiality is pervasive in real economies, not a special case. The agents embody inductive reasoning — maintaining portfolios of heuristics and betting on what has worked, exactly as real people navigate complex environments. The perpetual oscillation, never reaching equilibrium, is Beinhocker's "perpetual novelty" made visible.
Agents on a 50x50 torus grid relocate when the fraction of similar neighbors falls below a tolerance threshold. Based on Schelling (1971).
A 30% tolerance threshold — agents happy being a minority — produces 75% neighborhood homogeneity, a 49% amplification from the random baseline. A 50% threshold drives segregation to 90%. The micro-macro disconnect is dramatic: mild individual preferences generate extreme collective outcomes that no agent intended.
Emergent phenomena are characteristics of the system as a whole that arise endogenously out of the interactions of agents, rather than being imposed from the outside or simply being the aggregate of individual agent behaviors. Emergence is what makes complex systems more than the sum of their parts.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 5, p. 101| Experiment | Threshold | Density | Final Seg. | Amplification | Insight |
|---|---|---|---|---|---|
| 1. Baseline | 30% | 70% | 0.750 | +49% | Mild preference produces extreme segregation |
| 2. Moderate | 50% | 70% | 0.899 | +79% | Near-total segregation; 1,823 moves |
| 3. High Intolerance | 75% | 70% | 0.998 | +99% | 65,820 moves; perfect segregation |
| 4. Very Tolerant | 10% | 70% | 0.540 | +7% | Control: minimal effect below critical region |
| 5. Sparse | 30% | 50% | 0.778 | +59% | Sparse: many small clusters |
| 6. Dense | 30% | 95% | 0.753 | +52% | Dense: massive contiguous blocks |
Schelling segregation is the canonical demonstration of emergence in social systems. The agent rule is trivial (check neighbors, maybe move), yet the emergent spatial patterns are intricate and unpredictable. No central planner designed the segregation; it arises as a pure collective phenomenon. Small changes in individual tolerance produce qualitatively different macro outcomes — a hallmark of complex adaptive systems and a caution that individual good intentions do not guarantee collective good outcomes.
Lotka-Volterra dynamics in both ODE and spatial agent-based variants on a 50x50 toroidal grid. Prey, predators, and grass interact through consumption, reproduction, and starvation.
The ODE model guarantees eternal oscillation — populations approach but never reach zero. The spatial agent-based model can and does produce near-extinctions (predators dropped to 1 individual under fast starvation). This extinction gap is the most important qualitative difference: the deterministic equilibrium model misses the possibility that real systems can collapse.
Oscillations that do not settle down, punctuated equilibrium, and power laws — these are all signature behaviors of a complex adaptive economy at work. The real-world economy is a far more interesting place than the equilibrium world imagined by Traditional Economics.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 8, p. 185| Experiment | Model | Prey Range | Predator Range | Period | Insight |
|---|---|---|---|---|---|
| 1. Spatial Baseline | Agent-based | 35–1,075 | 15–138 | ~56 | Classic boom-bust cycling |
| 2. ODE Baseline | ODE (RK4) | 8–40 | 3–23 | 5.48 | Perfect orbits; no extinction possible |
| 3. More Predators | Agent, init=50 | 17–1,173 | 14–147 | ~37 | Transient differs, long-run similar |
| 4. Fast Starvation | Agent, starve=5 | 2–1,260 | 1–158 | ~20 | Near-extinction; predators at 1 |
| 5. High Prey Repro | Agent, repro=0.10 | 57–1,010 | 22–161 | ~22 | Faster recovery, shorter cycles |
| 6. Larger World | Agent, 100x100 | 200–5,582 | 8–464 | long | Spatial refugia stabilize dynamics |
The predator-prey model demonstrates that complex adaptive systems generate their own dynamics without external forcing. The boom-bust cycles are entirely endogenous, paralleling Beinhocker's argument that business cycles may arise from system structure rather than external shocks. The extinction gap between ODE and agent-based models validates his critique of equilibrium approaches: the smooth deterministic world misses catastrophic regime changes that discrete, stochastic systems naturally produce.
Bak-Tang-Wiesenfeld abelian sandpile model. Grains dropped one at a time onto a grid; cells topple when they exceed a threshold, redistributing grains to neighbors. Based on Bak, Tang & Wiesenfeld (1987).
The system spontaneously self-organizes to a critical state with power-law avalanche distribution (exponent ~1.26–1.30). The ratio of maximum to median avalanche size is ~600:1 — there is no "typical" event size. At criticality, 43.9% of cells sit at height 3 (one grain below threshold), a loaded-spring state that enables cascade propagation.
Power laws are a signature characteristic of complex adaptive systems. … In a power-law distribution there is no “typical” event size. Events of all scales occur, from the very small to the very large, with a specific mathematical relationship between size and frequency.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 8, p. 179| Experiment | Grid | Threshold | Ticks | Exponent | Insight |
|---|---|---|---|---|---|
| 1. Baseline | 50x50 | 4 | 10k | 1.289 | Power-law avalanches; max 6,647 |
| 2. Small Grid | 25x25 | 4 | 10k | 1.299 | Similar exponent; max capped at 972 |
| 3. Large Grid | 100x100 | 4 | 10k | 2.225 | Not yet critical; mean height only 1.0 |
| 4. Long Run | 50x50 | 4 | 50k | 1.256 | Better statistics confirm power law |
| 5. High Threshold | 50x50 | 8 | 10k | 1.711 | Sub-critical; insufficient loading |
| 6. Low Threshold | 50x50 | 2 | 10k | 1.190 | 85% of drops cause avalanches |
The sandpile is the purest demonstration of self-organized criticality — the system drives itself to a state where events of all sizes occur with no characteristic scale. Traditional risk models based on normal distributions systematically underestimate tail risk because they assume a "typical" event size exists. In power-law systems, extreme events are not outliers but an inherent, predictable consequence of the system's dynamics. This explains why financial crises, market crashes, and technological disruptions seem to "come out of nowhere" — they are the inevitable large avalanches in a system at criticality.
Business Plans as the evolutionary units of the economy. Each BP encodes Physical Technology (PT), Social Technology (ST), and Strategy on NK fitness landscapes. Fitness is multiplicative: PT × ST × market_fit. Based on Beinhocker, Chapter 14.
Multiplicative fitness creates a binding constraint problem: total fitness (~0.30–0.57) is far below any single component (~0.65–1.0) because the weakest component drags everything down. Stable preferences produce 81% more wealth but collapse diversity to near-zero (0.006). Too much innovation is as bad as too little — mutation rate 0.15 actually lowers PT and ST fitness below baseline.
Wealth is knowledge and its origin is evolution. … The economy is the ultimate open-ended evolutionary system. … Business Plans are the economic equivalent of DNA — they are the instructions, the code, for building economic organisms that can survive and reproduce in the economic environment.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 14, pp. 317–318| Experiment | Key Parameter | Total Wealth | Mean Fitness | Diversity | Insight |
|---|---|---|---|---|---|
| 1. Baseline | 50 BPs, mut=0.05 | 19.90 | 0.304 | 0.089 | Weakest component is binding constraint |
| 2. Stable Prefs | No pref shifts | 35.93 | 0.573 | 0.006 | +81% wealth; strategy converges to 1.0 |
| 3. Rapid Shifts | Pref shift 0.1 | 19.95 | 0.322 | 0.058 | Red Queen: 37% more innovation, same wealth |
| 4. Low Innovation | Mutation 0.01 | 28.11 | 0.483 | 0.039 | Higher wealth through optimization |
| 5. High Innovation | Mutation 0.15 | 19.87 | 0.318 | 0.322 | Excess mutation destroys accumulated knowledge |
| 6. Large Economy | 200 BPs | 90.71 | 0.417 | 0.145 | 4.6x wealth from 4x population (mild IRS) |
This simulation directly operationalizes Beinhocker's central thesis: wealth is "fit order" — the degree to which business plans match their environment. Business Plans serve as economic DNA, encoding the instructions for wealth creation. The multiplicative fitness function confirms that all three G-R conditions (Physical Technology, Social Technology, and Strategy) must work together. Creative destruction operates as Schumpeterian selection, with the bottom fraction constantly replaced by better-adapted plans. The exploration-exploitation dilemma mirrors real economies: too little innovation leads to lock-in; too much prevents optimization.
Firms compete via strategic portfolios across shifting market niches. Three strategy types — exploiters (concentrate resources), explorers (diversify broadly), and adaptive (dynamically reallocate based on feedback) — compete for market share on evolving fitness landscapes. Based on Beinhocker, Chapter 15.
Adaptive firms dominate with 63.2% market share in the baseline, confirming that a portfolio approach — dynamically reallocating resources based on feedback — outperforms both pure exploitation and pure exploration. Pure exploration is a death sentence across all environments (mean fitness 0.36 vs. 0.65 baseline). Exploiter monocultures are fragile: high fitness during stability, followed by catastrophic synchronized crashes when niches shift.
Strategy is not a detailed plan of action. Strategy is a portfolio of experiments — a population of competing business plans that are tested against the real world and selected for fitness. The key is to run many experiments, learn fast, scale the winners, and cut the losers.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 15| Experiment | Key Parameter | Exploiter Share | Adaptive Share | Mean Fitness | Insight |
|---|---|---|---|---|---|
| 1. Baseline | Mixed, shift=0.05 | 27.6% | 63.2% | 0.6514 | Adaptive firms dominate via portfolio reallocation |
| 2. Stable Market | shift=0.0 | 67.8% | 21.9% | 0.6375 | Exploiters win when landscape is static |
| 3. Volatile Market | shift=0.2 | 40.7% | 50.0% | 0.6106 | Adaptive leads; even explorers eliminated |
| 4. All Exploiters | exploit=1.0 | 91.2% | 0.0% | 0.6304 | Competency trap: synchronized crashes |
| 5. All Explorers | explore=1.0 | 0.0% | 0.0% | 0.3622 | Discover many, master none |
| 6. Large Economy | 100 firms, 10 niches | 60.6% | 29.0% | 0.5954 | Individual specialization, collective diversity |
These results directly support Beinhocker's Chapter 15 argument that strategy is evolutionary, not predictive. "Strategy is a portfolio of experiments": adaptive firms that maintain multiple bets and reallocate based on feedback consistently outperform single-bet strategies. "Robustness over optimality": the adaptive strategy is never the best at any single moment but performs well enough across all conditions to accumulate the highest long-run share. Markets function as massively parallel search algorithms — individual firms specialize while the system as a whole explores broadly.
Five policy regimes — laissez-faire, social democrat, innovation state, protectionist, and adaptive — compete in an evolutionary economy. 50 firms evolve on an NK fitness landscape (N=12, K=3) under different policy levers affecting entry barriers, compliance costs, mutation rates, market share caps, and safety net strength. Based on Beinhocker, Chapter 18.
The adaptive regime dominates on nearly every metric: highest mean GDP (9.83, 60% above laissez-faire), best long-run growth (near-flat vs. laissez-faire's −0.364 decline), lowest unemployment (15.3%), and highest final fitness (0.715). It discovered a hybrid configuration no fixed regime offered — low regulation, high innovation subsidies, and a strong safety net. Protectionism is the worst regime by far, confirming that blocking evolutionary entry is the cardinal policy sin.
The role of policy in a complex adaptive economy is not to engineer particular outcomes but to shape the fitness environment — to create the conditions under which evolutionary search can operate most effectively. Good policy enables adaptation; bad policy blocks it.
— Eric D. Beinhocker, The Origin of Wealth, Chapter 18| Experiment | Regime | Mean GDP | Mean Unemp | Mean Gini | Insight |
|---|---|---|---|---|---|
| 1. Baseline | Laissez-faire | 6.14 | 39.4% | 0.250 | High unemployment, weak safety net |
| 2. Explicit LF | Laissez-faire | 6.14 | 39.4% | 0.250 | Confirms baseline reproducibility |
| 3. Social Democrat | Social-democrat | 7.59 | 19.2% | 0.228 | Lowest Gini; most total innovations (2,021) |
| 4. Innovation State | Innovation-state | 8.13 | 23.5% | 0.245 | 2nd-highest GDP; high mutation rate |
| 5. Protectionist | Protectionist | 4.70 | 33.1% | 0.192 | Lowest GDP; entry barriers stifle search |
| 6. Adaptive | Adaptive | 9.83 | 15.3% | 0.266 | Dominates; discovers hybrid policy mix |
The simulation confirms Beinhocker's core arguments from Chapter 18. Policy shapes fitness landscapes, not outcomes — no regime could "pick" the optimal firm strategy, but regimes differed dramatically in how many firms survived to search and how fast they searched. Evolutionary dynamics require a variation supply: the strongest regimes maintained high firm populations through safety nets and respawning, while the weakest lost firms permanently. Adaptive governance outperforms fixed rules because complex systems require responsive, not ideological, governance. And protecting incumbents is the worst possible policy: it restricts the evolutionary process by blocking entry, shielding unfit firms, and reducing variation.
Cross-cutting themes from the synthesis, each demonstrated by one or more simulations.
The rational-expectations baseline produces the worst market: highest volatility (0.0833), price drift, near-zero volume (0.05). Learning agents, perpetually out of equilibrium, track fundamentals more closely with half the volatility (0.0423) and 100x more trading volume. Fat tails (Hill index 2.55) and volatility clustering (0.489) are the natural signature of adaptive agents.
Shipping delay dominates all other factors with super-exponential cost scaling: delay 2 costs $3,236; delay 5 costs $142,427 (44x). A single one-time demand change generates 30+ ticks of wild endogenous oscillation. The structure tax (44x from delay) dwarfs the human tax (1.6x from bounded rationality). Business cycles are structural, not exogenous.
At K=2 (Derrida parameter 1.010), the system maximizes information processing. At K=5, cascades engulf 93.4% of the network. The complexity catastrophe is scale-dependent: K=4 works at N=20 but paralyzes at N=150. Hierarchy cuts cascades from 63% to 36% at scale — it is computational infrastructure for managing complexity.
Cascade sizes follow a power law with exponent ~2.0, consistent with empirical extinction data. The system drives itself to criticality without tuning. Dense networks (conn. prob 0.15) produce perpetual catastrophe with mean cascade 56.9 and bimodal distribution. The connectivity-fragility paradox: more connections means more productive and more fragile simultaneously.
Pure exploitation yields peak 1.44 performance during stability but catastrophic collapse at disruption. Pure exploration yields steady 1.04 with perfect resilience. The tournament is a rigidity ratchet that purges adaptive capacity during stable periods. The 25% performance-resilience gap is the cost of insurance against an uncertain future. Diversity is not inefficiency.
41 comparisons between simulation outputs and published empirical data. 25 strong, 14 moderate, 2 weak matches.
| Model | Prediction | Real-World Evidence | Match |
|---|---|---|---|
| Stock Market | Fat tails (Hill exponent 2.55) | S&P 500 tail exponent 2.5–4.0 (Gopikrishnan et al. 1999, Cont 2001) | Strong |
| Volatility clustering (|return| AC = 0.49) | ARCH/GARCH effects in all equity markets (Engle 1982, Bollerslev 1986) | Strong | |
| Near-zero return autocorrelation | Real returns ~0 autocorrelation; simulation: -0.357 | Weak | |
| Volume-volatility correlation | Positive correlation documented universally (Karpoff 1987) | Moderate | |
| Endogenous wealth inequality (Gini 0.43) | US household wealth Gini ~0.85 (Wolff 2017); direction correct, magnitude lower | Moderate | |
| Excess kurtosis (42–211) | Empirical: 5–50 for daily returns (Cont 2001); simulation inflated | Moderate | |
| Beer Game | Bullwhip ratios 1.7–7.9x | Real supply chains: 1.0–8.0x (Lee et al. 1997, Cachon et al. 2007) | Strong |
| Super-linear cost scaling with delay | Convex lead-time cost documented (De Treville et al. 2004) | Strong | |
| Information sharing cuts cost 43% | VMI and POS sharing: 20–35% savings (Gavirneni et al. 1999) | Strong | |
| Bounded-rationality cost: 1.6–2.7x | Beer Game experiments: 2–3x median (Sterman 1989) | Strong | |
| Conservative ordering outperforms aggressive | Behavioral operations consensus (Schweitzer & Cachon 2000) | Strong | |
| Boolean Network | Phase transition at K=2, p=0.5 | Theoretical prediction confirmed; real gene networks K~2 (Aldana 2003) | Strong |
| Sharp phase boundary 2Kp(1-p)=1 | Derrida & Pomeau (1986) proof; computational confirmation | Strong | |
| Cascade bimodality at K=3 | Computational confirmation; limited direct empirical data | Moderate | |
| Hierarchy reduces cascades 30–40% | Qualitative support (Simon 1962, Perrow 1984) | Moderate | |
| Effective group size 5–9 | Hackman 2002, Dunbar 1992; extensive empirical support | Moderate | |
| Punctuated Eq. | Power-law cascade exponent ~1.8–1.9 | Fossil record extinctions 1.5–2.5 (Newman & Palmer 2003) | Strong |
| Dense networks are fragile | Financial network fragility (Haldane & May 2011) | Strong | |
| Large systems: rare but catastrophic events | Mass extinctions, financial crises, large-system outages | Strong | |
| Perpetual punctuated equilibrium cycles | Fossil record; organizational change (Gersick 1991) | Strong | |
| Rigids/Flex. | Tournament selection drives rigidity | Structural inertia theory (Hannan & Freeman 1984) | Strong |
| Optimal mix depends on volatility | Contingency theory (Burns & Stalker 1961) | Strong | |
| Exploration-exploitation tradeoff (25%) | March (1991); empirical confirmation (Uotila et al. 2009) | Strong | |
| El Farol Bar | Self-organization near threshold (57.3 vs 60) | Arthur (1994) original results; replicated in multiple frameworks | Strong |
| Prediction accuracy below 50% (self-referential paradox) | Consistent with minority game literature (Challet & Zhang 1997) | Strong | |
| Schelling Seg. | 30% threshold produces 75% segregation | Schelling (1971) original; confirmed computationally (Clark & Fossett 2008) | Strong |
| Non-linear threshold-segregation relationship | Empirical support from urban sociology (Bruch & Mare 2006) | Moderate | |
| Density affects dynamics more than outcomes | Computational confirmation; limited direct empirical tests | Moderate | |
| Predator-Prey | Endogenous boom-bust oscillations | Lynx-hare cycle (Elton & Nicholson 1942); microcosm experiments (Huffaker 1958) | Strong |
| Spatial refugia stabilize populations | Huffaker (1958) mite experiments; metapopulation theory (Levins 1969) | Strong | |
| Sand Pile | Power-law avalanche exponent ~1.2–1.3 | BTW theoretical prediction; confirmed experimentally (Frette et al. 1996) | Strong |
| Self-organization to criticality without tuning | SOC framework (Bak et al. 1987); observed in rice pile experiments | Strong | |
| No typical event size (max:median ~600:1) | Consistent with earthquake magnitude distributions (Gutenberg-Richter law) | Moderate | |
| Business Plans | Multiplicative fitness creates binding constraints | Complementarities in organizational design (Milgrom & Roberts 1995) | Moderate |
| Exploration-exploitation tradeoff in innovation rate | March (1991); empirical R&D intensity studies (Cohen & Levinthal 1990) | Weak | |
| Strategy | Adaptive portfolio strategy outperforms pure exploit/explore | March (1991); ambidexterity literature (O'Reilly & Tushman 2004) | Strong |
| Stable environments favor exploitation; volatile favor adaptation | Contingency theory (Burns & Stalker 1961; Hannan & Freeman 1984) | Strong | |
| Exploiter monocultures produce synchronized crashes | Competency trap (Levinthal & March 1993); industry concentration risk | Moderate | |
| Public Policy | Adaptive governance outperforms fixed regimes (+60% GDP) | Adaptive management literature (Holling 1978); experimental governance | Moderate |
| Entry barriers reduce growth more than tax rates | Ease of doing business (World Bank); firm entry studies (Djankov et al. 2002) | Strong | |
| Safety nets maintain evolutionary search population | Nordic model outcomes; active labor market policy (Card et al. 2010) | Moderate |
All 96 experiments were implemented as Python agent-based simulations, each with a command-line interface supporting parameter sweeps and JSON output. Simulations are deterministic and reproducible: every experiment uses seed 42, and all results can be regenerated from the CLI commands documented in each findings file.
Each model implements a specific mechanism from the complexity economics literature: the SFI stock market uses genetic algorithm evolution of forecasting rules; the Beer Game uses Sterman's empirically calibrated anchor-and-adjust heuristic; Boolean networks use Kauffman's random Boolean functions with the Derrida parameter; the punctuated equilibrium model combines Jain-Krishna autocatalytic networks with Bak-Sneppen extremal dynamics; the rigids-vs-flexibles model uses Harrington's tournament promotion; the Prisoner's Dilemma uses Axelrod's spatial evolutionary framework; Sugarscape follows Epstein & Axtell; technology evolution uses NK fitness landscapes; the integrated economy links four sub-models through composite health and dependency cascades; the El Farol Bar implements Arthur's inductive reasoning with competing prediction strategies; Schelling segregation uses threshold-based relocation on a torus grid; predator-prey compares ODE and spatial agent-based Lotka-Volterra dynamics; the sand pile implements the Bak-Tang-Wiesenfeld abelian sandpile for self-organized criticality; business plan evolution models Beinhocker's multiplicative fitness across Physical Technology, Social Technology, and Strategy on NK landscapes; the strategy evolution model tests portfolio vs. exploit vs. explore strategies across shifting market niches; and the public policy model compares five policy regimes (laissez-faire, social-democrat, innovation-state, protectionist, and adaptive) governing firms evolving on NK fitness landscapes.
Parameter sweeps were designed to isolate the effect of a single variable at a time (mutation rate, population size, connectivity, delay length, environmental stability), with the baseline configuration representing default or empirically calibrated values. Experiments range from 50 to 500 ticks depending on the model's convergence characteristics, with 200 ticks as the most common duration.
Empirical validation compares 41 simulation outputs against published measurements from peer-reviewed literature in finance, supply chain management, evolutionary biology, organizational science, network theory, urban sociology, and population ecology. Match quality is assessed as Strong (within empirical range), Moderate (correct direction, approximate magnitude), or Weak (qualitative match only).