Why individual decisions do not simply add up
The model above is a defined research construction with a method of estimation. The educational point is the testable channel: participant constraints connect flows and prices. Do not turn it into “every inflow is always multiplied by the same number.”
Feedback on a hypothetical balance sheet
Suppose an organisation holds assets worth 100 and debt of 80. Equity is 20. If assets are revalued to 90 with debt unchanged, equity falls to 10: a 10% asset fall is a 50% equity fall. This is fictional balance-sheet arithmetic, not an assessment of an institution.
Add a lender rule that debt/assets may not exceed 80%. After revaluation, debt must be reduced by selling assets and using proceeds to repay it. With no costs or further price movement, selling 50 leaves assets 40 and debt 30, a 75% ratio. Selling 40 is enough to reach the boundary: assets 50 and debt 40. The action can be much larger than the initial loss.
This is not yet a market model. If many similar organisations sell together, specify how sales affect price. Positive feedback may arise when falling prices cause sales that intensify the fall; negative feedback when buyers strengthen as prices fall. The sign is a behavioural hypothesis, not a property of the word “system”.
Adaptation changes the old rule
Participants may notice a pattern and change behaviour. If many act alike, an observed relation can weaken, strengthen or change form. A test on the old regime therefore need not transfer to the new one. This does not mean every regularity disappears immediately; test stability assumptions and whether participants had the data in real time.
Draw a directed chain: price change, collateral value, cash requirement, forced sale, new price change. Label each arrow separately. Some links follow contract terms, others observations, and others remain hypotheses.
Heterogeneity matters more than an average participant
Two owners of the same asset may have different horizons, obligations and financing access. One need not sell today; another must meet a cash demand. Averaging them can erase the stress mechanism. Start with [the balance sheet](/guides/balance-sheet/) and obligation maturities, not a psychological portrait of an “average investor”.
Game theory can record possible responses, but equilibrium depends on payoffs, information and repeated interaction. MIT teaching materials treat these elements separately and do not forecast competition in every industry.[2] See [competition and cooperation](/guides/game-theory-competition/).
How to read a complex-market model
Ask for the unit of analysis, balance constraints, price rule and out-of-sample check. Which parameters are estimated and which are illustrative? Does the model reproduce several observations at once, and are there alternative explanations? A striking picture is insufficient; a useful model makes conditional claims and shows when they fail.
What happens when one link is removed
Temporarily remove one arrow: suppose sales do not affect price, or collateral requirements do not depend on revaluation. Recalculate. If the conclusion remains, that link may not cause the effect; if it changes sharply, its estimate needs especially strong evidence.
This does not reproduce a real market, but disciplines explanation. State which participants are excluded, which prices are external and which constraints are fixed. The more explicit these choices, the easier it is to identify an observation that could disprove the explanation.
El Farol: when a common forecast cancels itself
In Brian Arthur’s El Farol attendance model, agents decide whether to attend a capacity-limited venue using a forecast of attendance. The forecast affects attendance, so one shared belief can destroy its own basis.[RCS1] For an exercise, use ten agents and comfortable capacity six. Each gets +1 when attendance is at most six, −1 when overcrowded and 0 when staying home. These are author-created payoffs, not a market estimate.
If all forecast five and attend, attendance is ten and each gets −1. If all forecast eight and stay home, attendance is zero and each gives up a possible gain. Forecast accuracy is therefore inseparable from the behaviour rule. Different methods—last observation, three-period average, alternation—may sustain diversity but do not guarantee comfort or equilibrium.
For a reproducible experiment specify initial history, methods per agent, error score, tie-break and update time. After each round score methods using actual attendance, then choose the next method. Compare with independent random attendance at probability 0.6 using a stated metric: mean payoff, overcrowding or dispersion. Mean attendance six does not mean no round was overcrowded.
Fitness landscape: a local peak is not global
In NK models, fitness is assigned to feature combinations and one feature’s contribution depends on the others. Coupled landscapes let one participant’s action change another’s conditions.[RCS2] This is model language, not biological proof of a firm’s future success.
Take two binary choices: technology and distribution channel. An artificial annual table is 00→8, 10→6, 01→5, 11→12. Neighbours differ in one choice. From 00 every single change is worse, so local improvement stops at 8, while the global maximum is 12 at 11. Reaching 11 through 10 requires temporarily accepting 6. This is not advice to tolerate losses: budget, completion probability and switching cost are omitted.
With independent contributions this incompatibility would not occur. Interaction makes improving each part alone a poor route to improving the whole. If the environment changes so 11 falls to 4 while 00 stays 8, yesterday’s peak is no longer best without changing the decisions. Comparing firms only by current profit misses transition cost and durability.
Simple rules, selection and preserving options
A simple rule needs input, action, constraint and test. Example: each round test one untried neighbour; keep it only if result improves after costs; spend no more than two reserve units on search. It is testable but will not find 11 from 00 in the table. An alternative permits one two-component experiment; compare its advantage with failed-search cost, not only its best endpoint.
Make selection explicit. Two strategy groups have shares 0.5 and 0.5 and positive fitness results 2 and 1. If next share is proportional to current share times result, the first becomes 2/3. If results are 0.5 and 2, it becomes 1/3. This models weight redistribution, not the quality of every survivor.
Exploration preserves several options; exploitation uses a known one. Long-term strategy is a budget of experiments, measurement horizon and stopping condition. A binary tree of depth d has 2^d leaves: at d=10, 1,024; at d=20, 1,048,576. Deeper search is costly before unknown environmental rules are counted.
A short change affects one component; a long change affects several. The latter may bypass a local peak but creates more unknowns and financing needs. Fixed results, slow updating and fast competitor response are different regimes. State permitted actions, limits, priorities, review date and exit condition. Decentralisation without necessary information exchange is not automatically adaptation.
The option to defer a project is separate from its main cash flow. In an author-created two-scenario model, immediate launch pays 30 or −10 with equal probability, expectation 10. Free information before launch, with refusal in the bad state, raises expectation to 15. A delay cost of 6 reduces it to 9, so waiting is worse. This is conditional option flexibility, not a universal DCF substitute.
Attendance, landscape and selection models answer different questions; do not merge them under “evolution”. A chess analogy does not prove human intuition beats computation, and a neuroscience story does not establish strategy returns. Transfer requires observable payoffs, update rules and a simple control.
Sources and boundaries
[1] [In Search of the Origins of Financial Fluctuations: The Inelastic Markets Hypothesis](https://www.nber.org/papers/w28967): a research model of flows and prices, not a universal law.
[2] [Lecture Notes: Game Theory for Managers](https://ocw.mit.edu/courses/15-040-game-theory-for-managers-spring-2004/pages/lecture-notes/): equilibrium, repeated games, information and incentives; not a company description.
[RCS1] W. Brian Arthur, “Inductive Reasoning and Bounded Rationality,” source cr-elfarol-arthur. The endogenous-attendance model is sourced; the numerical example and protocol are original.
[RCS2] Stuart A. Kauffman and Sonke Johnsen, “Coevolution to the Edge of Chaos,” source cr-fitness-kj. Formal interacting-landscape model; not financial validation.