The Vault

THE FRACTALS OF FINANCE | Research Series | Episode 5 of 9

The Null World

We built a market with no feedback. Every signature vanished. We turned feedback on. Every signature returned.

Over four episodes, we have built a circumstantial case. Markets carry memory. The memory is deep. The tails are fat. The fingerprint is universal. Every line of evidence points toward the same explanation: feedback between participants and price creates the statistical signatures that define real markets.

But circumstantial evidence, no matter how voluminous, is not proof. Correlation is not causation. The observation that memory, fat tails, and persistence always appear together does not, by itself, demonstrate that feedback is their cause. It is possible, in principle, that some other mechanism produces the same pattern.

To move from observation to proof, we need a controlled experiment. We need to build a world without feedback and test whether the fingerprint survives.

If feedback is the cause, removing it should remove the effect.

The Experiment

We built an agent-based model of a financial market. The model is deliberately simple. It contains a population of traders who buy and sell a single asset. Price is determined by the balance of supply and demand. Orders flow into the market, price adjusts, and the cycle repeats across thousands of simulated trading days.¹

The model allows us to control one variable with precision: the proportion of traders who condition their behaviour on recent price. The model contains two populations that exist in every real market. Divergent agents, the trend followers, condition their behaviour on recent price. They create positive feedback by allowing the output of the system, price, to become an input to their decisions. Convergent agents, the value investors and mean reverters, trade against recent moves. They create negative feedback, pulling price back toward equilibrium.

The null world contains no agents at all. It is a pure random process: independent, identically distributed returns drawn from a Gaussian distribution. This is the world that traditional finance assumed. A world of pure noise, where price changes carry no information about future price changes.

By adjusting the mix between divergent and convergent agents, we can control the balance of feedback in the system. A world with only convergent agents has negative feedback alone. Add divergent agents and positive feedback enters. The model gives us a dial that controls the intensity of the mechanism we suspect is responsible for everything we have observed.²

We then run the same statistical battery on each simulated world that we ran on sixty-eight real markets: autocorrelation of absolute returns, Hurst exponents, kurtosis, tail exponents, and sigma event counts.

The question is simple. Does the fingerprint survive when feedback is removed?

A World Without Feedback

We begin with the null world. Pure noise. No agents, no trades, no microstructure. Returns are drawn independently from a Gaussian distribution. No feedback of any kind. Every return is independent of what came before.

This is the null world. The world that traditional finance assumed was real.

The results are immediate and absolute.

The autocorrelation of absolute returns: zero. No memory.

The Hurst exponent: 0.53. Indistinguishable from the random walk value of 0.5. No persistence.

Excess kurtosis: negative 0.2. The distribution is Gaussian by construction. No fat tails.

Five-sigma events: zero. None. Across the entire simulation. The tails are exactly as thin as the bell curve predicts.

Every signature vanishes.

In a world without feedback, markets produce exactly the behaviour that traditional finance assumed: independent returns, Gaussian tails, no memory, no clustering, no persistence. The random walk is not wrong as a description of this world. It is a perfect description of this world.

This world does not exist.

Turning Feedback On

Now we populate the market. We introduce a realistic ecology of agents: convergent agents who fade recent moves and anchor to fundamental value, and divergent agents who follow recent trends. Thirty percent of the population are divergent. Seventy percent are convergent.

The effect is not gradual. It is not subtle. It is a phase transition.

The autocorrelation of absolute returns leaps from zero to 0.287. Memory appears instantly. It is not a faint signal buried in noise. It is massive, persistent, and structurally identical to what we observe in real markets.

The Hurst exponent leaps from 0.53 to 0.609. Persistent memory, well above the random walk value of 0.5.

Excess kurtosis leaps from zero to 27. The distribution transforms from Gaussian to fat-tailed, with tails extending far beyond the bell curve.

Five-sigma events: thirteen per ten thousand days. In the same simulation length that produced zero under pure noise.

Every signature reappears. All of them. Simultaneously.

Figure 5.1: Two panels showing the autocorrelation of absolute returns in simulated markets. The left panel, labelled “Feedback OFF,” shows the result for the null world (pure noise, blue) and a convergent-only market (purple): both produce flat or near-flat lines. No memory. No persistence. A dead signal. The right panel, labelled “Feedback ON,” shows the result when trend followers are introduced. The green line (30% divergent agents) shows positive autocorrelation that decays slowly across hundreds of lags. The dark red line shows the average across all 68 real markets for comparison. The simulated feedback-on world and the real world produce the same shape. The contrast between the two panels is absolute. The left panel is silence. The right panel is structure. The only difference between them is feedback.

The left panel of Figure 5.1 is the most important null result in this series. It shows what a market looks like when feedback is absent: a dead flat line. No memory at any lag. No structure at any timescale. The autocorrelation function of a pure noise process, exactly as the random walk theory predicted.

The right panel shows what happens the moment feedback enters the system. The flat line erupts into a structured, slowly decaying function that persists across hundreds of lags. The shape matches real markets. The mechanism is feedback and nothing else.

Three Worlds

Figure 5.2 places three return distributions side by side. The null world. The feedback world. The real world.

Figure 5.2: Three panels showing return distributions on identical logarithmic probability scales, pooled across ten independent simulation runs of 10,000 trading days each. The left panel, “No Feedback (Simulated),” shows a distribution that hugs the Gaussian curve perfectly. Kurtosis is zero. Five-sigma events: zero. The tails follow the Gaussian curve precisely. This is the null world. The centre panel, “Feedback ON (Simulated),” shows a radically different shape. The distribution develops heavy tails stretching far beyond the Gaussian boundary. Kurtosis is 27.1. Five-sigma events: 13 per 10,000 days. The tails persist where the bell curve says they cannot. This is the world with feedback. The right panel, “S&P 500 (Real),” shows actual market data. Kurtosis is 16.4. Five-sigma events: 37. The shape of the real distribution shares the same fat-tailed structure as the feedback world. The no-feedback world looks nothing like reality. The feedback world looks like reality. The conclusion is visual and immediate.

Look at the three panels. The left panel is smooth, symmetric, and obedient. It follows the Gaussian curve with textbook precision. It is the world traditional finance imagined.

The right panel is peaked, heavy-tailed, and rough. It defies the Gaussian at every extreme. It is the real world.

The centre panel, the simulated world with feedback, looks like the right panel. Not like the left.

Feedback creates reality. Its absence creates a fiction.

What the Experiment Proves

The logic of this experiment is the logic of controlled causation.

We held everything constant except one variable: feedback. When feedback was absent, in a world of pure noise with no agents, every signature of real market behaviour vanished. When feedback was present, through the introduction of divergent and convergent agents, every signature returned.

This is not correlation. This is a controlled manipulation. We changed one input and observed a total transformation in the output. Memory, fat tails, persistence, and volatility clustering all appear and disappear together, controlled by a single switch: the presence or absence of feedback.³

The null hypothesis, that the fingerprint can exist without feedback, is rejected. In a world without feedback, the fingerprint does not merely weaken. It does not partially appear. It vanishes entirely. Every trace of it disappears.

The alternative hypothesis, that feedback causes the fingerprint, is supported. When feedback is introduced, the fingerprint appears in full, with all signatures present simultaneously, in a model that contains no fundamental information, no macroeconomic shocks, no news, no central banks, no earnings announcements, and no external events of any kind.

Feedback alone is sufficient to produce every signature we observe in real markets.

Nothing else is required. No external shocks. No complex fundamental dynamics. No elaborate microstructure. Just participants observing price and reacting to it. The simplest possible form of feedback, applied to the simplest possible market, reproduces the statistical DNA of sixty-eight real markets across eight asset classes and four decades.

What This Means

The experiment resolves the central question of this series.

Episodes 1 through 4 documented the fingerprint: memory, persistence, fat tails, universality. Those episodes showed what exists. Episode 5 shows why it exists.

Markets are not random. They carry deep, persistent memory in their volatility structure. Their tails are fat, populated by extreme events that the bell curve declares impossible. These signatures appear universally across every market on earth.

The cause is feedback. The observation of price by participants, and the conditioning of their behaviour on that observation, creates a self-reinforcing loop that generates every signature simultaneously. Remove the loop and the signatures vanish. Restore the loop and the signatures return.

This finding has consequences that extend far beyond academic interest. If feedback is the engine that drives market behaviour, then every model, every risk framework, and every portfolio construction method that assumes independence is built on a foundation that does not exist. The random walk was not a simplification. It was a denial of the mechanism that produces the phenomena it failed to explain.

The null world is the world finance assumed. The feedback world is the world that exists.

Episode 5 demonstrated that feedback is sufficient to produce the fingerprint. Episode 6 asks a more precise question: what happens as feedback intensity changes? We sweep the dial from zero to maximum and watch the signatures evolve. The results reveal something remarkable. The fingerprint does not appear gradually. It emerges through a phase transition, a critical threshold beyond which the system transforms completely.

How much feedback does it take to break the random walk? Less than you think.

Endnotes

References

The agent-based model follows the heterogeneous agent framework established by William Brock and Cars Hommes, “Heterogeneous Beliefs and Routes to Chaos in a Simple Asset Pricing Model,” Journal of Economic Dynamics and Control, 1998; and Thomas Lux and Michele Marchesi, “Scaling and Criticality in a Stochastic Multi-Agent Model of a Financial Market,” Nature, 1999. Our implementation uses 1,000 agents divided into three populations. Fundamentalists submit orders based on deviation from fundamental value with Gaussian noise (negative feedback). Divergent agents (trend followers) compute an exponentially weighted moving average of recent returns across heterogeneous lookback periods (5, 10, 20, 35, 50, 70, 90, and 120 days), with agents distributed in log-weighted proportion favouring shorter horizons. Their demand is proportional to signal strength (positive feedback). Convergent agents (mean reverters) use the same heterogeneous lookback structure but trade against recent returns (negative feedback). The null world is a pure Gaussian random process with no agents. Price adjusts proportionally to net order flow: P_t = P_{t-1} x exp(lambda x net_demand / N). Each simulation runs for 10,000 trading days, averaged across 20 independent runs per configuration. Full source code is provided in fractals_abm.py for independent replication.
The configurations tested: (a) Null world: pure Gaussian noise, no agents. (b) Convergent only (0% divergent): fundamentalists and mean reverters, negative feedback only. (c) 30% divergent: moderate trend-following alongside 70% convergent agents. The comparison between null, convergent-only, and divergent configurations demonstrates that the fingerprint requires specifically positive (divergent) feedback, not merely the presence of agents trading on price.
The logic of causal inference through controlled manipulation is standard in experimental science. The key requirements are: (a) manipulation of a single independent variable (feedback intensity), (b) observation of a dependent variable (statistical signatures), (c) control of confounding factors (constant agent count, noise levels, simulation parameters), and (d) replication (20 runs per configuration with reported standard deviations). Our experiment satisfies all four criteria. For discussion of agent-based modelling as a tool for causal inference in economics, see: J. Doyne Farmer and Duncan Foley, “The Economy Needs Agent-Based Modelling,” Nature, 2009; and Blake LeBaron, “Agent-Based Computational Finance,” Handbook of Computational Economics, 2006.
The convergent-only configuration (0% divergent agents) produces: ACF(1) of absolute returns = 0.253, Hurst exponent = 0.559, excess kurtosis variable across runs (median near zero), five-sigma events near zero. While convergent agents create some transient memory at short lags from their mean-reversion activity, they do not produce the persistent, slowly-decaying memory structure observed in real markets. Negative feedback (contrarian behaviour) acts as a damping mechanism, suppressing the extreme events and tail behaviour that characterise real markets. This is consistent with the theoretical prediction of Brock and Hommes (1998).

Methodology

Agent-based model specification. The model contains N = 1,000 agents in three populations: fundamentalists, divergent agents (trend followers), and convergent agents (mean reverters). Fundamentalists trade toward a fixed fundamental value with demand proportional to price deviation plus Gaussian noise. Divergent agents compute an exponentially weighted signal from recent returns across eight heterogeneous lookback windows (5 to 120 days), with demand proportional to signal strength. Convergent agents use the same heterogeneous lookback structure but with inverted signals, fading recent returns. The null world is a pure Gaussian i.i.d. process with no agents. The market-clearing price adjusts proportionally to net order flow: P_t = P_{t-1} x exp(lambda x net_demand / N), where lambda is a price impact parameter. Full source code is provided in the supplementary file fractals_abm.py for independent replication.
All statistical tests were applied identically to simulated and real data: ACF of raw and absolute returns (lags 1 to 252), Hurst exponents via R/S method, excess kurtosis, Hill tail exponent with k = sqrt(N), and sigma event counts at thresholds 3 through 7. Each configuration was simulated 20 times with different random seeds. Reported values are means across the 20 runs. Standard deviations are reported in the supplementary data and are small relative to effect sizes, confirming the robustness of the findings.

Figures

Figure 5.1: Left panel shows ACF of absolute returns for the null world (blue, pure noise) and the convergent-only market (purple). Both produce flat or near-flat autocorrelation, with convergent agents showing only brief transient memory at lag 1. Right panel shows ACF for the 30% divergent configuration (green) alongside the real 68-market average (dark red). The divergent configuration produces persistent memory decaying slowly over hundreds of lags, matching the shape of real market memory.
Figure 5.2: Return distributions on log-probability scale for three worlds, each pooled across ten independent simulation runs: (1) No feedback simulated, kurtosis = 0.0, zero 5-sigma events. (2) Feedback ON simulated (30% trend followers), kurtosis = 27.1, 13 five-sigma events per 10,000 days. (3) S&P 500 real data, kurtosis = 16.4, 37 five-sigma events per 10,000 days. The no-feedback distribution matches the Gaussian curve. The feedback and real distributions share the same heavy-tailed structure.

This research series is drawn from The Fractals of Finance: Determinism, Adaptation and the Geometry of Markets

The book explores the full architecture of feedback, fat tails, and fractal structure in financial markets, and what it means for how we trade, invest, and understand risk.

Available now on Amazon in paperback, hardcover, and Kindle.

Want a practical field manual for trading trends and capturing outliers?

The Aussie Turtles Trend Following Guide: A Field Manual for Hunting Outliers adapts the timeless principles of the original Turtle traders into a systematic, rules-based approach for modern markets. Co-authored with Adam Havryliv.