The Convergent–Divergent Spectrum
The striking feature of financial markets is that two opposing forces, those that oppose the prevailing move and those that amplify it, coexist permanently. The statistical signature of this coexistence is the shape of the return distribution itself.
The Question This Series Left Unanswered
Throughout this research series, we built a case piece by piece. Episode 1 proved that markets carry memory: the autocorrelation of absolute returns persists for months in all sixty-eight markets, while raw return autocorrelation sits at zero. Episode 2 measured the depth of that memory using Hurst exponents and found every market deep in persistent territory, averaging H = 0.866. Episode 3 opened the tails and found the bell curve shattered: 5,791 times more five-sigma events than the Gaussian allows, with a mean tail exponent of 3.33. Episode 4 demonstrated that all three signatures form a single, universal fingerprint that does not respect asset class boundaries. Episode 5 proved causation: in a simulated world of pure noise, every signature vanished; when divergent and convergent agents were introduced, every signature returned. Episode 6 revealed that this was not a gradual effect. The transition between regimes is a phase transition, a sudden reorganisation of market behaviour at a critical threshold of roughly twenty-five to thirty percent divergent agent participation. Episodes 7, 8, and 9 closed the case: the fingerprint has persisted through four decades, five structural crises, and a complete technological revolution. The mechanism is feedback. The verdict is unanimous.
But one question lingered. If markets contain both convergent agents (who oppose the prevailing move) and divergent agents (who amplify the prevailing move), what happens as you sweep across the full spectrum of possible population mixes? Does the return distribution change gradually, or does it, too, exhibit the phase transition behaviour documented in Episode 6?
This appendix answers that question. And in doing so, it reveals something profound about the shape of every liquid market you have ever traded.
The Spectrum
Episodes 5 and 6 established the agent-based model with two populations: convergent agents who oppose the prevailing move, dampening price swings and pulling volatility back toward its unconditional level, and divergent agents who amplify it, chasing momentum and reinforcing trends through feedback. Episode 5 tested discrete configurations: feedback off, feedback on. Episode 6 swept divergent participation from zero to ninety-five percent and discovered the phase transition at twenty-five to thirty percent.
But those episodes held the convergent population fixed and varied the divergent fraction. They did not explore the full spectrum: what happens when you vary the balance between convergent and divergent agents across the entire range, from a market dominated entirely by agents opposing the prevailing move to one dominated entirely by agents amplifying it?
This appendix maps that spectrum. And it begins with a question about who, exactly, populates these markets.
Two Forces, One Market
Bouchaud and Potters observed that the vast majority of trades exert a directional force on price. A buyer pushes price up. A seller pushes it down. The question is whether that force reinforces the prevailing move or opposes it. A trend-following CTA adding to a winning position exerts divergent force. A pension fund rebalancing into an underweight asset exerts convergent force. Most market activity falls cleanly into one camp or the other.
A small minority of trades sit closer to neutral. A delta-hedged options position is designed to be insensitive to direction. A pairs trade offsets a long against a short. A volatility seller collecting theta profits from the passage of time, not from price movement. These trades exist, and they exert minimal net directional force on the underlying market.
But even these apparently neutral trades are not truly inert. The delta hedge must be rebalanced as price moves, and that rebalancing creates gamma exposure that can amplify or dampen volatility depending on positioning. Options market makers who are short gamma must buy into rallies and sell into declines, creating divergent feedback. Those who are long gamma do the opposite. The hedging flows from an options book can be the most powerful feedback mechanism in a market, as the portfolio insurance cascades of Black Monday demonstrated in 1987. The neutral trade, on close inspection, is rarely as neutral as it appears.
The concept of a “pure noise” market, one driven entirely by random, directionless trading, was the control case in Episode 5. It produced a perfect Gaussian: no memory, no fat tails, no persistence. It was useful as a theoretical baseline. But it does not describe any real market, because real markets are populated by participants whose trades overwhelmingly fall into one of two categories: those that oppose the prevailing move and those that amplify it.
This places every real market somewhere on a spectrum between two extremes:
At one end: 100% convergent. Every participant opposes the prevailing move. They buy into declines and sell into advances. Price swings are damped before they develop. Volatility is suppressed. Extreme moves are absorbed.
At the other end: 100% divergent. Every participant amplifies the prevailing move. They buy because price is rising and sell because price is falling. Moves feed on themselves. Trends persist. Volatility clusters. Extreme events multiply.
Real markets live somewhere between these extremes. The question is: what does the return distribution look like at each point along this spectrum?
The Experiment
We extended the agent-based model from Episodes 5 and 6. Where Episode 6 swept divergent agent participation from zero to ninety-five percent while maintaining a fixed convergent population, this appendix sweeps the balance between convergent and divergent agents across the entire spectrum in 10% increments, from 100% convergent through to 100% divergent. At each step, we simulated 150,000 daily returns (five independent runs of 30,000 days each) and measured the resulting distribution.
The model incorporates the key dynamics established in the series. Convergent agents dampen volatility shocks, pulling variance back toward its unconditional level. They also compress extreme returns, reducing the probability of tail events. Divergent agents do the opposite: they amplify shocks through GARCH-style volatility feedback, introduce serial correlation through momentum effects, and generate heavier-tailed innovations. The population mix determines the effective strength of each mechanism.
For each configuration, we measured four statistics: kurtosis (the peakedness and tail weight of the distribution), the percentage of observations beyond three standard deviations, the Hill tail exponent (a direct measure of tail heaviness), and the autocorrelation structure of absolute returns (the memory signature from Episode 1).
The Transformation
Figure A.1 shows the result. Each panel displays the return distribution at a different point along the spectrum, with the matched Gaussian distribution overlaid in dashed orange. The kurtosis value is displayed in the upper right of each panel.
Figure A.1: Return distributions across the convergent–divergent spectrum. Each panel shows the simulated return histogram (coloured) against the matched Gaussian (dashed orange). Kurtosis values increase nonlinearly as divergent agents gain population share.
Read the panels left to right, top to bottom, and watch the transformation unfold. In the convergent-dominated panels (top row), the distribution hugs the Gaussian closely. Kurtosis sits near 3.0, the Gaussian benchmark. Returns are well-behaved. Tail events are rare. This is a market where agents consistently oppose the prevailing move: price swings are corrected quickly, and the central limit theorem holds.
As divergent agents enter the population, the distribution begins to change. The peak sharpens. The tails extend. By the time you reach the 30/70 convergent-divergent split, the kurtosis has risen to 4.0 and the characteristic leptokurtic shape is clearly visible. In the bottom row, divergent-dominated configurations push kurtosis to 5.2, then 8.4, then 9.3. The distribution has been fundamentally reshaped.
But the critical observation is not that the distribution changes. It is how it changes.
The Phase Transition in the Distribution
Figure A.2 plots kurtosis and tail event frequency against the divergent agent fraction. If the transition were gradual, these curves would be approximately linear. They are not.
Figure A.2: Kurtosis (left) and tail event frequency (right) across the spectrum. Both metrics accelerate nonlinearly, consistent with the phase transition behaviour documented in Episode 6. The Gaussian baseline is shown in dashed orange.
The pattern is unmistakable. Kurtosis remains near Gaussian through the convergent-dominated regime, barely moving from 3.09 to 3.31 as the divergent fraction rises from 0% to 40%. Then it begins to accelerate. At 70% divergent, kurtosis reaches 4.0. At 80%, it jumps to 5.24. At 90%, it leaps to 8.44. The final step to 100% divergent adds another point to reach 9.31.
This is not a linear slide. It is a hockey stick. The system resists departure from Gaussian behaviour while convergent agents maintain critical mass, then departs rapidly once divergent agents exceed a threshold. This is precisely the phase transition signature documented in Episode 6, now visible in the distribution itself.
Why This Matters
The nonlinear response resolves an apparent tension in the series. Episode 6 documented sharp phase transitions in market dynamics. This appendix shows the population sweep is gradual (we move in 10% steps). But these are not contradictory. Think of heating water: the temperature rises gradually, but the phase transition from liquid to gas occurs at a critical threshold. Similarly, the population mix changes gradually, but the distribution transformation accelerates at a critical divergent fraction.
This is characteristic of systems near a critical point. Small changes in the control parameter (population mix) produce disproportionately large changes in the observable (distribution shape). The market is not a dial you turn smoothly. It is a system that flips.
Three Regimes, One Chart
Figure A.3 strips the sweep down to its essence by overlaying three representative configurations: pure convergent, mixed population, and pure divergent.
Figure A.3: Three regimes overlaid. Convergent-only (cyan) produces a tight, near-Gaussian distribution. Mixed population (gold, 30% conv / 70% div) produces the leptokurtic shape observed in real markets. Divergent-only (red) produces a wide, heavy-tailed distribution. The Gaussian reference is shown in dashed orange.
This is the figure that confirms what practitioners have observed for decades but struggled to explain. The leptokurtic distribution, the tall peak with fat tails that characterises every liquid market, is not a single phenomenon. It is two phenomena superimposed.
The tall peak is a convergent phenomenon. When convergent agents dominate, they oppose the prevailing move, absorbing swings and compressing returns toward zero. This creates the characteristic sharp peak that rises above the Gaussian reference. Most of the time, the dampening force wins. Returns cluster near zero.
The fat tails are a divergent phenomenon. When divergent agents gain temporary critical mass (through herding, momentum cascades, or panic), they amplify moves beyond anything the Gaussian predicts. These are the crashes, the rallies, the five-sigma events that occur thousands of times more frequently than the bell curve allows.
The leptokurtic shape exists because both forces coexist simultaneously in every liquid market. It is the signature of competing feedback mechanisms. Remove convergent agents and you lose the peak. Remove divergent agents and you lose the tails. Only when both are present do you get the distribution that every trader recognises but that the Gaussian model cannot produce.
The Volatility Signature
The distribution tells part of the story. The time series tells the rest. Figure A.5 displays absolute returns across 2,000 trading days for each of the three regimes.
Figure A.5: Absolute return series across regimes. Convergent-only (top) shows uniform, low-amplitude volatility. Mixed (middle) shows moderate clustering. Divergent-only (bottom) shows the dramatic volatility bursts and clustering observed in real markets.
Under pure convergent dynamics, volatility is uniform. Large and small moves are evenly distributed across time. There is no clustering, no persistence, no memory. This is the textbook random walk.
Under divergent dynamics, the picture transforms completely. Volatility arrives in bursts. Large moves cluster together, separated by periods of relative calm. This is the volatility clustering that Mandelbrot identified in cotton prices in 1963, that Engle formalised with ARCH in 1982, and that Episode 1 of this series documented across all 68 futures contracts. The mechanism is feedback: a large move triggers further large moves through the divergent agents who amplify momentum and chase shocks.
The Memory Signature
Episode 1 established that the autocorrelation of absolute returns is the fingerprint of feedback. Markets with feedback display persistent, slowly decaying autocorrelation in their volatility. Markets without feedback display none. Figure A.6 shows this signature across the three regimes.
Figure A.6: Autocorrelation of absolute returns. Convergent-only (left) shows no significant memory. Mixed (centre) shows weak short-range memory. Divergent-only (right) reproduces the slow, persistent decay observed in real market data — the same signature documented across all 68 contracts in Episode 1.
The convergent-only panel is empty. No memory. No persistence. The market forgets yesterday’s volatility immediately. This is precisely what the random walk hypothesis predicts and precisely what Episode 1 showed does not exist in any of the 68 real markets we tested.
The divergent-dominated panel reproduces the characteristic slow decay: high autocorrelation at short lags, declining gradually across dozens of lags, remaining statistically significant well beyond what any short-memory process could produce. This is the long-memory signature. It matches the empirical evidence from Episode 1, where the mean ACF(1) of absolute returns across all 68 contracts was 0.353.
The simulation confirms the mechanism. Volatility memory does not arise from complex microstructure or exotic stochastic processes. It arises from divergent feedback. Agents who amplify the prevailing move create persistence in volatility because today’s large move attracts further momentum trading, which generates tomorrow’s large move. The memory is endogenous. It is a direct consequence of the feedback structure.
The Numbers
Table A.1 summarises the full sweep. Four statistics tell the story: kurtosis measures the departure from Gaussian shape, tail events count the frequency of extreme moves, the Hill exponent quantifies tail heaviness directly, and standard deviation captures overall volatility.
Table A.1: Statistical summary across the convergent–divergent spectrum. Highlighted rows indicate departure from Gaussian behaviour (K > 4.5). Gaussian reference: K = 3.0, tail events = 0.27%.
The table confirms the visual evidence. Kurtosis trebles from 3.09 to 9.31 across the spectrum, but the increase is concentrated in the final three steps. Tail events beyond three standard deviations increase nearly sevenfold, from 0.26% to 1.78%. The Hill exponent drops from 5.85 (thin tails, near-Gaussian) to 3.12 (heavy tails, consistent with power-law behaviour). Every metric tells the same story: the departure from normality accelerates nonlinearly with divergent agent concentration.
What This Means for Markets
This appendix establishes three findings that extend the core arguments of the series.
First, the leptokurtic distribution is a dual-feedback signature. The tall peak and fat tails are not a single statistical anomaly. They are two distinct phenomena generated by two competing forces. The peak is convergent. The tails are divergent. This decomposition is not metaphorical. It is mechanistic. The agent-based model produces the leptokurtic shape only when both agent types are present.
Second, the phase transition documented in Episode 6 is visible in the distribution itself. The kurtosis does not rise linearly with the divergent fraction. It accelerates at a critical threshold, consistent with a phase transition. The market resists departure from Gaussian behaviour while convergent forces maintain critical mass, then transforms rapidly once that critical mass is lost.
Third, every liquid market sits in the mixed zone. Neither extreme (pure convergent or pure divergent) matches the empirical evidence from the 68 contracts analysed in this series. Real markets display moderate kurtosis (typically 4–15), significant but not overwhelming tail events, and persistent but bounded volatility memory. All of these are characteristics of the mixed-population regime. The market is a battlefield between two forces, and the leptokurtic distribution is the ceasefire line.
For Practitioners If you trade a liquid market, you are trading a system where convergent and divergent feedback coexist. The convergent force is why most days are unremarkable, small moves, mean reversion, business as usual. The divergent force is why some days are extraordinary, crashes, squeezes, momentum cascades that the Gaussian model says should not happen. Risk management that assumes Gaussian returns ignores the divergent force. It undercounts tail events by factors of thousands (as Episode 1 documented: 5,791 times more five-sigma events than the Gaussian predicts). Understanding that the distribution is shaped by competing feedback mechanisms is the first step toward building risk models that actually match the market you are trading. |
The Shape of All Markets
Every liquid financial market that has ever been studied, across every asset class, in every country, over every time period with sufficient data, displays the same leptokurtic return distribution. Tall peak. Fat tails. The shape has been documented so many times it is sometimes called a “stylised fact” of finance, a pattern so universal it is treated as axiomatic.
This appendix provides a mechanistic explanation for that universality. The leptokurtic shape is not accidental. It is not an artefact of aggregation, microstructure, or measurement error. It is the inevitable consequence of a market populated by participants who exert both convergent and divergent feedback. Since Bouchaud’s framework establishes that every trade either opposes or amplifies the prevailing move, and since every liquid market contains participants of both types, the leptokurtic distribution is not just common. It is necessary.
The tall peak tells you that most of the time, the market’s convergent forces win. They absorb shocks, dampen swings, and maintain order. The fat tails tell you that some of the time, divergent forces overwhelm convergent ones, and feedback cascades generate the extreme events that define financial history.
Both forces are always present. The distribution is their shadow.
Methodology Notes
The agent-based model uses a GARCH(1,1) framework where the population mix determines effective model parameters. Convergent agents reduce the GARCH alpha coefficient (shock sensitivity) and add mean-reversion in variance, pulling conditional variance back toward its unconditional level. They also dampen extreme returns through a threshold compression mechanism. Divergent agents increase alpha, increase persistence (beta), reduce the degrees of freedom of the innovation distribution (producing heavier-tailed shocks), and introduce serial correlation through momentum effects.
Each configuration was simulated five times with independent random seeds, generating 30,000 observations per run for a total of 150,000 daily returns per population mix. Kurtosis is reported as raw (non-excess) kurtosis where the Gaussian benchmark is 3.0. Hill tail exponents are estimated from the top 5% of absolute returns. All code is available in the companion repository for full replication.
The model is deliberately minimal. It captures the essential dynamics (convergent dampening, divergent amplification, GARCH volatility feedback) without unnecessary complexity. This parsimony is a strength: the leptokurtic distribution emerges from the simplest possible representation of competing feedback forces, suggesting it is a robust property of the mechanism rather than an artefact of model specification.
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.
Available now on Amazon in paperback, hardcover, and Kindle.