The Vault

THE FRACTALS OF FINANCE | Research Series | Episode 1 of 9

The Random Walk Is Dead

How we tested the most fundamental assumption in finance, and watched it fail in every market on earth.

For more than a century, the dominant theory in finance rested on a single, elegant assumption: price changes are independent. Yesterday tells you nothing about tomorrow. The market has no memory. Each tick arrives clean, uncontaminated by what came before.

This idea was not merely academic. It was foundational. It supported the Efficient Market Hypothesis, the Black-Scholes options framework, the Capital Asset Pricing Model, and the Value at Risk standards that governed institutional risk management across the world. Trillions of dollars in capital allocation, regulation, and portfolio construction depended on this one premise.

Price changes are independent.

We tested that assumption across sixty-eight futures markets spanning eight asset classes, six continents, and forty-one years of daily data. The results were not ambiguous. They were not marginal. They did not require sophisticated interpretation.

The assumption failed in every single market.

The Comfortable Fiction

The random walk hypothesis entered finance through Louis Bachelier in 1900, gained mathematical formalism through Paul Samuelson in the 1960s, and achieved near-universal acceptance through Eugene Fama’s articulation of the Efficient Market Hypothesis in 1970.¹

The logic was appealing. If markets are efficient, then all available information is already reflected in price. New information arrives unpredictably. Therefore price changes must be unpredictable. Therefore returns must be independent. Therefore the past cannot inform the future.

The conclusion seemed to follow naturally from the premise. And for decades, it appeared to hold. If you measured the autocorrelation of daily returns, the statistical relationship between today’s return and tomorrow’s, you found almost nothing. Returns looked random. The independence assumption appeared safe.

But this test was asking the wrong question.

Measuring the autocorrelation of raw returns tests whether direction is predictable. It asks: if the market went up today, does it go up tomorrow? The answer, broadly, is no. Direction carries almost no serial dependence. On this point, the random walk appeared correct.

What nobody was measuring, or at least what nobody was broadcasting, was the autocorrelation of magnitude. Not the direction of returns, but their size. Not whether the market went up or down, but whether it moved violently or quietly.

That question produces a completely different answer.

The Test

We assembled daily returns from sixty-eight continuous futures contracts provided by Commodity Systems Incorporated, spanning September 1984 to January 2026. The dataset covers equities, bonds, currencies, energy, metals, grains, softs, and meats: every major tradeable asset class on earth. In total, the dataset comprises 647,922 clean trading days after quality filtering.²

For each market, we computed two autocorrelation functions out to 252 lags, one full trading year.

The first measured autocorrelation of raw returns. This is the standard test. It asks whether today’s return predicts tomorrow’s return, or the return ten days later, or fifty days later, or two hundred days later.

The second measured autocorrelation of absolute returns. This asks a different question entirely. It asks whether today’s magnitude, regardless of direction, predicts tomorrow’s magnitude, and next week’s, and next month’s. It measures whether the market remembers how large its movements were.

If markets truly follow a random walk, both measures should hover near zero at every lag. Independence means no memory of any kind. Not in direction. Not in magnitude.

One of these measures obeyed the theory. The other destroyed it.

The Evidence

Figure 1.1: Autocorrelation functions for three markets drawn from different asset classes: the S&P 500 (equities), Gold (metals), and Soybeans (grains). The left column shows autocorrelation of raw daily returns. The right column shows autocorrelation of absolute daily returns. Each plot extends to 252 lags, one full trading year. The dashed red lines mark the 95% significance boundary. In the left column, raw return autocorrelation hugs zero at every lag. The market shows no memory of direction. In the right column, absolute return autocorrelation begins near 0.5 and decays slowly over hundreds of days. The market remembers its own intensity for months. The contrast is stark. Direction is memoryless. Magnitude is not. This pattern appears in all three markets despite their having nothing in common: different exchanges, different participants, different fundamentals, different continents.

The left side of Figure 1.1 is what finance expected. Raw returns show essentially zero autocorrelation. The mean across all sixty-eight markets is effectively zero. Direction is memoryless. On this point, the efficient market hypothesis holds.

The right side is what finance ignored.

Absolute returns show massive, persistent autocorrelation. The mean across all sixty-eight markets at lag one is 0.35, more than two orders of magnitude larger than the raw return autocorrelation. And it does not vanish after a few days. It persists across weeks, months, and into a full trading year. At lag 252, one year later, the autocorrelation of absolute returns is still positive and statistically significant in the majority of markets.

The market forgets where it went. It never forgets how hard it moved.

This is not a subtle finding. It is not a marginal statistical artefact that requires careful interpretation. The contrast is visible to the naked eye. Place the two charts side by side and the conclusion is immediate. The left panel is noise. The right panel is structure.

The phenomenon is known in the academic literature as volatility clustering: the tendency for large moves to follow large moves and small moves to follow small moves. It was first documented empirically by Mandelbrot in 1963 and formalised by Engle’s ARCH model in 1982 and Bollerslev’s GARCH in 1986.³

Yet its full implications were never absorbed by the mainstream frameworks that continued to assume independence.

What the autocorrelation of absolute returns reveals is memory. Not the kind of memory that allows prediction of direction. Something deeper. The market remembers its own state of excitation. A violent day leaves an imprint that persists for months. A quiet period suppresses activity long after the initial calm. The system carries forward its own history in a way that the random walk explicitly forbids.

Sixty-Eight Markets. One Pattern.

The natural objection is specificity. Perhaps the S&P 500 has memory because of its size, its derivative markets, its algorithmic participants. Perhaps Gold retains memory because of central bank activity. Perhaps Soybeans are special because of seasonal planting cycles.

We tested every market in our universe. Every single one.

Figure 1.2: Autocorrelation functions of absolute returns for all sixty-eight futures markets, overlaid on a single chart. Each thin line represents one market, coloured by asset class: blue for equities, green for bonds, gold for currencies, red for energy, grey for metals, brown for grains, pink for softs, and orange for meats. The bold navy line is the cross-market average. Every line begins well above zero and decays slowly across 252 lags. Not a single market hugs the zero axis. The cross-market average starts above 0.35 at lag one and remains above 0.15 at lag 100. The visual message is immediate: this is not a phenomenon specific to equities, or to developed markets, or to liquid markets. It appears universally. In Orange Juice, in the Japanese Yen, in Lean Hogs, in the Euro Bund, in Crude Oil, and in every other market we tested.

Figure 1.2 is the image that ends the argument.

Sixty-eight autocorrelation curves, one for each market, all arcing upward from zero and decaying slowly across a full trading year. Not a single curve hugs the zero axis. Not one market behaves as the random walk requires.

One hundred percent of markets show more than twenty statistically significant lags in absolute return autocorrelation. Ninety-nine percent show more than one hundred significant lags. The weakest memory in the dataset still extends across months. The strongest extends beyond a year.

These markets share nothing. Soybeans and sovereign bonds are governed by entirely different fundamentals. Rubber and the Russell 2000 have no common participants. Cotton and the Canadian Dollar operate in separate economic universes. Live Cattle and the Long Gilt respond to different forces entirely.

Yet they all carry the same fingerprint.

The only factor these markets share is structure. Every one of them is a complex adaptive system in which participants condition their behaviour on recent prices. Trend followers buy because price rose. Stop losses trigger because price fell. Margin calls force liquidation because volatility spiked. Algorithms amplify because signals fired. Human psychology anchors because the recent past feels more real than the distant past.

The common factor is feedback.

And feedback creates memory.

What This Means

If markets carried no memory, volatility would be stable. Large moves would be followed by average moves. Quiet periods would contain the same probability of disruption as turbulent ones. Risk would be constant and measurable. The bell curve would describe reality. Portfolio construction based on stable variance would work.

None of this is true.

Markets carry deep, persistent memory in their volatility structure. This memory means that risk is not constant. It clusters, builds, and releases in patterns that extend across months. It means that the calm periods are not safe. They are accumulations of stored energy. It means that the turbulent periods are not anomalies. They are the system expressing what it has been carrying.

The random walk is not approximately wrong. It is structurally wrong. It describes a world that has never existed in any market, in any asset class, on any continent, across any period of the last four decades.

The random walk is dead.

The question that follows is the one this series exists to answer. If markets carry memory, what creates it? If volatility clusters, what mechanism drives the clustering? If every market on earth shares the same fingerprint despite sharing nothing else, what common force is responsible?

We believe the answer is feedback. And we intend to prove it.

Episode 1 established that markets carry memory. Episode 2 asks the harder question: how long does that memory last? Using Hurst exponents, a tool originally designed to measure the persistence of flooding on the Nile, we quantify the depth of market memory across all sixty-eight markets. The results will show that every market on earth sits deep in persistent territory, far from the random walk baseline.

The market never forgets. Episode 2 measures how far back it remembers.

Endnotes

References

Louis Bachelier, “Théorie de la spéculation” (1900), introduced the random walk model for speculative prices. Paul Samuelson formalised the connection between efficiency and martingale behaviour in “Proof That Properly Anticipated Prices Fluctuate Randomly,” Industrial Management Review, 1965. Eugene Fama’s “Efficient Capital Markets: A Review of Theory and Empirical Work,” Journal of Finance, 1970, established the Efficient Market Hypothesis as the dominant paradigm.
Data sourced from Commodity Systems Incorporated (CSI). All contracts are ratio (proportional) back-adjusted continuous futures, which preserves percentage returns across contract rolls. The universe comprises 68 contracts: 13 equity indices, 10 bonds, 8 currencies, 8 energy, 8 metals, 8 grains, 8 softs, and 5 meats. The sample spans September 1984 to January 2026. Returns are computed as log returns: r_t = ln(P_t / P_{t-1}). Quality filters exclude zero-volume days and flag limit-move observations. Total clean observations: 647,922 trading days across all contracts.
Benoit Mandelbrot, “The Variation of Certain Speculative Prices,” Journal of Business, 1963, first documented fat tails and volatility clustering in cotton prices. Robert Engle, “Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation,” Econometrica, 1982, introduced the ARCH model. Tim Bollerslev, “Generalized Autoregressive Conditional Heteroscedasticity,” Journal of Econometrics, 1986, extended ARCH to the GARCH framework. These models capture volatility clustering but do not explain its cause.

Methodology

Autocorrelation Function (ACF). For a return series {r_t} of length N, the sample autocorrelation at lag k is computed as: ACF(k) = [Σ(r_t – μ)(r_{t+k} – μ)] / [Σ(r_t – μ)²], where μ is the sample mean and summation runs over available overlapping observations. We compute ACF for both raw returns r_t and absolute returns |r_t| at lags 1 through 252. Under the null hypothesis of independence (white noise), ACF values are approximately normally distributed with mean zero and standard deviation 1/√N. We use the 95% significance threshold of ±1.96/√N.
The contrast between flat raw-return ACF and slowly-decaying absolute-return ACF is one of the most robust “stylised facts” in empirical finance. It was systematically documented by Rama Cont in “Empirical Properties of Asset Returns: Stylized Facts and Statistical Issues,” Quantitative Finance, 2001, and by Zhuanxin Ding, Clive Granger, and Robert Engle in “A Long Memory Property of Stock Market Returns and a New Model,” Journal of Empirical Finance, 1993. Ding, Granger and Engle showed that autocorrelation of |r_t|^d is maximised at d ≈ 1 (absolute returns) and persists over very long lags, exhibiting power-law decay consistent with long-range dependence.
The power-law decay of absolute return autocorrelation, ACF(|r|, k) ~ k^(2H-2) where H is the Hurst exponent, suggests fractional integration or long memory processes. This is explored further in Episode 2. For foundational work on long memory in financial time series, see: Andrew Lo, “Long-Term Memory in Stock Market Prices,” Econometrica, 1991; Clive Granger and Zhuanxin Ding, “Some Properties of Absolute Return: An Alternative Measure of Risk,” Annales d’Économie et de Statistique, 1995; and Stephen Taylor, Modelling Financial Time Series, Wiley, 1986.
The concept of volatility clustering, that large changes tend to be followed by large changes and small changes by small changes, was first noted by Mandelbrot (1963) and has since been confirmed in virtually every financial market studied. Our finding of universal absolute-return memory across 68 diverse futures markets extends this literature by demonstrating the phenomenon with a breadth of coverage rarely seen in published studies. Most existing work examines 5 to 20 markets, typically within a single asset class.

Figures

Figure 1.1: Autocorrelation functions computed for S&P 500 E-mini (ES), Gold (GC2), and Soybeans (S2) using the full available sample for each contract (approximately 10,400 observations each). Left panels show ACF of raw log returns; right panels show ACF of absolute log returns. Lags 1 to 252. Red dashed lines indicate ±1.96/√N significance bounds.
Figure 1.2: Autocorrelation functions of absolute returns for all 68 contracts in the dataset, computed at lags 1 to 252. Lines coloured by asset class. Bold navy line represents the unweighted cross-market mean ACF at each lag. All individual curves begin above the significance threshold and remain positive across the full lag range.

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.