Do Retail Investors Make Markets Multifractal? A Dual-Listed Test

Stock returns are not themselves predictable day to day, but their size is: wild days cluster with wild days, and this memory fades slowly, as a power law. Beyond that, markets are multifractal, meaning small and large price moves follow different scaling rules. This is usually summarized by a "spectrum width," a single number for how unevenly roughness is spread across move sizes. Researchers routinely attribute the wider spectra of Asian markets to their large share of individual, retail investors. Nobody has tested that attribution in a way that separates investor mix from everything else that differs between markets. There is also a sharper worry. Retail-heavy markets are full of policy hype cycles and synchronized panics. Those produce sudden jumps in the volatility level, and such jumps are known to fake long memory. So the "retail multifractality" in the literature might be an artifact of these jumps rather than a property of how retail traders behave.

We propose to settle this with Chinese companies listed both in Shanghai/Shenzhen (A-shares, dominated by retail investors) and in Hong Kong (H-shares, dominated by institutions). The two listings share one set of cash flows, news and management. They differ mainly in who trades them. For each pair we would track the multifractal spectrum as a time series, the way clinicians track a biomarker across a hospital stay rather than reading it once at admission. We would then ask three questions. Is the A-leg's spectrum persistently wider than the H-leg's? Does that gap survive removing detected volatility-level breaks? Does the gap narrow when the A-share investor base becomes more institutional?

Core Claim

If retail participation genuinely generates multifractal volatility, three things should hold. First, within dual-listed pairs, the part of the spectrum width that comes from temporal correlation should be larger on the A-leg. Second, this gap should remain after the series are adjusted for discrete variance breaks. Third, the gap should shrink in A-shares that become eligible for cross-border institutional buying through the Stock Connect programs, compared with A-shares that do not. If instead the gap vanishes after break adjustment, the conclusion reverses: retail influence works through regime shifts, and multifractality in these markets is a symptom of those breaks rather than of persistent behavioral memory. Either outcome is informative. Each would correct a causal story that current work tells only qualitatively.

What is new is the combination of three things: a within-firm comparison that holds fundamentals fixed, a spectrum estimated as a time series rather than once per epoch, and an explicit test against the structural-break explanation.

Prior Work and Where This Sits

Three studies motivate the question. Application of Multi-Fractal Detrended Fluctuation Analysis (duplicate record here; Application of Multi-Fractal Detrended Fluctuation Analysis) applies multifractal detrended fluctuation analysis (MF-DFA) to 14 indices over a single 2006–2013 window. MF-DFA measures how fluctuations of different sizes grow with the length of the time window. The paper finds that small moves are persistent while large moves are anti-persistent, ranks markets by spectrum width, and explains Asian–Western differences through retail participation. It does so only verbally, with one static spectrum per index and no investor data. Analysis of Price Differences Between A and H Shares uses the same 63 dual-listed firms we would build on. It reaches an adjusted $R^2$ of 55.21% and finds a positive coefficient of 0.2484 on the A/H 90-day volatility ratio, which supports the idea that segmentation by investor type changes volatility. Its volatility measure is a plain rolling standard deviation with no notion of memory or scaling. Stylized facts of price gaps in limit order books and Scaling and memory in the non-Poisson process of limit order cancelation show long memory (DFA exponents 0.7–0.8 and 0.80–0.97) and multifractality in Shenzhen order-book data. Both treat the order flow as one undifferentiated population.

Two studies constrain the design. Bootstrap-Based Test for Volatility Shifts in GARCH against Long-Range Dependence shows that discrete variance shifts mimic long memory and that standard asymptotic change-point tests are badly miscalibrated in heavy-tailed finite samples. We therefore use its residual-bootstrap test for change points rather than the asymptotic version. Locally stationary long memory estimation estimates a time-varying memory parameter $d(u)$ with wavelets. Applied to yen/dollar volatility, it finds cyclical behavior and shifts around crises. It is our main methodological neighbor and our strongest descriptive baseline. It tracks one memory number, not the shape of the spectrum, and it uses no information about who is trading.

The clinical longitudinal literature supplies the method for the forecasting part. Association between albumin changes and prognosis in older patients with acute myocardial infarction shows that the in-hospital change in a marker predicts outcomes better than its admission value, and that the relationship is S-shaped. It models this with restricted cubic splines, flexible curves that can bend at estimated thresholds, and finds inflection points at −3.27 and 0.92 g/L. The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness shows that a static ratio loses its predictive value in the most dysregulated patients. We borrow this hierarchy of marker summaries: a single snapshot, a simple difference score, a spline of the change, and the full trajectory. We then ask which level of summary of the spectrum carries information about future tail events. This analogy shapes the methods only. It is not a claim that markets resemble patients.

Research Program

The basic object is the spectrum width $\Delta\alpha_{i,L}(t)$. It is computed by MF-DFA on intraday 5-minute absolute returns of firm $i$'s listing $L \in \{A, H\}$, over a trailing window ending on day $t$. A-shares give about 48 five-minute returns per day, so a 60-trading-day window holds roughly 2,900 points. That is enough for moments $q \in [-4, 4]$. For negative $q$, which magnify the smallest fluctuations and become unstable when those fluctuations approach zero, we cross-check with a wavelet-leader estimator.

Each window is decomposed with two kinds of surrogate series. A shuffled surrogate keeps the distribution of returns but destroys their time order. A phase-randomized surrogate keeps the correlations but makes the distribution Gaussian. Comparing the original with these surrogates splits the width into a fat-tail part and a correlation part. Only the correlation part is the behavioral-memory quantity we care about.

The central estimand is the within-pair gap

$$G_i(t) = \Delta\alpha^{\text{corr}}_{i,A}(t) - \Delta\alpha^{\text{corr}}_{i,H}(t),$$

where the superscript denotes the correlation component. We regress $G_i(t)$ on firm fixed effects and on time-varying controls: the A/H realized-volatility ratio, the A/H turnover ratio, and the A/H price premium. A positive intercept that survives these controls is the "retail widens the spectrum" effect, net of plain volatility level.

The break-adjusted version repeats everything after the bootstrap CUSUM procedure locates variance breaks within each window and rescales each segment to unit variance. The fraction of $G$ that remains after adjustment answers Gap 3 directly.

Before touching real data, we calibrate the whole pipeline on synthetic series with matched sample sizes. These include GARCH with planted breaks and no true memory, FIGARCH with true memory, and a multifractal random walk. Calibration has to show that the break adjustment removes spurious width in the first case without erasing genuine width in the others. Otherwise the main test cannot be interpreted.

Baselines

The naive reference point is the volatility gap itself. It is the A/H realized-volatility ratio, the exact regressor that Analysis of Price Differences Between A and H Shares reports with a coefficient of 0.2484 inside an adjusted $R^2$ of 55.21% for the price premium. If the spectrum gap $G$ carries no information beyond this ratio, then multifractality adds nothing to the demographic story. Beating this baseline means one of two things. Either $G$ stays significant with the ratio as a control, or $G$ improves out-of-sample prediction of the A/H premium beyond the ratio alone, on the same firms, the same 2014–2017 period, and the same stepwise specification extended by one variable.

The strongest descriptive competitor is the time-varying memory parameter $d_{i,L}(u)$ from Locally stationary long memory estimation. It is run on the same intraday series with its recursive one-sided kernel and its bandwidth tuned by its own criterion. If the A-minus-H difference in $d(u)$ explains as much as $G$, then a single memory number suffices and the spectrum adds nothing. That paper reports cyclicality and crisis shifts qualitatively but no accuracy figure; no reported result found on Lacuna for its predictive or discriminative performance. We will therefore compare the two directly on the same tasks.

For forecasting, the strongest classical competitors are GJR-GARCH and Realized GARCH fitted to each listing. Regime-aware Financial Volatility Forecasting via In-Context Learning treats GJR-GARCH as its strongest classical baseline and reports about 27% lower high-volatility MAE than GJR-GARCH on the S&P 500. Deep Learning Enhanced Realized GARCH reports beating Realized GARCH on 28 of 31 Oxford-Man indices. No reported result found on Lacuna for any of these models on dual-listed A/H stocks. We will refit them ourselves with the same rolling estimation windows and evaluation dates.

The key ablations remove one ingredient at a time. The first drops the surrogate decomposition and uses the raw width. The second drops the break adjustment. The third replaces the trajectory of $\Delta\alpha$ with its snapshot value.

Why not just…?

Why not just compare volatility levels or kurtosis between A and H legs? Spectrum width is known to rise with fat tails. A wider A-leg spectrum could therefore just restate that A-shares have fatter tails and higher volatility. This is a real threat, and it is why the estimand is the correlation component of the width, with realized volatility and turnover included as controls. If $G$ loses significance once the A/H kurtosis ratio is also added, we will report that the demographic effect runs through the distribution of returns, not through their memory.

Why not just use a single time-varying Hurst exponent or $d(u)$? The 14-index study shows that small and large moves can scale in opposite directions. A single exponent averages over exactly that inversion, so retail and institutional legs could share the same $d$ while differing at extreme $q$. This is an empirical question. The head-to-head comparison against $d(u)$ in the baselines settles it, and if $d(u)$ does as well, we will adopt it as the simpler marker.

Why not just fit GARCH-family models with a regime-switching or break term and read retail effects from the persistence parameter? GARCH persistence close to one, as reported around the 2015 crash in Information Spillover Effect Changes of Major Financial Markets, cannot by itself tell genuine long memory apart from breaks. GARCH also assumes memory decays exponentially, which the order-book DFA exponents above contradict. For pure forecasting, however, GARCH or HAR-type models may well beat any spectrum-based predictor. We therefore treat forecasting as secondary and report it honestly rather than leaning on it.

Proposed Experiments

The first experiment runs the synthetic calibration described above. It must be passed before anything else is interpreted. It reports the false-positive rate of nonzero correlation-width on break-only GARCH series, before and after adjustment, and the power to detect width in FIGARCH and multifractal random-walk series.

The second experiment estimates $G_i(t)$ on the dual-listed panel. It starts from the AH paper's 63 firms and its Bloomberg sample and extends both the firm set and the period where intraday data allow. It reports the fixed-effects intercept with and without controls, and with and without break adjustment. The decisive check comes here. A positive, break-robust intercept supports genuine retail-driven memory. An intercept that collapses after adjustment supports the break account.

The third experiment exploits the 2014 and 2016 Stock Connect eligibility expansions as shifts in investor mix. It compares newly eligible A-shares with ineligible ones in a difference-in-differences design. The demographic hypothesis predicts that $\Delta\alpha^{\text{corr}}_A$ narrows for eligible stocks while their H-legs do not change. As a placebo, we check that eligibility does not move the matched H-shares.

The fourth experiment treats $\Delta\alpha(t)$ as a longitudinal marker. It asks whether it predicts next-month tail events, defined as realized volatility in the top decile of the stock's own history or a drawdown beyond its 95th percentile. The comparison runs across the clinical hierarchy of summaries: snapshot, difference score, restricted-cubic-spline change, and full trajectory. All of these are added to Realized GARCH forecasts and evaluated by AUC and log loss on a time-ordered 70/30 split. Index-level replication on the Oxford-Man Realized Library tests whether the 2015 crash appears as an inflection in the trajectory before it appears in the level.