[Paper Review] Multifractal dynamics of stock markets

Dariusz Grech, Łukasz Czarnecki|arXiv (Cornell University)|Dec 17, 2009

Complex Systems and Time Series Analysis2 references6 citations

TL;DR

This study investigates multifractal dynamics in stock indices from developed (S&P500) and emerging (WIG) markets using multifractal detrended fluctuation analysis (MF-DFA). It demonstrates that non-stationarity—especially from extreme events like crashes and sub-trends—distorts multifractal scaling, but removing such events restores monotonic h(q) behavior and widens the f(α) spectrum, revealing clearer multifractal structure.

ABSTRACT

We present a comparative analysis of multifractal properties of financial time series built on stock indices from developing (WIG) and developed (S&P500) financial markets. It is shown how the multifractal image of the market is altered with the change of the length of time series and with the economic situation on the market. We emphasize that the proper adjustment of scaling range for multiscaling power laws is essential to obtain the multifractal image of time series. We analyze in this paper multifractal properties of real financial time series using Hölder $f(α)$ representation and multifractal-DFA method. It is also investigated how multifractal properties of stocks change with variety of "surgeries" done on the initial real financial time series. This way we reveal main phenomena on the market influencing its multifractal dynamics. In particular, we focus on examining how multifractal picture of real time series changes when one cuts off extreme events like crashes or rupture points, and how fluctuations around the main trend in time series influence the multifractal behavior of financial series in the long-time horizon for both developed and developing markets.

Motivation & Objective

• To compare multifractal properties of financial time series from developed (S&P500) and emerging (WIG) markets.
• To investigate how the length of time series and economic conditions affect the observed multifractal structure.
• To determine the role of extreme events and non-stationarities in distorting multifractal scaling behavior.
• To assess how 'surgery' on time series—such as removing crashes or sub-trends—improves multiscaling properties.

Proposed method

• Employed multifractal detrended fluctuation analysis (MF-DFA) to compute q-dependent Hurst exponents h(q) from financial time series.
• Used the Legendre transform to derive the singularity spectrum f(α) from h(q), with α = dh(q)/dq and s(q) = qh(q) - 1.
• Applied a q-th moment fluctuation function F_q(s) to amplify small and large fluctuations, enabling detection of multifractal scaling.
• Analyzed time series in segmented regimes (high vs. low volatility) to isolate the impact of non-stationarity on multifractal structure.
• Performed 'surgery' on data by removing extreme returns (e.g., crashes) and re-integrating to generate modified time series.
• Compared multifractal spectra f(α) and h(q) behavior before and after surgical interventions to assess improvements in scaling.

Experimental results

Research questions

• RQ1 How do multifractal properties of S&P500 and WIG indices differ across developed and emerging markets?
• RQ2 To what extent does the length of a financial time series affect the reliability of its observed multifractal scaling?
• RQ3 How do extreme events such as market crashes distort the multifractal spectrum f(α) and the h(q) function?
• RQ4 What is the impact of non-stationary sub-trends (e.g., opposite directional movements) on the multifractal structure of financial indices?
• RQ5 Can removing extreme events restore monotonic h(q) behavior and a well-defined parabolic f(α) spectrum?

Key findings

• Finite-length financial time series exhibit artificial multifractality due to statistical limitations, especially in the estimation of h(q) for large q.
• The singularity spectrum f(α) for the full S&P500 index shows a non-monotonic, twisted shape, indicating non-stationary distortions from extreme events.
• When the S&P500 is split into low- and high-volatility regimes, both sub-periods display monotonic h(q) and reversed parabolic f(α) spectra, with wider spectra in the low-volatility regime.
• Time series with abrupt events or opposing sub-trends ('bad' data) show non-monotonic h(q) and distorted f(α), indicating degraded multiscaling.
• After removing extreme events via 'surgery', the h(q) function becomes monotonic and the f(α) spectrum widens significantly, restoring a clear multifractal structure.
• The width of the f(α) spectrum increases after removing crashes, indicating enhanced multifractality, confirming that non-stationarity obscures true multifractal dynamics.