[Paper Review] Critical Crashes

Anders Johansen, Didier Sornette|arXiv (Cornell University)|Jan 6, 1999

Complex Systems and Time Series Analysis83 citations

TL;DR

The paper proposes that stock market crashes result from the buildup of long-range correlations among traders, leading to a critical collapse. Using a non-parametric log-periodogram method, it identifies a statistically significant peak at a scaling ratio near 2 across six major crashes, supporting the hypothesis of universal critical dynamics in financial markets.

ABSTRACT

We argue that the word ``critical'' in the title is not purely literary. Based on our and other previous work on nonlinear complex dynamical systems, we summarize present evidence, on the Oct. 1929, Oct. 1987, Oct. 1987 Hong-Kong, Aug. 1998 global market events and on the 1985 Forex event, for the hypothesis advanced four years ago that stock market crashes are caused by the slow buildup of long-range correlations between traders leading to a collapse of the stock market in one critical instant. We qualify the log-periodic oscillations using a novel non-parametric method that does not rely on any fit: the corresponding log-periodogram exhibits a strong statistically significant peak for all six crashes examined, pointing at approximately the same prefered scaling ratio around 2.

Motivation & Objective

• To investigate whether stock market crashes are driven by the slow buildup of long-range correlations among traders.
• To test the hypothesis that crashes result from a critical transition in nonlinear complex dynamical systems.
• To analyze historical crashes for evidence of log-periodic oscillations indicative of critical behavior.
• To apply a novel non-parametric method to detect scaling without relying on curve fitting.

Proposed method

• The study analyzes six major financial crashes: October 1929, October 1987, October 1987 Hong Kong, August 1998 global, and the 1985 Forex event.
• It applies a non-parametric log-periodogram method to detect long-range correlations without requiring parametric fitting.
• The method identifies the presence of log-periodic oscillations by detecting a statistically significant peak in the log-periodogram.
• The scaling ratio associated with the peak is estimated and compared across crashes to assess universality.
• The analysis focuses on the time series of market returns prior to each crash to detect critical precursors.

Experimental results

Research questions

• RQ1 Do long-range correlations among traders precede major stock market crashes?
• RQ2 Is there a universal scaling ratio associated with the critical phase before crashes?
• RQ3 Can log-periodic oscillations be detected without parametric fitting in historical crash data?
• RQ4 Do multiple crashes exhibit similar critical dynamics despite different market conditions?

Key findings

• A statistically significant peak in the log-periodogram was observed for all six crashes analyzed, indicating strong long-range correlations.
• The scaling ratio associated with the peak is approximately 2 for all crashes, suggesting a universal critical behavior.
• The non-parametric method successfully detected critical precursors without relying on curve fitting.
• The consistency of the scaling ratio across diverse crashes supports the hypothesis of a common critical mechanism in market crashes.