Kim-Tsurumi (2000) Korean Currency Crisis and Regime Change: A Multivariate GARCH Model with Bayesian Approach

Summary

Kim and Tsurumi (2000) apply Bollerslev's (1990) Constant Conditional Correlation (CCC) multivariate GARCH (MGARCH) model to four daily Korean financial series — KOSPI spot and futures, Won/Dollar spot and non-deliverable forward (NDF) — spanning October 1996 to April 1998 and encompassing the 1997 currency crisis. A structural break is detected by maximizing a Bayesian posterior for the unknown changepoint date, derived via Laplace approximation from split-sample MGARCH log-likelihoods. The MGARCH posterior mode places the break at 20 October 1997, roughly three weeks before the official 8 November 1997 Won devaluation, with individual-series Bayesian posteriors suggesting the disintegration began as early as August 1997 for the futures and NDF markets.

Key Claims

• All four series are integrated of order one; returns $y_{it} = 100\ln(s_{it}/s_{it-1})$ are modelled with constant means.
• Marginal GARCH(1,1) estimates yield $\hat\alpha_i + \hat\beta_i \approx 1$ for all four series (integrated GARCH, IGARCH), consistent with infinite unconditional variance.
• Full-sample CCC-MGARCH correlation between KOSPI futures and spot is $\hat\rho_{KF,KS} = 0.8597$; full-sample log-likelihood is $-2100.77$.
• Individual-series Bayesian posterior break modes (grid search, 5 July–10 December 1997 in 5-day steps): KS (KOSPI spot) = 30 Sep 1997; KF (KOSPI futures) = 11 Aug 1997; WS (Won/Dollar spot) = 22 Oct 1997; NF (NDF) = 14 Aug 1997.
• MGARCH posterior break mode: 20 October 1997 (pre-break LL = −889.14; post-break LL = −1144.72; vs. pooled LL = −2100.77).
• The futures (KF) and NDF (NF) markets showed structural change earliest (~August 1997), preceding the spot markets — consistent with informed trading.
• The break predates the official IMF bail-out announcement and Won devaluation by several weeks.

Concepts Introduced or Extended

• Conditional Correlation Models — application of Bollerslev (1990) CCC-MGARCH to Korean financial markets.
• GARCH and BEKK-GARCH — IGARCH(1,1) marginal models for each series; Laplace-approximation Bayesian structural break detection in a multivariate GARCH framework.
• Structural Break Testing — Bayesian posterior for unknown break date derived from split-sample likelihoods via Laplace approximation; likelihood-ratio (LR) test as a classical complement.

Entities Mentioned

• Suduk Kim
• Hiroki Tsurumi
• Tim Bollerslev

Quotes

"The results indicate that there was a structural change in the Korean financial markets, especially the futures and NDF markets, as early as August 1997, well before the official devaluation date of November 8, 1997."

My Take

The Laplace-approximation approach to Bayesian break detection in MGARCH is pragmatic: it avoids full Markov chain Monte Carlo (MCMC) for the nuisance parameters while delivering a posterior over the break date. A limitation is that the constant correlation assumption (CCC) may itself be violated around a crisis — dynamic conditional correlation (DCC) or BEKK might better capture the correlation spike typical of financial stress. The finding that derivatives markets (futures, NDF) led spot markets in signalling the crisis aligns with the literature on price discovery, but the study cannot distinguish informed trading from liquidity-driven anticipation. The IGARCH finding may partly reflect the structural break itself (Lamoureux-Lastrapes 1990), which the break model partially addresses.