Estimating interest rate setting behaviour in Korea: a constrained ordered choices model approach

ABSTRACT

We study the Bank of Korea’s interest rate setting behaviour using an array of constrained ordered choices models, where the Monetary Policy Committee revises the target policy interest rate only when the current market interest rate deviates from the optimal rate by more than certain threshold values. Our models explain changes in the monetary policy stance well for the monthly frequency Korean data since January 2000. We find important roles for the output gap and the foreign exchange rate in understanding the Bank of Korea’s rate decision-making process. We also implement out-of-sample forecast exercises with September 2008 (Lehman Brothers Bankruptcy) for a split point. We demonstrate that out-of-sample predictability improves greatly for the rate cut and the rate hike decisions using SE-adjusted inaction bands.

KEYWORDS:

• Monetary policy
• Bank of Korea
• probit model
• robit model
• logit model
• target RP rate
• interbank call rate
• Taylor rule

JEL CLASSIFICATION:

• E52
• E58
• C51
• C52

Acknowledgement

The authors thank Kyung Soo Kim, Michael Stern and seminar participants at the 2014 KEA International Conference and Auburn University for helpful comments. Special thanks to the referee who gave constructive suggestions. The views expressed herein are those of the authors and do not necessarily reflect the official views of the Bank of Korea. When reporting or citing this paper, the authors’ names should be explicitly stated.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Since the BOK officially employed the inflation targeting system in 1998, they have been implementing monetary policies by setting policy interest rates such as the target RP rate.

² Nominal interest rates are bounded below by 0%.

³ Hu and Phillips (2004b) also investigated the Bank of Canada’s monetary policy behaviour using a similar methodology. Phillips, Jin, and Hu (2005) corrected the errors in Hu and Phillips (2004b) with regard to the convergence rates of maximum likelihood estimates.

⁴ A referee pointed out that forecasting cut or hike decisions might be more important than predicting stay decisions correctly. For this purpose, the referee suggested to use the robit model that may help improve the fit of tails.

⁵ We allow the inaction band $\left[{\tau }_L,{\tau }_U\right]$ to be asymmetric because we do not impose any restriction on the thresholds. We may assume ${\tau }_L=-{\tau }_U$ for symmetric bands when ${\tau }_L$ is restricted to be less than zero.

⁶ Note that its coefficient is restricted to be $-1$ since we are interested in the divergence measure of newly set optimal interest rate from the current market interest rate.

⁷ The robit model approximates the logit model when the degree of freedom is seven. See Liu (2005) for details.

⁸ The target RP rate and the call rate correspond to the target federal funds rate and the effective federal funds rate in the US, respectively, prior to the recent US financial crisis.

⁹ For the quadratically detrended gap, we demeaned and detrended the real industrial production using an intercept, linear trend and quadratic trend. See Clarida, Galí, and Gertler (2000), among others, who employed the same method. We separated HP cyclical components of the monthly real industrial production using 125 000 for the smoothing parameter.

¹⁰ Adding additional thresholds, we may extend the model to incorporate these 50 and 100 basis points changes. Since these are quite rare events (5 out of 162 observations), a trichotomous specification seems to be a more efficient choice.

¹¹ However, this caveat does not apply to out-of-sample forecast when one uses point estimates to formulate the conditional expectation (see Kim, Jackson, and Saba 2009).

¹² The BOK switched from the total CPI inflation to the core CPI inflation for the period between 2000 and 2006. They returned to the total CPI inflation in 2007. Replacing ${\pi }_t$ with the core inflation yields similar results because these two inflation measures exhibit quite similar dynamics over time.

¹³ Results with other specifications are available upon requests.

¹⁴ Results from other models are qualitatively similar.

¹⁵ Results from other specifications are available upon requests.

¹⁶ For example, correct prediction rates for C, S and H actions from Taylor B model were 18.75%, 96.28% and 6.67%, respectively, when we employed a point estimate-based inaction band. Model Taylor H1 performed similarly, yielding 31.25%, 99.24% and 6.67%, respectively.