South Korea Trend Inflation

Trend inflation is the long-run persistent component of inflation, namely the underlying rate of price increase that remains after removing transitory supply shocks, seasonal fluctuations, base effects, and measurement noise. It represents the inflation regime toward which observed inflation gravitates over the medium to long term, and can be interpreted as the economy's 'steady-state' inflation rate.

Observed inflation contains two distinct components. One is a slowly evolving trend driven by monetary policy credibility, the anchoring of expectations, and structural economic forces, and the other is a volatile transitory component driven by supply shocks (oil prices, food prices, exchange rate pass-through), tax changes, and measurement idiosyncrasies. Since the trend is a latent component that cannot be directly observed in market prices, it is extracted using an unobserved-components model. Each month's published value is the one-sided estimate of that trend from inflation data up to that month, and the series is the trend itself, not the long-run anchor of the expected-inflation curve, which blends the trend with the inflation target and survey expectations.

A rise in trend inflation indicates strengthening underlying price pressure, while a fall indicates its easing. Under a credible inflation-targeting regime, trend inflation gravitates toward the announced target over time, and persistent deviations from the target may indicate erosion of central bank credibility, structural shifts in price-setting behavior, or supply-side forces that monetary policy cannot easily offset (Carvalho et al. 2023).

Extracted from headline CPI year-over-year inflation using the Unobserved Components Stochastic Volatility (UCSV) model of Stock and Watson (2007). The published series starts in December 1995, the first month whose expanding window holds 60 monthly observations.

(1) State-Space Setup. The observation and trend equations are expressed in the following state space:

The log-volatility processes follow:

where $\tau_t$ is trend inflation, $h_t$ is observation log-volatility, and $g_t$ is trend innovation log-volatility.

(2) Gibbs Sampling. Estimation uses the precision-based Gibbs sampler of Chan and Jeliazkov (2009), which exploits the tridiagonal structure of the precision matrix for $O(T)$ sampling. The nonlinear SV observation equation is linearized via the 10-component Gaussian mixture approximation to $\log(\chi^2(1))$ from Omori et al. (2005). The sampler cycles through five Gibbs blocks, drawing $\tau_{1:T}$ via precision-based sampling, $h_{1:T}$ and $g_{1:T}$ via the KSC mixture, the mixture indicators via a multinomial draw, and $\sigma^2_h, \sigma^2_g$ from an inverse-gamma conjugate prior.

(3) MCMC Settings. The sampler uses 2,000 burn-in draws, 5,000 posterior samples, an IG prior with $\nu_0 = 10$ and $E[\sigma^2] = 0.02$, and a diffuse initial state variance $V_0 = 10.0$. The sampler is initialized from a 12-month moving average of inflation, which damps monthly seasonal variation in its starting values.

(4) One-sided expanding-window estimate. The published trend is not a two-sided posterior path over the full sample. For each month the sampler is run afresh on the expanding window of inflation observations up to and including that month, and the published value is the posterior mean of the trend in the window's final month, so no later observation enters it. The series begins only once the window reaches its minimum length of 60 monthly observations, and the months before that are not estimated. Every window uses the same random seed, so consecutive months share common random numbers and their month-to-month moves are not independent Monte Carlo noise. The initial trend of every window has the same diffuse prior, centered on average inflation over the first year of the sample, so the earliest months lean more on that start value and on the fixed priors than later months do. Each value is one-sided in the data, but the prior mean of the volatility innovation variances is calibrated on the full sample and the input is the latest vintage of the price index, so the series is a quasi-real-time estimate, not a real-time one (Orphanides and van Norden 2002). It is a model estimate of an unobserved trend, neither an observed statistic nor a central bank's estimate, and a past month's value changes only when the inflation data up to that month are revised or rebased or when the sampler itself changes.

Trend inflation provides a benchmark for measuring the inflation gap in policy rate rules that is not distorted by transitory shocks. In the Taylor (1993) rule and its modern extensions, the optimal policy rate responds to the inflation gap, namely the deviation of current inflation from the target. Using trend inflation rather than headline inflation to compute this gap yields a structurally cleaner measure, enabling policymakers to assess whether the current policy rate is appropriate for the medium-term inflation trajectory.

Changes in trend inflation are associated with shifts in the monetary policy regime. The concept originates from the permanent-transitory decomposition of macroeconomic time series, applied to inflation by Beveridge and Nelson (1981) and refined by Stock and Watson (2007). Cogley and Sargent (2005) and Cogley, Primiceri, and Sargent (2010) demonstrate that movements in trend inflation track changes in the policy regime. During periods of high and rising trend inflation (e.g., the 1970s in the U.S.), monetary policy responded insufficiently to inflation, violating the Taylor principle ($\partial i / \partial \pi > 1$), and inflation rose in a self-reinforcing manner. The subsequent decline in trend inflation during the Volcker disinflation and the Great Moderation reflected a credible commitment to price stability.

For Korea, trend inflation dynamics have been shaped by several structural factors. Rapid productivity growth during the industrialization period had a disinflationary effect, the Asian Financial Crisis of 1997–98 produced a sharp but temporary inflation spike, and after the adoption of formal inflation targeting in 1998 Korean inflation gradually converged toward advanced-economy norms. Since 2020, the global inflation surge driven by supply chain disruptions and energy price shocks has tested the anchoring of Korean trend inflation.

Trend inflation also provides the appropriate benchmark for computing 'core' inflation gaps and for calibrating the inflation component in financial conditions indices (FCIs). Chan, Koop, and Potter (2013) show that the UCSV trend outperforms traditional core measures (trimmed mean, weighted median, ex-food-and-energy) in forecasting future headline inflation, because it adaptively filters out different types of transitory shocks rather than mechanically excluding fixed categories.

The UCSV model's allowance of time-varying volatility in both the trend and the transitory component is empirically important for trend extraction. Inflation volatility varies substantially across regimes, from the Great Inflation of the 1970s to the Great Moderation and through the post-COVID inflation surge, and a model that imposes constant volatility produces biased trend estimates during regime transitions. The model's stochastic volatility structure further provides an additional informational dimension, where the estimated variance of trend innovations ($\exp(g_t)$) measures the degree of uncertainty about the inflation regime itself. High trend innovation volatility signals that the inflation regime may be shifting, which has implications for the conduct and communication of monetary policy.

Changes in trend inflation mark shifts in the persistent component of realized inflation, with direct implications for nominal bond yields, real return outlooks, and inflation-linked asset valuations.

For bond investors, trend inflation is a key determinant of the long-run level of nominal yields. A secular decline in trend inflation, as observed in Korea and other advanced economies from the late 1990s through the 2010s, is associated with a structural downshift in equilibrium bond yields. Cieslak and Povala (2015) show that decomposing inflation into a trend and a cyclical component substantially improves bond return predictability, as the trend component captures the persistent shifts in yield levels that drive long-horizon returns.

The stochastic volatility structure of the UCSV model provides a natural measure of inflation uncertainty, which is itself a priced risk factor in bond markets. Wright (2011) demonstrated that inflation uncertainty is positively correlated with term premiums, and when the inflation regime is uncertain, investors demand higher compensation for holding nominal bonds. The UCSV's time-varying volatility estimates can therefore serve as inputs for term premium models and inflation risk premium estimation.

For asset allocators, trend inflation informs the real return outlook for different asset classes. When trend inflation is rising, nominal bonds face headwinds (capital losses from rising yields), while commodities, real estate, and inflation-protected securities tend to outperform. Conversely, declining trend inflation is favorable for long-duration nominal bonds and growth equities whose valuations benefit from lower discount rates.

The KRED trend inflation series summarizes the persistent component of realized inflation. It complements, rather than replaces, market-based breakeven and survey-based measures of inflation expectations for Korean institutional investors assessing inflation risk exposure.

KRTRINF is the trend of KRED's unobserved-components stochastic-volatility model of headline CPI inflation (Stock and Watson 2007), published as a one-sided estimate, the posterior mean of the trend in the final month of an expanding window refitted for each month, and never as a full-sample posterior path. The trend's integration order is imposed by the random-walk trend equation rather than estimated, and the endpoint posterior mean of such a trend is a one-sided, exponentially weighted average whose time-varying weight, set by the volatility signal-to-noise ratio, dominates the path's serial correlation (Cogley 2002), so the statistical object is a persistence summary of that one-sided path and not an integration order of the data.

The augmented Dickey and Fuller (1979) regression with the lag augmentation of Said and Dickey (1984), the nonparametric Phillips and Perron (1988) test and the KPSS stationarity test of Kwiatkowski et al. (1992) were deliberately not run on this path, because a verdict from them would test the random-walk trend the model imposes rather than a property of the data, and the random-walk-versus-constant character of such a trend is not pinned down by the likelihood (Stock and Watson 1998). The GLS-detrended test of Elliott, Rothenberg, and Stock (1996) was not run for the same reason, and neither was the modified-statistic battery of Ng and Perron (2001), since both confirm an integration reading this archetype does not take. The level break search of Bai and Perron (1998) was likewise not run, because a break in the mean of an estimated trend reflects the estimator's response to its inputs rather than a change in the economy (Perron 1989). A level portmanteau such as Ljung and Box (1978) is not applied either, since on this path it reads the estimator's weighting rather than evidence about the data-generating process.

What the matrix mandates in their place is a descriptive persistence summary and a statement about revision. Over the 366 months from 1995-12-01 to 2026-05-01 the first-order autoregressive coefficient of the one-sided trend is 0.964 and the implied half-life is 18.96 months, reported as a property of this series and not as an order of integration, and read with the small-sample downward bias of near-unit autoregressive estimates in mind (Andrews 1993). The series begins at 1995-12-01 because no month is estimated before the expanding window reaches its minimum length. The revision cell records that a stored month is recomputed only when the inflation data up to that month are revised or rebased or when the sampler itself changes, and that each load replaces the whole history, so a fixed past month can move across updates although no smoothing enters it, and no dated vintage panel exists from which the size of that movement could be measured (Orphanides and van Norden 2002).

Three limits are stated rather than implied. No order of integration is assigned, by ruling and not by an inconclusive test (Hamilton 2018). The trend is one-sided in the data but it is not a real-time estimate, since the prior on the volatility innovations is calibrated on the full sample and the input is the latest vintage of the price index, which makes it quasi-real-time evidence in the sense of Orphanides and van Norden (2002), and it is a model estimate, neither an observed statistic nor a central bank's estimate. And every window uses the same random seed, so consecutive months share common random numbers, their month-to-month moves are not independent Monte Carlo noise, and the extent to which shared simulation error enters the persistence measured above is not quantified here, while the half-life is measured in months.