---
ticker: "KRMEDCPI"
title: "South Korea Weighted-Median CPI"
unit: "%"
frequency: "Monthly"
source: "Bryan and Cecchetti (1994); Ball and Mazumder (2019)"
release: "Updated Monthly"
category: "Consumer & Producer Prices"
country: "KR"
language: "en"
canonical: "https://kred.dev/en/series/KRMEDCPI"
license: "https://creativecommons.org/licenses/by-nc-nd/4.0/"
latest_value: 1.99
latest_date: "2026-09-01"
first_date: "2017-01-01"
observations_total: 117
observations_shown: 117
---

# South Korea Weighted-Median CPI

## Overview

The weighted midpoint of item-level price changes, filtering transitory outliers to track the persistent core of inflation.

## Key Figures

|  | Value | Date |
|---|---|---|
| Latest | 1.99 | 2026-09-01 |
| Change from previous | 0.00 | 2026-08-01 |
| Change over one year | +0.06 | 2025-09-01 |
| Highest on record | 4.32 | 2022-09-01 |
| Lowest on record | 0.21 | 2019-06-01 |
| Period covered | 2017-01-01 – 2026-09-01 |  |
| Observations | 117 |  |

## Recent observations

| Date | Value | Change |
|---|---|---|
| 2017-01-01 | 1.72 |  |
| 2017-02-01 | 1.52 | -0.20 |
| 2017-03-01 | 1.64 | +0.12 |
| 2017-04-01 | 1.39 | -0.25 |
| 2017-05-01 | 1.35 | -0.04 |
| 2017-06-01 | 1.54 | +0.19 |
| 2017-07-01 | 1.17 | -0.36 |
| 2017-08-01 | 1.52 | +0.35 |
| 2017-09-01 | 1.33 | -0.20 |
| 2017-10-01 | 0.95 | -0.38 |
| 2017-11-01 | 0.90 | -0.05 |
| 2017-12-01 | 1.06 | +0.16 |
| 2018-01-01 | 0.97 | -0.09 |
| 2018-02-01 | 1.00 | +0.03 |
| 2018-03-01 | 1.06 | +0.06 |
| 2018-04-01 | 1.39 | +0.32 |
| 2018-05-01 | 1.52 | +0.13 |
| 2018-06-01 | 1.41 | -0.11 |
| 2018-07-01 | 1.10 | -0.31 |
| 2018-08-01 | 1.19 | +0.09 |
| 2018-09-01 | 1.12 | -0.08 |
| 2018-10-01 | 1.07 | -0.05 |
| 2018-11-01 | 0.98 | -0.09 |
| 2018-12-01 | 0.90 | -0.08 |
| 2019-01-01 | 0.74 | -0.16 |
| 2019-02-01 | 0.63 | -0.10 |
| 2019-03-01 | 0.52 | -0.11 |
| 2019-04-01 | 0.44 | -0.08 |
| 2019-05-01 | 0.33 | -0.11 |
| 2019-06-01 | 0.21 | -0.12 |
| 2019-07-01 | 0.55 | +0.34 |
| 2019-08-01 | 0.85 | +0.30 |
| 2019-09-01 | 0.40 | -0.45 |
| 2019-10-01 | 0.42 | +0.03 |
| 2019-11-01 | 0.31 | -0.11 |
| 2019-12-01 | 1.05 | +0.74 |
| 2020-01-01 | 1.22 | +0.17 |
| 2020-02-01 | 0.94 | -0.28 |
| 2020-03-01 | 0.75 | -0.19 |
| 2020-04-01 | 0.42 | -0.33 |
| 2020-05-01 | 0.49 | +0.06 |
| 2020-06-01 | 0.53 | +0.04 |
| 2020-07-01 | 0.68 | +0.15 |
| 2020-08-01 | 0.70 | +0.02 |
| 2020-09-01 | 0.61 | -0.09 |
| 2020-10-01 | 0.64 | +0.03 |
| 2020-11-01 | 0.76 | +0.12 |
| 2020-12-01 | 0.87 | +0.11 |
| 2021-01-01 | 0.67 | -0.20 |
| 2021-02-01 | 0.69 | +0.02 |
| 2021-03-01 | 1.14 | +0.45 |
| 2021-04-01 | 1.45 | +0.31 |
| 2021-05-01 | 1.36 | -0.09 |
| 2021-06-01 | 1.46 | +0.10 |
| 2021-07-01 | 1.37 | -0.09 |
| 2021-08-01 | 1.38 | +0.01 |
| 2021-09-01 | 1.34 | -0.04 |
| 2021-10-01 | 1.83 | +0.49 |
| 2021-11-01 | 1.85 | +0.02 |
| 2021-12-01 | 1.98 | +0.13 |
| 2022-01-01 | 2.46 | +0.48 |
| 2022-02-01 | 2.85 | +0.39 |
| 2022-03-01 | 2.79 | -0.06 |
| 2022-04-01 | 2.76 | -0.03 |
| 2022-05-01 | 3.42 | +0.66 |
| 2022-06-01 | 3.69 | +0.26 |
| 2022-07-01 | 3.93 | +0.24 |
| 2022-08-01 | 4.32 | +0.39 |
| 2022-09-01 | 4.32 | +0.01 |
| 2022-10-01 | 4.00 | -0.32 |
| 2022-11-01 | 4.18 | +0.18 |
| 2022-12-01 | 3.74 | -0.44 |
| 2023-01-01 | 3.61 | -0.13 |
| 2023-02-01 | 3.63 | +0.02 |
| 2023-03-01 | 3.41 | -0.22 |
| 2023-04-01 | 3.44 | +0.03 |
| 2023-05-01 | 3.13 | -0.31 |
| 2023-06-01 | 2.58 | -0.55 |
| 2023-07-01 | 2.54 | -0.03 |
| 2023-08-01 | 2.57 | +0.03 |
| 2023-09-01 | 2.54 | -0.03 |
| 2023-10-01 | 2.66 | +0.12 |
| 2023-11-01 | 2.37 | -0.29 |
| 2023-12-01 | 2.21 | -0.16 |
| 2024-01-01 | 2.00 | -0.21 |
| 2024-02-01 | 2.03 | +0.03 |
| 2024-03-01 | 1.99 | -0.04 |
| 2024-04-01 | 1.94 | -0.05 |
| 2024-05-01 | 1.99 | +0.05 |
| 2024-06-01 | 1.78 | -0.21 |
| 2024-07-01 | 1.75 | -0.03 |
| 2024-08-01 | 1.56 | -0.20 |
| 2024-09-01 | 1.31 | -0.24 |
| 2024-10-01 | 1.16 | -0.16 |
| 2024-11-01 | 1.07 | -0.09 |
| 2024-12-01 | 1.48 | +0.41 |
| 2025-01-01 | 1.15 | -0.33 |
| 2025-02-01 | 1.26 | +0.11 |
| 2025-03-01 | 1.22 | -0.04 |
| 2025-04-01 | 1.54 | +0.32 |
| 2025-05-01 | 1.53 | -0.01 |
| 2025-06-01 | 1.65 | +0.12 |
| 2025-07-01 | 1.56 | -0.10 |
| 2025-08-01 | 1.46 | -0.10 |
| 2025-09-01 | 1.93 | +0.46 |
| 2025-10-01 | 1.90 | -0.03 |
| 2025-11-01 | 1.61 | -0.29 |
| 2025-12-01 | 1.57 | -0.04 |
| 2026-01-01 | 1.71 | +0.14 |
| 2026-02-01 | 1.71 | -0.01 |
| 2026-03-01 | 1.99 | +0.28 |
| 2026-04-01 | 1.92 | -0.06 |
| 2026-05-01 | 1.95 | +0.03 |
| 2026-06-01 | 1.98 | +0.03 |
| 2026-07-01 | 1.99 | +0.01 |
| 2026-08-01 | 1.99 | 0.00 |
| 2026-09-01 | 1.99 | 0.00 |

## Definition

KRMEDCPI is the weighted median CPI for Korea, a limited-influence estimator of underlying inflation computed each month from the cross-sectional distribution of item-level consumer-price changes rather than from any model of inflation dynamics (Bryan and Cecchetti 1994; Bryan, Cecchetti, and Wiggins 1997). The statistic is the 50th weighted percentile of that distribution, namely the price change of the item containing the half point of the cumulative weight distribution (Bryan and Pike 1991; Dolmas 2005; Meyer and Venkatu 2014). It descends from the reading of core inflation as the persistent component of measured inflation once transitory relative-price noise is removed, and from the robust-statistics argument that a central order statistic estimates that component more efficiently than the mean when the cross-section is fat-tailed (Eckstein 1981; Cecchetti 1997; Roger 1997; Aucremanne 2000).

Item price changes are measured year over year on non-seasonally-adjusted item indices, the published variant of the Bank of Japan's core indicators rather than the original month-over-month convention on seasonally adjusted components (Shiratsuka 1997; Hogen, Kawamoto, and Nakahama 2015; Shiratsuka 2015). The year-over-year basis removes item-level seasonality without a model-based adjustment step, but it carries the known caveat that year-over-year changes are affected by the previous year's developments and lag turning points in underlying inflation, a limitation noted by Shiratsuka (1997) drawing on Kimura (1995).

Weights follow the official vintage calendar rather than a single fixed vintage. Each month is aggregated under the official item-weight vintage that governed the headline index in that month, five vintages whose applicability segments begin at 2012-01, 2015-01, 2017-01, 2020-01, and 2022-01, mapped onto the current back-linked item panel through a name and code concordance and renormalized over the items with a valid price change. This restores the per-base-year practice of the Bank of Japan's published indicators and of the long-sample Japanese evidence (Hogen, Kawamoto, and Nakahama 2015; Shiratsuka 1997). The residual approximation is disclosed rather than hidden, as a small share of the older vintages' weight mass has no surviving counterpart on the current panel and is renormalized away, and the per-month coverage share and item count are published columns of the series itself. The published span begins at 2017-01, the human-signed start under a rule requiring every month from the start onward to pass the coverage and reconstruction thresholds of a zero-trim reconstruction arbiter that must reproduce official headline inflation within fixed tolerances. Four months from 2019-09 to 2019-12 exceed the per-month reconstruction tolerance because a free-education policy left a reconstruction hole in the discontinued education items, and they are retained under an outer-bounded, human-signed exception recorded in the engineering documentation.

Four properties established by the verification battery are disclosed prominently. The series runs 0.5793pp below headline inflation on average, with a Newey and West (1987) t statistic of −3.518, the textbook downward level bias of a weighted median on a right-skewed price-change distribution (Ball and Mazumder 2019), so it reads as a direction and momentum gauge rather than a level gauge. It tracks the 37-month centered trend of headline inflation worse than both official exclusion cores, with an RMSE of 0.7395 against 0.6475 and 0.5490, and the statistical insignificance of that deficit on the 95-month evaluation window is a small-sample power limit rather than evidence the deficit has closed (Clark 2001; Rich and Steindel 2007). The unit-root verdict on the published span is a clean I(1), with the three standard procedures in agreement, and is reported exactly as the battery returns it. The out-of-sample battery is not runnable on the 114-month span, so the series is a descriptive gauge of the current month's price-change distribution and never a forecast (Atkeson and Ohanian 2001; Meyer, Venkatu, and Zaman 2013).

A companion weighted trimmed-mean estimator was evaluated on the same battery and is withheld under the publish gate, because its twelve-month mean-reversion criterion failed on the published span while the weighted median passed it under the codified one-sided convention (Cogley 2002; Bryan and Cecchetti 1994).

## Methodology

The estimator is computed from the same sorted item-level cross-section each month, and every constant it depends on is fixed in code, with nothing re-estimated in production.

**(1) Item-level year-over-year price change.** Each item's price change is computed as $\pi_{i,t} = 100\,(P_{i,t}/P_{i,t-12} - 1)$ on non-seasonally-adjusted item indices, a ratio that is valid only when both observations lie on the same linked index series (Shiratsuka 1997; Shiratsuka 2015; Hogen, Kawamoto, and Nakahama 2015).

**(2) Per-month official weight vintage.** Each month receives the official item-weight vintage that governed the headline index in that month, five vintages whose applicability segments begin at 2012-01, 2015-01, 2017-01, 2020-01, and 2022-01 per the official rebase and update calendar, mapped onto the current back-linked item panel through a name and code concordance. The mapped weights are renormalized to unit mass over the items with a valid price change in the month, so the weights entering the estimator always cumulate to one, and the per-month coverage share of retained official weight together with the item count is published as columns of the output series (Bryan, Cecchetti, and Wiggins 1997; Dolmas 2005).

**(3) Stable sort and cumulative weights.** The valid items are sorted in ascending order of price change with a stable sort that keeps input order among ties, and the renormalized weights are accumulated into a cumulative weight for each sorted item (Bryan, Cecchetti, and Wiggins 1997).

**(4) Weighted median.** KRMEDCPI is the price change of the item containing the 0.5 point of the cumulative weight distribution, and when a cumulative weight lands exactly on the boundary the midpoint of the two adjacent items' price changes is taken, the unique rule consistent with the limit of the symmetric trim (Bryan and Pike 1991; Bryan and Cecchetti 1994; Roger 1997; Aucremanne 2000; Meyer and Venkatu 2014).

**(5) Reconstruction arbiter and derived start.** A zero-trim Lowe aggregate with price-backdated weights under each month's assigned vintage must reproduce official headline year-over-year inflation within fixed per-month and per-segment tolerances, which validates the item universe, the weight mapping, and the year-over-year computation jointly (Silver 2007; Wynne 2008). The published span begins at the earliest month from which every later month passes both the coverage threshold and the reconstruction tolerances once a human-signed, outer-bounded exception covering the four months from 2019-09 to 2019-12 is applied, namely 2017-01.

## Applications in Economics

Central banks watch limited-influence estimators because monthly headline inflation mixes the persistent component relevant for policy with transitory relative-price noise, a decomposition that goes back to the original definition of core inflation as the trend rate on the economy's long-run path (Eckstein 1981; Cecchetti 1997). When the cross-sectional distribution of item price changes is fat-tailed, the weighted mean is an inefficient estimator of that trend and limited-influence estimators are more efficient ones, the founding argument of the trimmed-mean literature within which the weighted median is the limiting member (Bryan and Cecchetti 1994; Bryan, Cecchetti, and Wiggins 1997). The same robust-statistics reading was developed for New Zealand and Belgium, where medians and trimmed means were proposed as robust estimators of the central tendency of inflation (Roger 1997; Aucremanne 2000).

The relation to the official exclusion cores is a running theme of the evaluation literature. Fixed-exclusion cores remove the same items every month regardless of where the shocks fall, whereas a trimmed mean removes whichever items are extreme in the current month, and the permanent exclusion of food and energy has been criticized on precisely that ground (Bullard 2011). Comparative studies find that no single core measure dominates and that trimmed means and medians are competitive with, and often better than, exclusion cores at tracking trend inflation (Clark 2001; Rich and Steindel 2007; Silver 2007; Wynne 2008; Detmeister 2011). An alternative tradition identifies core inflation as the component of measured inflation with no long-run effect on output (Quah and Vahey 1995).

In central-bank practice, underlying-inflation dashboards of exactly this kind are maintained for the United States, the euro area, and Japan (Amstad, Potter, and Rich 2017; Ehrmann et al. 2018; Hogen, Kawamoto, and Nakahama 2015; Shiratsuka 2015). Within an inflation-targeting regime the operational case for such gauges is that policy should respond to the persistent part of inflation relative to the target, in Korea the 2 percent consumer-price target, while communication about headline versus core manages the risk that persistent shocks are wrongly discounted (Bernanke and Mishkin 1997; Svensson 1997; Mishkin 2007; Blinder 1997). The trend object these gauges approximate is the slowly moving component that unobserved-components models extract from the same data (Stock and Watson 2007; Stock and Watson 2016).

On this site the weighted median is read against the two official Korean exclusion cores and the model-based trend as the smoothness and momentum reading of the item-level cross-section, with its downward level bias kept in view rather than netted out.

## Applications in Financial Markets

Fixed-income and inflation-market participants price nominal bonds, inflation-linked bonds, and inflation swaps off a decomposition of yields into expected inflation, real rates, and risk premia, in which the expected-inflation input is a persistent object rather than the latest monthly print (Ang, Bekaert, and Wei 2008; D'Amico, Kim, and Wei 2018; Duffee 2018). Underlying-inflation gauges supply the persistent reading against which breakeven rates are judged, and mispricing episodes in indexed markets show that the market-implied inflation path can deviate substantially from fundamentals (Fleckenstein, Longstaff, and Lustig 2014; Bauer and McCarthy 2015).

For reading monthly releases, the informational question is whether a surprise moves the persistent component or only the noise, which is why long-maturity yields react to macroeconomic news through revisions in long-run inflation expectations (Gurkaynak, Sack, and Swanson 2005). Survey measures, market prices, and time-series models compete as expected-inflation inputs, with surveys hard to beat in United States data, and the underlying-inflation gauges published here are the cross-sectional entry in that toolkit (Ang, Bekaert, and Wei 2007; Faust and Wright 2013; Stock and Watson 2007).

The forecasting evidence for such gauges is mixed and its limits are stated plainly. A median CPI improves some United States inflation forecasts (Meyer, Venkatu, and Zaman 2013), and core measures predict total inflation better than headline in some samples (Crone et al. 2013), but simple benchmarks that use only lagged inflation are notoriously hard to beat (Atkeson and Ohanian 2001). For this Korean series the out-of-sample comparison is not runnable on the short published span, so the weighted median enters valuation work as a descriptive measure of the current persistent component, not as a forecasting model (Meyer and Venkatu 2014).

## Statistical Tests

The sample is 114 monthly observations spanning 2017-01 to 2026-06, the span certified by the reconstruction arbiter under the human-signed start rule, and the unit-root tests are run under a specification with constant and linear trend. The three procedures agree and a clean I(1) verdict is assigned. The augmented Dickey-Fuller test in lag-augmented form fails to reject a unit root at p = 0.8212 (Dickey and Fuller 1979; Said and Dickey 1984), the nonparametric test also fails to reject at p = 0.6163 (Phillips and Perron 1988), and the stationarity-null test rejects stationarity at p = 0.023 (Kwiatkowski et al. 1992). The same clean verdict holds for headline inflation and for both official exclusion cores on the identical window, and the label is taken directly from the battery output rather than asserted.

The portmanteau statistic on the first difference returns Q(12) = 10.86 with p = 0.5413, the weakest serial dependence in the suite and consistent with the series being the smoothest candidate (Ljung and Box 1978). The multiple-break procedure locates five mean breaks, at 2018-12, 2020-08, 2022-01, 2023-06, and 2024-11, with the 2022-01 date shared exactly by one official exclusion core and the full set segmenting the pandemic disinflation, the subsequent surge, and its unwinding (Bai and Perron 1998, 2003). The monthly-dummy F test on the first difference finds no residual seasonality, F = 0.4916 with p = 0.9046, consistent with the year-over-year basis adopted exactly to remove item-level seasonality (Shiratsuka 1997).

The estimator carries a significant downward level bias. The mean gap to headline inflation is −0.5793pp with a Newey and West (1987) t statistic of −3.518 and p = 0.0004, the textbook behavior of a weighted median on a right-skewed price-change distribution, documented for the United States analogue by Ball and Mazumder (2019), so the series reads as a direction and momentum gauge rather than a level gauge (Rich and Steindel 2007).

Against the 37-month centered moving average of headline inflation the RMSE is 0.7395 over the 95-month evaluation window, worse than both official exclusion cores at 0.6475 and 0.5490, and the deficit is insignificant at the 10 percent level under the Diebold and Mariano (1995) test with the small-sample correction of Harvey, Leybourne, and Newbold (1997), p = 0.3993 and p = 0.1094 (Clark 2001). That insignificance is read as a power limit of n = 95 rather than as evidence the deficit closed, since the same deficit tested significant on the longer superseded sample, and both facts are stated together. The offsetting strength is smoothness, as the full-span standard deviation of 0.9695 against 1.4610 for headline makes the series the smoothest in the suite (Bryan, Cecchetti, and Wiggins 1997).

The mean-reversion regression of future changes in headline inflation on the current gap follows Cogley (2002) and is evaluated under a codified one-sided convention, a directional test of a positive coefficient at the 5 percent level with the two-sided reading always reported alongside. At the twelve-month horizon the coefficient is 0.9399 with a Newey and West (1987) t statistic of 1.937 over 101 overlapping observations, significant under the one-sided convention while the two-sided reading is marginal and falls just short of the 5 percent threshold, and because the two readings differ on this sample both are stated. The series remains the only candidate in the suite, including both official exclusion cores, to pass the one-sided test at the twelve-month horizon. At the 36-month horizon the joint Wald test of the zero-intercept unit-slope restriction rejects for every series including both official cores, at 112.47 for this series and 74.98 and 35.69 for the cores, a surge-window artifact of 77 overlapping three-year observations that is uninformative for ranking candidates (Rich and Steindel 2007).

The out-of-sample family is recorded as not runnable, because the recursive design requires a 180-month initial estimation window and the published span is 114 months, and no substitute was improvised. Simple benchmarks that use only lagged inflation are notoriously hard to beat (Atkeson and Ohanian 2001; Crone et al. 2013), so no forecasting claim attaches to the series, which is published as a descriptive gauge of the current month's price-change distribution, never a forecast (Meyer, Venkatu, and Zaman 2013; Meyer and Venkatu 2014).

## Frequently Asked Questions

### What is the weighted-median consumer price index?

Item-level price changes for the month are ordered by size, weights are accumulated in that order, and the change at the point where cumulative weight reaches one half is taken. It is a limited-influence measure, constructed so that a few extreme items cannot pull the reading around.

### How does the weighted-median CPI differ from core CPI?

Fixed exclusion removes the same categories every month whether or not they are the source of that month's dispersion. The weighted median discards both ends of the distribution automatically, whichever items happen to lie there, so nothing has to be designated in advance.

### Can the weighted-median CPI be read as an inflation forecast?

No. It describes the underlying component present in current price data and carries no predictive claim.