---
ticker: "KRGDPNOW"
title: "South Korea Real-GDP Nowcast"
unit: "%"
frequency: "Monthly"
source: "Giannone, Reichlin, and Small (2008); Bańbura and Modugno (2014); Mariano and Murasawa (2003); Bok, Caratelli, Giannone, Sbordone, and Tambalotti (2018)"
release: "Updated Monthly"
category: "Recession & Nowcasting"
country: "KR"
language: "en"
canonical: "https://kred.dev/en/series/KRGDPNOW"
license: "https://creativecommons.org/licenses/by-nc-nd/4.0/"
latest_value: 0.83
latest_date: "2026-09-30"
first_date: "2013-02-28"
observations_total: 151
observations_shown: 120
---

# South Korea Real-GDP Nowcast

## Overview

Reads monthly activity data into a near real-time estimate of current-quarter growth, available ahead of the official release.

## Key Figures

|  | Value | Date |
|---|---|---|
| Latest | 0.83 | 2026-09-30 |
| Change from previous | 0.00 | 2026-08-31 |
| Change over one year | +0.25 | 2025-09-30 |
| Highest on record | 1.04 | 2022-04-30 |
| Lowest on record | -0.26 | 2020-07-31 |
| Period covered | 2013-02-28 – 2026-09-30 |  |
| Observations | 151 |  |

## Recent observations

| Date | Value | Change |
|---|---|---|
| 2015-11-30 | 0.80 | +0.01 |
| 2015-12-31 | 0.80 | 0.00 |
| 2016-02-29 | 0.76 | -0.04 |
| 2016-03-31 | 0.71 | -0.05 |
| 2016-04-30 | 0.59 | -0.12 |
| 2016-05-31 | 0.61 | +0.02 |
| 2016-06-30 | 0.62 | +0.01 |
| 2016-07-31 | 0.79 | +0.16 |
| 2016-08-31 | 0.78 | 0.00 |
| 2016-09-30 | 0.79 | +0.01 |
| 2016-10-31 | 0.78 | -0.02 |
| 2016-11-30 | 0.77 | 0.00 |
| 2016-12-31 | 0.74 | -0.04 |
| 2017-02-28 | 0.66 | -0.07 |
| 2017-03-31 | 0.69 | +0.03 |
| 2017-04-30 | 0.85 | +0.16 |
| 2017-05-31 | 0.89 | +0.04 |
| 2017-06-30 | 0.91 | +0.02 |
| 2017-07-31 | 0.93 | +0.02 |
| 2017-08-31 | 0.91 | -0.02 |
| 2017-09-30 | 0.90 | -0.01 |
| 2017-10-31 | 0.95 | +0.05 |
| 2017-11-30 | 0.95 | 0.00 |
| 2017-12-31 | 0.96 | +0.01 |
| 2018-02-28 | 0.85 | -0.11 |
| 2018-03-31 | 0.82 | -0.03 |
| 2018-04-30 | 0.86 | +0.04 |
| 2018-05-31 | 0.84 | -0.02 |
| 2018-06-30 | 0.84 | 0.00 |
| 2018-07-31 | 0.89 | +0.05 |
| 2018-08-31 | 0.85 | -0.04 |
| 2018-09-30 | 0.80 | -0.05 |
| 2018-10-31 | 0.69 | -0.10 |
| 2018-11-30 | 0.70 | 0.00 |
| 2018-12-31 | 0.68 | -0.02 |
| 2019-02-28 | 0.62 | -0.06 |
| 2019-03-31 | 0.61 | -0.01 |
| 2019-04-30 | 0.63 | +0.03 |
| 2019-05-31 | 0.64 | 0.00 |
| 2019-06-30 | 0.63 | -0.01 |
| 2019-07-31 | 0.68 | +0.05 |
| 2019-08-31 | 0.68 | 0.00 |
| 2019-09-30 | 0.66 | -0.02 |
| 2019-10-31 | 0.69 | +0.03 |
| 2019-11-30 | 0.72 | +0.03 |
| 2019-12-31 | 0.72 | 0.00 |
| 2020-02-29 | 0.78 | +0.06 |
| 2020-03-31 | 0.71 | -0.07 |
| 2020-04-30 | 0.29 | -0.41 |
| 2020-05-31 | -0.04 | -0.33 |
| 2020-06-30 | -0.02 | +0.01 |
| 2020-07-31 | -0.26 | -0.23 |
| 2020-08-31 | -0.21 | +0.04 |
| 2020-09-30 | -0.16 | +0.06 |
| 2020-10-31 | 0.31 | +0.47 |
| 2020-11-30 | 0.34 | +0.03 |
| 2020-12-31 | 0.44 | +0.10 |
| 2021-02-28 | 0.61 | +0.17 |
| 2021-03-31 | 0.68 | +0.07 |
| 2021-04-30 | 0.72 | +0.04 |
| 2021-05-31 | 0.84 | +0.12 |
| 2021-06-30 | 0.88 | +0.04 |
| 2021-07-31 | 1.03 | +0.15 |
| 2021-08-31 | 1.01 | -0.02 |
| 2021-09-30 | 0.98 | -0.03 |
| 2021-10-31 | 0.83 | -0.14 |
| 2021-11-30 | 0.86 | +0.02 |
| 2021-12-31 | 0.89 | +0.03 |
| 2022-02-28 | 1.00 | +0.11 |
| 2022-03-31 | 1.03 | +0.03 |
| 2022-04-30 | 1.04 | +0.01 |
| 2022-05-31 | 0.99 | -0.05 |
| 2022-06-30 | 1.00 | +0.01 |
| 2022-07-31 | 0.90 | -0.10 |
| 2022-08-31 | 0.81 | -0.08 |
| 2022-09-30 | 0.79 | -0.02 |
| 2022-10-31 | 0.80 | +0.01 |
| 2022-11-30 | 0.75 | -0.06 |
| 2022-12-31 | 0.71 | -0.03 |
| 2023-02-28 | 0.50 | -0.22 |
| 2023-03-31 | 0.49 | -0.01 |
| 2023-04-30 | 0.54 | +0.05 |
| 2023-05-31 | 0.58 | +0.04 |
| 2023-06-30 | 0.59 | +0.02 |
| 2023-07-31 | 0.73 | +0.14 |
| 2023-08-31 | 0.71 | -0.02 |
| 2023-09-30 | 0.71 | -0.01 |
| 2023-10-31 | 0.70 | -0.01 |
| 2023-11-30 | 0.69 | -0.01 |
| 2023-12-31 | 0.67 | -0.02 |
| 2024-02-29 | 0.64 | -0.03 |
| 2024-03-31 | 0.66 | +0.02 |
| 2024-04-30 | 0.70 | +0.04 |
| 2024-05-31 | 0.67 | -0.03 |
| 2024-06-30 | 0.66 | -0.01 |
| 2024-07-31 | 0.59 | -0.07 |
| 2024-08-31 | 0.57 | -0.02 |
| 2024-09-30 | 0.56 | -0.01 |
| 2024-10-31 | 0.51 | -0.05 |
| 2024-11-30 | 0.52 | +0.01 |
| 2024-12-31 | 0.53 | 0.00 |
| 2025-02-28 | 0.42 | -0.10 |
| 2025-03-31 | 0.43 | +0.01 |
| 2025-04-30 | 0.41 | -0.02 |
| 2025-05-31 | 0.42 | +0.01 |
| 2025-06-30 | 0.44 | +0.02 |
| 2025-07-31 | 0.58 | +0.14 |
| 2025-08-31 | 0.58 | 0.00 |
| 2025-09-30 | 0.58 | 0.00 |
| 2025-10-31 | 0.75 | +0.17 |
| 2025-11-30 | 0.70 | -0.05 |
| 2025-12-31 | 0.71 | +0.01 |
| 2026-02-28 | 0.59 | -0.12 |
| 2026-03-31 | 0.63 | +0.03 |
| 2026-04-30 | 0.76 | +0.14 |
| 2026-05-31 | 0.71 | -0.05 |
| 2026-06-30 | 0.72 | 0.00 |
| 2026-07-31 | 0.80 | +0.09 |
| 2026-08-31 | 0.83 | +0.02 |
| 2026-09-30 | 0.83 | 0.00 |

## Definition

The South Korea real-GDP nowcast (KRGDPNOW) is the pseudo-real-time smoothed estimate produced by a mixed-frequency dynamic factor model (MF-DFM) that estimates quarterly real-GDP QoQ growth from monthly real-activity indicators in near real time. The model treats a single latent business-conditions factor $f_t$ as jointly driving the monthly and quarterly indicators and reads quarterly GDP as a weighted moving average of an unobserved monthly GDP-growth flow (Mariano and Murasawa 2003). The whole specification is the Korean-data implementation of the nowcasting procedure Giannone, Reichlin, and Small (2008) established.

The panel consists of eight series and begins in July 2008, the first month at which all the Korean inputs become stably available. The monthly block holds seven series, namely (i) employment SA, (ii) industrial production SA, (iii) retail sales SA, (iv) services production SA, (v) the KOSTAT coincident composite cyclical component, (vi) the Bank of Korea BSI manufacturing business-conditions outlook, and (vii) the Bank of Korea CSI consumer-sentiment index, and the quarterly block holds real GDP QoQ SA.

The horizon is defined by the position of the evaluation date. A row is labeled backcast when the evaluation date is past the close of the target quarter but before the GDP release, nowcast when the evaluation date lies inside the target quarter, and forecast when the evaluation date precedes the start of the target quarter, a classification that follows the standard Giannone, Reichlin, and Small (2008) introduced.

The series is not a point forecast but the conditional expectation of GDP growth given the information set at each evaluation date. A rise in the nowcast means the newly arriving indicators point to an upward revision of the current-quarter growth trajectory and a fall points to a downward revision, and the value itself measures the central tendency of the distribution.

## Methodology

With the latent factor $f_t$, the monthly observation vector $y_t^m$, and the quarterly observation $y_t^q$, the model is

$$y_t^m = \Lambda^m f_t + \varepsilon_t^m$$

$$y_t^q = \Lambda^q \sum_{j=0}^{2} w_j f_{t-j} + \varepsilon_t^q$$

$$f_t = A_1 f_{t-1} + u_t,\quad u_t \sim \mathcal{N}(0, Q)$$

$$\varepsilon_{i,t} = \rho_i \varepsilon_{i,t-1} + \eta_{i,t},\quad \eta_{i,t} \sim \mathcal{N}(0, \sigma_i^2)$$

with the quarterly-flow temporal-aggregation weights $w_j$ taken from the Mariano and Murasawa (2003) quarterly-flow specification.

Estimation is by the EM algorithm of the mixed-frequency state-space model with the following specification. **(1) Input transforms.** The four trending Korean monthly indicators (employment SA, industrial production SA, retail sales SA, services production SA) enter as log first differences in percent. The KOSTAT coincident composite cyclical component is already detrended by construction and is fed in levels. The BSI manufacturing index and CSI consumer-sentiment index are survey indices and enter in levels. Quarterly real GDP already ships as a QoQ growth rate and is fed in directly as the quarterly input. **(2) Standardization and specification.** Standardization rescales each observed variable to in-sample mean zero and unit standard deviation to remove scale dependence, and the specification uses one factor, AR(1) factor dynamics, AR(1) idiosyncratic disturbances, with quarterly GDP fed in separately as the quarterly input. **(3) EM fit.** The Bańbura and Modugno (2014) EM algorithm iterates to convergence under a 500-iteration cap, initialized by a PCA-based deterministic start.

The publication-lag mask hides observations whose label date exceeds the visibility cutoff $d - L_i$ at evaluation date $d$. The lag values used here are 28 days for GDP, 13 days for the employment index, 30 days for industrial production, retail sales, services production, and the coincident-composite cyclical, and 0 days for the BSI and CSI surveys (both released on the same day they apply to). The pseudo-real-time backtest walks a monthly grid from 2013-01-01 forward and refits the model on the lag-masked panel at each evaluation date. Annual re-estimation refits the parameter vector at the first evaluation date of each calendar year and uses the same parameters to smooth subsequent intra-year panels, an operational rhythm consistent with the New York Fed Nowcast specification in Bok et al. (2018).

The news decomposition implements the Bańbura and Modugno (2014) Kalman-smoother weight decomposition directly, and the decomposition convention follows Bok et al. (2018, §III). The change in the nowcast

$$\Delta \hat{y}_{q,t} = \hat{y}_{q,t}^{(d_t)} - \hat{y}_{q,t}^{(d_{t-1})}$$

between evaluation dates $d_{t-1}$ and $d_t$ decomposes as the sum over source series $i$ of the data surprise $\nu_{i,t}$ (the observed value minus the prior-evaluation-date model expectation) times the Kalman gain $w_{i,t}$ from the $d_{t-1}$ model, so that

$$\Delta \hat{y}_{q,t} = \sum_i w_{i,t} \nu_{i,t}$$

holds, where the additivity of the decomposition is an exact identity.

Three naive benchmarks are reported alongside the model. **(1) Random Walk (RW)** is

$$\text{RW}(q) = y_{q-1}$$

, the lag-one autoregressive baseline that Atkeson and Ohanian (2001) presented as a stiff benchmark for U.S. inflation forecasting and that is borrowed here for the GDP-growth series. **(2) AR(2)** fits a two-lag autoregression on the quarterly GDP series available at the evaluation date and emits a one-step-ahead forecast; this is the simplest univariate time-series benchmark that does not use any of the side information. **(3) Last-observed** emits the most recently published quarterly GDP as the nowcast; by construction this baseline has zero error at the backcast horizon.

The sample begins in July 2008 and the backtest grid begins on 2013-01-01. The EM algorithm is deterministic but BLAS-sensitive, so the model is run with OPENBLAS_NUM_THREADS=1 and the cross-platform snapshot harness gates the nowcast time series against BLAS drift.

## Applications in Economics

Nowcasting is the procedure that extracts the conditional expectation of GDP by explicitly modeling the fact that the information set expands heterogeneously over time (Giannone, Reichlin, and Small 2008). KRGDPNOW is the implementation that applies that specification directly to Korean data. Bańbura, Giannone, and Reichlin (2011) consolidated this tradition as the standard reference, and Bańbura et al. (2013) treated the role of the data flow and vintages in the nowcasting environment in detail.

KRGDPNOW's specification sits on an established dynamic-factor-model lineage. The core decomposition adds the Mariano and Murasawa (2003) quarterly-flow temporal aggregation to the dynamic-factor coincident-index skeleton Stock and Watson (1989) established, reading quarterly GDP through an unobserved monthly factor, a construction that Mariano and Murasawa (2010) generalized to richer panels and for which the generalized dynamic factor model of Forni et al. (2000) supplies the external anchor. The Kalman-filtering identification and the steady-state representation follow the standard state-space form Hamilton (1994) consolidated, and the choice of the latent-factor dimension is consistent with traditional panel factor analysis under the Bai and Ng (2002) information criteria.

Estimation uses the Bańbura and Modugno (2014) arbitrary-missing-data EM algorithm, which implements, via EM, the maximum-likelihood estimation of dynamic factor models Doz, Giannone, and Reichlin (2011) developed and the large-panel quasi-maximum-likelihood approach Doz, Giannone, and Reichlin (2012) generalized. This treatment matters operationally because the seven monthly Korean series and the quarterly GDP series run at different publication lags. Because the EM treatment lets all observations enter the likelihood at their natural timing, there is no need to truncate the sample to the slowest series to force a balanced panel, and a single missing month in industrial production does not waste the information the other series carry for that month.

KRGDPNOW's operational specification is coherent with the Federal Reserve Bank of New York Nowcast. The New York Fed's formal specification, published in Bok et al. (2018), was maintained until its 2024 retirement, and KRGDPNOW is the Korean adaptation of it, matching in latent-factor count, AR(1) dynamics, AR(1) idiosyncratic disturbances, Mariano-Murasawa quarterly-flow temporal aggregation, EM fit discipline, and the Kalman-weight news-decomposition identity. Among other operational specifications, the Atlanta Fed GDPNow that Higgins (2014) develops is regression-based and the Cleveland Fed specification that Knotek and Zaman (2014) describe reflects judgment, whereas KRGDPNOW adopts a state-space specification because the Kalman smoother handles the heterogeneity of the evaluation-date information sets natively.

The pseudo-real-time backtest substitutes a publication-lag mask in the Korean environment, which lacks vintage data. Croushore (2011) consolidated the operational definition of vintage data, and Stark (2008) showed that vintage data provide a no-look-ahead environment for nowcast evaluation. Because Korea does not release official vintages in an operationally usable form, KRGDPNOW applies a lag mask at each evaluation date, an operational constraint of the Korean data environment. The lag mask ignores data revisions but models the timing of data arrival exactly, so it preserves part of the information that vintage data carry.

In the site's backtest, KRGDPNOW shows an RMSFE improvement of roughly 35% over the random walk and 10% over AR(2) at the nowcast horizon. This magnitude is the same order as the RMSFE improvement of around 30% that Giannone, Reichlin, and Small (2008) show on euro-area data, and it is comparable to the RMSFE improvement Bok et al. (2018) show on U.S. data. At the backcast horizon the last-observed baseline has zero error by construction, so the comparison is against the random walk and AR(2), and as the horizon moves toward the forecast region the information set narrows so the accuracy gain naturally shrinks.

KRGDPNOW measures the central tendency of the GDP distribution, and its measurement target is clearly distinct from those of the adjacent indicators. The series is the conditional expectation of GDP growth given the information set at each evaluation date, and the tail moments of that distribution are estimated by the site's GDP-at-Risk channel KRGAR (Adrian, Boyarchenko, and Giannone 2019). KRADS (Aruoba, Diebold, and Scotti 2009) is a coincident momentum factor, whereas KRGDPNOW is the conditional expectation of GDP growth itself, so the two measure different targets, and both are produced because the separated reading of a momentum gauge and of GDP growth is a meaningful information split in macro assessment.

## Applications in Financial Markets

KRGDPNOW sits naturally as the central-tendency input to Korean macro risk assessment. Where the site's GDP-at-Risk channel KRGAR (Adrian, Boyarchenko, and Giannone 2019) estimates the tail of the GDP distribution, KRGDPNOW estimates the center of that distribution from the information set at each evaluation date, and reading the two objects together separates the central tendency of the macro environment from its downside risk. In the nowcasting tradition Giannone, Reichlin, and Small (2008) established, this object is the foundational input that precedes every risk assessment.

In the Korean bond market KRGDPNOW carries meaning through two channels. First, GDP growth is a central input to the Bank of Korea policy reaction function. In the reaction function Taylor (1993) established the output gap is a primary determinant of the policy rate alongside inflation, and the GDP-growth information KRGDPNOW carries is usable for short-run updates of the policy-rate path. Second, the "growth trajectory" that the Bank of Korea Monetary Policy Decision statements customarily cite is a description that depends on quarterly data, and the monthly updates KRGDPNOW provides refresh the policy-reaction-function input at the monetary policy decision date roughly one quarter sooner.

The news decomposition is the tool that quantifies how much each data input revised the GDP nowcast at the time of a market shock (Bok et al. 2018). The site provides this in two forms, where the time-series form overlays realized GDP on the nowcast path across evaluation dates and the news-decomposition form refreshes the per-source-series contributions of the most recent twelve months each month and provides them directly to the user. This makes it possible to read explicitly how many basis points a one-month surprise in industrial production moved the GDP nowcast, which is used to assess the market reaction at the time of a data release.

The New York Fed Nowcast was the market-standard U.S. GDP nowcast object. This specification, documented in Bok et al. (2018), continues to be cited as the academic standard after its 2024 retirement, and KRGDPNOW is the Korean adaptation of it. Where the Higgins (2014) GDPNow is a different, regression-based U.S. implementation, KRGDPNOW is the Korean state-space implementation, and its difference from the Cleveland Fed specification Knotek and Zaman (2014) describe is of the same kind, in that KRGDPNOW excludes judgmental overlays.

Three operational caveats apply. First, Korea has no operationally usable official GDP vintages, so the site's pseudo-real-time backtest uses a lag mask, an operational approximation of the vintage standard Croushore (2011) and Stark (2008) define. That this approximation does not bias inference in the usual application of the equal-accuracy test Diebold and Mariano (1995) define is established by Doz, Giannone, and Reichlin (2011) together with the asymptotic justification of EM estimation. Second, because the panel begins in July 2008, the 1997 Asian financial crisis and the 2003 credit-card crisis cannot be evaluated with this indicator, and the backtest grid begins on 2013-01-01. Third, the 2020Q2 COVID-19 shock is a tail observation of the distribution, so it is flagged separately in RMSFE aggregation, following the out-of-sample outlier practice Stock and Watson (2016) recommend. These three limits are general limits common to every nowcast specification, and they are treated under the same discipline in the standard nowcasting reference Bańbura, Giannone, and Reichlin (2011) and the data-flow treatment of Bańbura et al. (2013).

## Statistical Tests

KRGDPNOW is a quarter-on-quarter real-GDP nowcast, a model-derived growth-rate object tested over 455 observations of the published path from 2013-01-31 to 2026-06-30 under a constant-plus-trend specification. The augmented Dickey and Fuller (1979) test with the lag augmentation of Said and Dickey (1984) rejects a unit root at p = 0.025, the Phillips and Perron (1988) test rejects at p = 0.000, and the KPSS test does not reject stationarity at p > 0.10 (Kwiatkowski et al. 1992), so the three procedures concur on a clean I(0) classification.

On the clean I(0) verdict the level is the stationary object, so its mean-break dates are citable. The Bai and Perron (1998) procedure computed by the algorithm of Bai and Perron (2003) locates three mean shifts at 2019-03-31, 2021-03-31, and 2023-03-31, all falling on quarter boundaries, read as level movements of a stationary series rather than stochastic-trend breaks (Perron 1989). On the stationary level the Ljung and Box (1978) portmanteau rejects the no-autocorrelation null at p = 0.000 at both the lag-12 and lag-24 rungs, confirming the high persistence of the nowcast path.

Because the nowcast has a realized target, its honest supplement is forecast evaluation against outturns over 145 quarterly forecast-realization pairs. The Diebold and Mariano (1995) test with the Harvey, Leybourne, and Newbold (1997) small-sample correction, computed on the loss differential against a naive random-walk benchmark under a Newey and West (1987) HAC, returns a mean loss differential of −0.5738 and a statistic of −2.53 with p = 0.012, so the nowcast significantly beats the naive benchmark. The Mincer and Zarnowitz (1969) efficiency regression of the realized outturn on the nowcast returns an intercept of 0.5554 and a slope of 0.1507, and the joint Wald test of unbiasedness gives F = 4.68 with p = 0.096, failing to reject unbiasedness at the five percent level on this sample.

Because the nowcast targets a seasonally-adjusted-at-source quarter-on-quarter growth rate, no calendar-seasonal component is defined, so the seasonal-frequency battery is not applicable and was not run (Ghysels and Osborn 2001).

## Frequently Asked Questions

### What does the real GDP nowcast estimate?

Real growth for a quarter already under way but not yet published, estimated close to real time by a mixed-frequency dynamic factor model fed with monthly activity indicators.

### How does a GDP nowcast differ from a forecast?

A nowcast estimates a quantity that has already been realised but not yet measured; a forecast concerns a period that has not happened. The nowcast's role ends the moment the official statistic is released.

### What does the nowcast news decomposition show?

It attributes each revision of the estimate to the individual release that caused it, so a reader can separate not merely that the estimate moved but which indicator moved it and by how much.