Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
58,607 characters · 11 sections · 30 citation commands
A Korean Macroeconomic Database for Data-Rich Policy Analysis and U.S.--Korea Dependence
Large macroeconomic panels have become standard inputs in empirical macroeconomics. Early data-rich studies showed that large sets of indicators improve forecasting and help summarize aggregate fluctuations, but this line of work often relied on datasets that required substantial manual curation. FRED-MD, developed by mccracken2015fredmd, changed this practice by providing a public monthly database with standardized transformations and documentation, and it has become a leading benchmark for transparent and reproducible macroeconomic research. Related efforts have also emerged outside the United States. For Canada, fortingagnon2022large provide a large macroeconomic database with public real-time vintages; for the euro area and its member countries, barigozzi2024large develop a harmonized dataset for comparative macroeconomic and policy analysis; and for Korea, kimandswanson2018 assemble real-time macroeconomic data for GDP backcasting, nowcasting, and forecasting. These studies differ in scope and are often organized around specific forecasting exercises rather than a general-purpose monthly macroeconomic database.
This paper constructs KRED, a FRED-MD-compatible monthly macroeconomic database for Korea. Its primary purpose is to provide a transparent and reusable foundation for data-rich policy analysis in a small open economy that is highly exposed to external monetary and financial conditions. This perspective is closely related to the broader small-open-economy literature that emphasizes the importance of foreign disturbances and external financial conditions for domestic dynamics; see, for example, justiniano2010smallopen,rey2015dilemma. It is also directly connected to the literature on the international transmission of U.S. monetary policy shocks, beginning with kim2001usmp. For Korea and Asia, related evidence shows that external monetary conditions matter for Korean capital flows and trade and for the lending behavior of Asian banks more generally; see ree2014safehaven and lee2022spilloversasia. At the same time, close alignment with the FRED-MD architecture makes KRED useful for comparative work and for cross-country models that combine Korea with other economies in a standardized way. Such standardization is particularly valuable when country-specific panels are brought into multicountry VARs and more structured matrix- and tensor-based dynamic models; see, for example, canova2009estimating,canova2013panel,hill2021tensor,li2021multilinear,tsay2024matrix,wang2024tensorar,luo2025bayesian. In this sense, KRED is designed not only as a domestic macroeconomic database for Korea but also as an empirical infrastructure for studying international macro-financial dependence.
Constructing such a database for Korea is nontrivial because relevant series are dispersed across multiple public platforms rather than provided through a single integrated system. KRED consolidates data from the ECOS system of the Bank of Korea, the KOSIS portal from the Ministry of Data and Statistics, and administrative labor-market statistics from the Ministry of Employment and Labor within a reproducible workflow. To facilitate transparency, replication, and future empirical work, the data repository is publicly available at \url{https://github.com/crbaek/KRED} and will be updated regularly. The initial release contains 125 monthly series dating back to 1960:01, and our empirical illustrations use a balanced panel of 104 series over 2009:06--2025:12. Principal-components analysis extracts four factors that explain about 30% of total variation. These factors correspond to financial conditions, real activity, housing and real-estate credit, and labor-market and price pressures, and their diffusion indices summarize major Korean macroeconomic episodes, including the pandemic contraction and the 2022--2023 tightening cycle.
We illustrate the empirical value of KRED through two applications that are directly relevant for Korean and Asian macroeconomic analysis. First, factor-augmented VARs in the spirit of bernanke2005measuring show that U.S. monetary tightening transmits strongly to Korea, and the preferred specification yields a more coherent inflation response than a low-dimensional VAR. Second, because KRED shares a common grouped architecture with FRED-MD, it is well suited to structured cross-country analysis. We use this feature to estimate a grouped U.S.--Korea tensor autoregression based on country--group factors. The tensor specification does not dominate separate-country VARs in point-forecast accuracy, but it yields a sharper structural result: cross-country dependence is concentrated in financially oriented blocks, with stronger transmission from the U.S. financial block to Korea than in the reverse direction, while spillovers in real activity and housing are much weaker.
These results position KRED as more than a data archive. It provides a transparent public database for Korean macroeconomic research, a practical input for data-rich policy analysis, and a standardized building block for comparative work on cross-country macro-financial analysis. This combination is particularly relevant in an Asian setting, where external monetary conditions and cross-border financial linkages are central to the interpretation of domestic macroeconomic fluctuations.
The remainder of the paper is organized as follows. Section 2 describes the construction of KRED and the main design choices. Section 3 presents the factor analysis and diffusion indices. Section 4 studies monetary policy shocks using FAVAR specifications. Section 5 introduces a grouped U.S.--Korea tensor autoregression and uses it to study cross-country macro-financial dependence. Section 6 concludes. The Appendix reports a supplementary real-activity impulse-response contrast and the mapping between KRED series and their FRED-MD counterparts.
KRED serves two related purposes: it is a general-purpose monthly macroeconomic database for Korea, and it is a standardized country block aligned as closely as possible with FRED-MD. This section documents the data sources, grouping and transformation rules, and the main departures required by Korean institutional features.
Our data come from three public sources: the Economic Statistics System (ECOS) of the Bank of Korea (\url{https://ecos.bok.or.kr/}), the KOSIS portal (\url{https://kosis.kr/index/index.do}) from the Ministry of Data and Statistics, and employment statistics from the Ministry of Employment and Labor (\url{https://laborstat.moel.go.kr/}). KRED contains 125 time series, of which 104 are used in the empirical analysis over 2009:06--2025:12. Variables are organized into eight groups: (i) output and income, (ii) labor market, (iii) housing, (iv) consumption, orders, and inventories, (v) money and credit, (vi) interest rates and exchange rates, (vii) prices, and (viii) the stock market. Transformation codes follow FRED-MD as closely as possible, and Appendix (ref) reports both the transformation codes and the series-level mapping to FRED-MD counterparts.
While the overall structure aligns with FRED-MD, several important differences reflect the unique characteristics of Korean macroeconomic data:
To illustrate the empirical content of KRED, we begin with a static factor analysis in the spirit of mccracken2015fredmd. We work with a balanced monthly panel covering 2009:06--2025:12 ($T=199$) and $q=104$ series. Relative to the FRED-MD benchmark set, the balanced panel excludes the following series due to incomplete coverage over the sample window: USGOOD, CES1021000001, USCONS, MANEMP, DMANEMP, NDMANEMP, \texttt{SRVPRD}, \texttt{USTPU}, \texttt{USWTRADE}, \texttt{USTRADE}, \texttt{USFIRE}, \texttt{USGOV}, \texttt{HOUST}, \texttt{HOUSTNE}, \texttt{HOUSTMW}, \texttt{HOUSTS}, \texttt{HOUSTW}, \texttt{RETAILx}, \texttt{ACOGNO}, \texttt{TOTRESNS}, \texttt{DTCOLNVHFNM}, and \texttt{EXKRCNx}.
Let $x_t\in\mathbb{R}^{q}$ denote the vector of transformed observations at month $t$, and stack the panel as \[ X=\bigl(x_1,\ldots,x_T\bigr)\in\mathbb{R}^{q\times T}. \] We use the static approximate factor model \[ X=\Lambda F+E, \qquad \Lambda\in\mathbb{R}^{q\times r},\quad F\in\mathbb{R}^{r\times T}, \] so that $x_t=\Lambda f_t+e_t$. We first center each series in $X$ to have zero time mean and form \[ S :=\frac{1}{T}XX'\in\mathbb{R}^{q\times q}. \] Let $S=VDV'$ be the eigen-decomposition with eigenvalues $d_1\ge\cdots\ge d_q\ge 0$, and let $V_r\in\mathbb{R}^{q\times r}$ collect the first $r$ eigenvectors. We estimate loadings and factors by \[ \widehat\Lambda(r)=\sqrt{q}\,V_r, \qquad \widehat F(r)=\frac{1}{q}\widehat\Lambda(r)'X\in\mathbb{R}^{r\times T}, \] which implies $(1/q)\widehat\Lambda(r)'\widehat\Lambda(r)=I_r$. With $\widehat f_t(r)$ the $t$-th column of $\widehat F(r)$ and $\widehat\lambda_i(r)'$ the $i$th row of $\widehat\Lambda(r)$, the fitted common component and residual are \[ \widehat x^{\,c}_{it}(r)=\widehat\lambda_i(r)'\widehat f_t(r), \qquad \widehat e_{it}(r)=x_{it}-\widehat\lambda_i(r)'\widehat f_t(r), \quad i=1,\ldots,q,\ \ t=1,\ldots,T. \] For more details, see, for example, bai2008large, forni2000generalized, stock-watson-DFM.
To select the number of factors $r$, we use the information criteria of bai2002determining, \[ \mathrm{IC}(r) = \log\!\left(\frac{1}{qT}\sum_{i=1}^q\sum_{t=1}^T \widehat e_{ti}^2(r)\right) + r\cdot g(q,T), \] with penalty functions \[
\] where $q\wedge T=\min(q,T)$. In our application, the minimizers are $\widehat r=3,3,5$ under $g_1,g_2,g_3$, respectively. Figure (ref) complements this choice by displaying the eigenvalue profile (bars) together with the cumulative variance share.
We proceed with $r=4$ factors. This choice balances parsimony and coverage: four factors explain 29.79% of the total variation (compared with 34.50% for the FRED-MD benchmark), while eight factors raise the cumulative share to 46.37%, close to the FRED-MD benchmark (47.6%). Figure (ref) in the Appendix plots the four estimated factor scores.
To quantify how each factor maps into economically meaningful blocks of variables, we follow the incremental $R^2$ decomposition used in mccracken2015fredmd. For each series $i$ and each $k=1,\ldots,r$, regress $x_{ti}$ on the first $k$ estimated factors and denote the resulting coefficient of determination by $R_i^2(k)$. Define the incremental explanatory power of factor $k$ for series $i$ as $$ mR_i^2(1)=R_i^2(1), \qquad mR_i^2(k)=R_i^2(k)-R_i^2(k-1),\quad k=2,\ldots,r. $$ Large values of $mR_i^2(k)$ indicate that factor $k$ adds substantial explanatory content for series $i$ beyond what is already captured by the previous factors. Figure (ref) displays the cross-sectional distribution of $mR_i^2(k)$ by group, and Table (ref) reports the top contributors in $mR_i^2(k)$ for $k=1,\ldots,4$. Two patterns stand out. First, the first two factors are jointly spanned by real-activity variables (Group 1) and interest-rate/spread variables (Group 6), but with opposite emphasis: Factor 1 is dominated by yield-curve and credit-spread measures, whereas Factor 2 is anchored by manufacturing and goods-production indicators. Second, Factor 3 remains sharply concentrated in housing variables, with additional contributions from credit aggregates, while Factor 4 is led by labor-market quantities together with consumer and producer prices. These concentration patterns, together with the top-$mR_i^2$ entries in Table (ref), motivate the interpretations below.
\noindentFactor 1: Financial conditions (term structure and credit spreads).\; The first factor is led by slope and spread variables such as T3YFFM, AAAFFM, T1YFFM, TB6SMFFM, T10YFFM, and \texttt{BAAFFM}, while manufacturing indicators such as \texttt{IPMANSICS}, \texttt{INDPRO}, and \texttt{CUMFNS} enter with smaller incremental gains. This composition points to a financial-conditions factor that summarizes the shape of the yield curve and corporate credit premia. Because medium- and long-maturity slope measures dominate the factor, it is best interpreted as a forward-looking macro--financial indicator rather than a coincident output index.
\noindentFactor 2: Real activity (manufacturing/goods cycle).\; The second factor is anchored by production and utilization measures, including IPMANSICS, IPCONGD, IPFINAL2, INDPRO, and CUMFNS, with \texttt{IPDCONGD} and \texttt{IPDMAT} reinforcing the goods-sector interpretation. Term-spread variables still enter the top-$mR_i^2$ list, but with clearly smaller contributions than in Factor 1. This pattern indicates a broad manufacturing and goods-cycle factor that captures cyclical comovement in output and utilization, with financial variables reflecting their close linkage to production demand.
\noindentFactor 3: Housing and real-estate credit (early-cycle demand).\; The third factor remains housing-centered: PERMIT and the regional permit series dominate the top-$mR_i^2$ list. At the same time, REALLN and BUSLOANS enter prominently, and PCEPI together with selected wage variables also contribute. The resulting pattern suggests a housing/construction factor with an associated credit margin. It captures a forward-looking real-estate block in which permit issuance and property-related lending move together, alongside selected expenditure-price and wage measures.
\noindentFactor 4: Labor market and prices (cost-pressure dimension).\; The fourth factor is led by labor-market variables---CE16OV, PAYEMS, CLF16OV, UNRATE, UEMP5TO14, \texttt{UEMPMEAN}, and \texttt{AWOTMAN}---while price measures such as \texttt{CUSR0000SAC}, \texttt{WPSFD49502}, and \texttt{WPSFD49207} provide a secondary but meaningful contribution. Economically, this factor represents labor utilization, slack, and associated cost conditions. The joint dynamics of employment, unemployment-duration, and price variables is consistent with a latent pressure-slack dimension underlying the Phillips curve relationship.
With all four factors included, Figure (ref) plots $R_i^2(4)$ in decreasing order. The best-explained series are drawn from four connected blocks: yield-curve and credit-spread measures, manufacturing production and utilization, housing permits, and labor-market quantities. This ranking is informative. Yield spreads and corporate premia summarize expected policy and risk compensation bernanke1995inside,gurkaynak2005sensitivity, housing permits reflect a forward-looking investment margin strauss2013does, and manufacturing output and utilization record the coincident goods-sector cycle stock1989new,gilchrist2012credit. The prominent appearance of employment and labor-force variables further indicates that the updated four-factor structure captures a broad macro--financial core rather than a narrowly defined activity block. Variables farther down the ranking are increasingly heterogeneous and sector-specific.
Following mccracken2015fredmd, we construct factor-based diffusion indices to summarize low-frequency dynamics implied by the estimated common components. Rather than tracking the share of series increasing at each date, the factor-based index cumulates the latent factor itself: $$ \widehat{\mathrm{FDI}}_{t}(k)=\sum_{s=1}^{t}\widehat f_{s}(k),\qquad t=1,\ldots,T, $$ where $\widehat f_{s}(k)$ is the $k$-th element of the factor score at time $s$. This cumulation accentuates persistent expansions and contractions. For interpretability, we orient Factor 3 so that higher values correspond to stronger housing activity.
Figure (ref) plots the four diffusion indices and highlights several episodes. Financial-conditions diffusion ($F_1$) falls sharply during 2009--2011, trends upward through the mid- and late-2010s, turns down in 2022--2023, and rises strongly in 2024--2025, reflecting large cyclical swings in the common term-spread and credit-premium component. Real-activity diffusion ($F_2$) rises during the early post-crisis recovery, trends downward through the mid-2010s, records a pronounced but short-lived upswing in 2022--2023, and reverses by 2025. It is therefore best read as the common manufacturing and goods-production margin rather than headline aggregate activity. Housing-and-real-estate-credit diffusion ($F_3$) is weak through the first half of the sample but trends upward from about 2016, reaches a local high around 2023, and then eases. Its smoother medium-run trajectory, relative to $F_1$ and $F_2$, suggests that permit issuance and property-related credit respond to housing-specific supply, financing, and policy conditions rather than one-for-one with the broader macro cycle.
The pandemic episode appears less as a synchronized collapse than as timing differences across factors. Around 2020, $F_1$ reaches a local high, $F_2$ is muted, $F_3$ continues its recovery after only a brief interruption, and $F_4$ remains positive but volatile. More persistent adjustment emerges after 2022. $F_1$ moves down first, the manufacturing block turns sharply, housing softens with delay, and labor-market-and-prices diffusion ($F_4$) declines after 2023, consistent with gradual easing in shared labor-market and cost pressures.
These diffusion patterns suggest a sequencing in which financial conditions are the most cyclical and forward-looking margin, manufacturing exhibits episodic swings, housing follows a slower medium-run cycle with a credit component, and labor-market and price pressures adjust most gradually. The four-factor structure therefore summarizes major Korean macroeconomic episodes while separating macro--financial, real, housing, and cost-pressure forces.
To further assess the practical value of KRED for data-rich policy analysis, we estimate factor-augmented vector autoregressions (FAVARs) in the spirit of bernanke2005measuring and study monetary policy transmission mechanism in Korea. The FAVAR framework conditions on a large information set. It therefore permits impulse-response analysis for a broad set of macroeconomic and financial variables that would be difficult to study jointly in a standard VAR.
To maintain comparability with bernanke2005measuring, we consider two FAVAR specifications that mirror their empirical design. Korea is a small open economy, so we also use a four-variable baseline VAR that augments the standard three variable system--real activity, inflation, and the domestic policy rate--with the U.S.\ policy rate. The U.S.\ federal funds rate (US.FFR, taken from FRED-MD) enters both FAVAR specifications for the same reason. It proxies for external monetary and financial conditions that are plausibly exogenous to Korea at a monthly frequency.
We compare three closely related models that exploit the informational content of KRED while remaining comparable to standard VAR practice.
Interest rates enter in levels across all models. CPI and IP follow common VAR and FAVAR conventions.\footnote{CPI is treated as an inflation-rate measure with transformation code 5 to maintain comparability with bernanke2005measuring. The unemployment rate UNRATE is included in levels with transformation code 1, following bernanke2005measuring. We also do not apply additional transformations to US.FFR or KR.MIR.} FAVARs are estimated using the standard two-step principal-components approach. We first extract factors from the large panel and then estimate a VAR in the observed block augmented by the estimated factors. In this section, factors are extracted after excluding any observed variables that enter the VAR block. This prevents the same series from affecting the system both directly and through the estimated factors and keeps the identification scheme aligned with the intended information set. Structural shocks are identified recursively. In all specifications, \texttt{US.FFR} is ordered first to reflect the small-open-economy assumption that Korean variables do not contemporaneously affect U.S.\ policy at a monthly frequency. \texttt{KR.MIR} is ordered last so that domestic policy can respond within the month to innovations in the preceding block.
The reported impulse responses include KR.MIR, IP (INDPRO), CPI (CPIAUCSL), real GNI, real \texttt{GDP}, the unemployment rate (\texttt{UNRATE}), real \texttt{M2} (\texttt{M2REAL}), total housing permits (\texttt{PERMIT}), the 6-month and 5-year term spreads (\texttt{TB6SMFFM} and \texttt{T3YFFM}), the KRW--USD exchange rate (\texttt{EXKRUSx}), and the stock market index (\texttt{KOSPI}).
The analysis yields two main findings. First, information extracted from a large-scale dataset is useful for studying the Korean economy, and it helps address the price puzzle that often arises in conventional VARs. The price puzzle refers to a positive short-run response of prices or inflation to a contractionary monetary policy shock. Second, factor augmentation relaxes the dimensionality constraint of small VARs and allows impulse responses to be traced for a broader set of macroeconomic and financial variables. The resulting responses align with standard transmission mechanisms and provide an empirically coherent account of monetary tightening.
Figure (ref) reports responses to a 25-basis-point innovation in KR.MIR. The broad transmission mechanism is similar across specifications. KR.MIR rises on impact and then mean-reverts gradually. Real activity, measured most clearly by IP, weakens with a lag, and financial variables such as M2REAL and KOSPI also move in a contractionary direction. For brevity, we do not report PERMIT, whose response is weak and imprecise, suggesting limited sensitivity of permit issuance to short-rate fluctuations at the monthly horizon and the possible importance of housing-specific supply and regulatory frictions. The main difference across specifications lies in the inflation response. In the preferred specification, \texttt{CPI} moves little on impact and then declines gradually, yielding a comparatively clean disinflation profile. In the alternative specification, by contrast, \texttt{CPI} displays a noticeable positive short-run response before turning negative at longer horizons. Responses of \texttt{GNI_real} and \texttt{GDP_real} are directionally consistent with monetary tightening, but they are weaker and less stable than the response of \texttt{IP}.
Figure (ref) reports responses to a 25-basis-point increase in US.FFR under the preferred and alternative FAVARs. The two specifications deliver qualitatively similar real-side responses. U.S. tightening induces contractionary effects in Korea: IP, GDP_real, and GNI_real decline, CPI falls over the medium horizon, and KR.MIR increases with a lag, indicating an endogenous domestic policy response to external conditions. Term spreads, especially \texttt{TB6SMFFM} and \texttt{T3YFFM}, compress on impact in a manner consistent with a flatter yield curve. By contrast, financial-price variables such as \texttt{EXKRUSx}, \texttt{KOSPI}, and, to a lesser extent, \texttt{M2REAL}, are more sensitive to specification. We therefore report these responses but do not attach a strong structural interpretation to their exact paths.
Figure (ref) compares the preferred and alternative FAVARs with the VAR benchmark and highlights the role of factor augmentation in identification. For a KR.MIR shock, the inflation response differs markedly across models. The VAR benchmark exhibits a short-run price puzzle, reflected in a positive short-run response of CPI on impact, which also persists in the alternative FAVAR. By contrast, the preferred FAVAR eliminates the price puzzle and yields a smooth disinflation profile. For a US.FFR shock, both FAVAR specifications imply more clearly disinflationary responses of CPI than the VAR benchmark, with the preferred specification producing the smoother profile and the alternative the larger response. These differences are consistent with the view that KRED factors absorb omitted-information components relevant for identifying monetary policy shocks. Some open-economy responses, especially those of EXKRUSx and KOSPI, remain specification-sensitive and should be interpreted with caution.
The FAVAR results deliver two empirical messages. First, KRED-based factor augmentation is useful for tracing monetary policy transmission in Korea, especially in the preferred specification, which yields the most coherent inflation response to a domestic policy shock. Second, U.S. monetary tightening transmits strongly to Korea, weakening real activity and inducing a delayed increase in KR.MIR. These findings support the view that KRED is not only a data repository but also a useful input for modern, data-rich macro--financial analysis.
This section uses the common grouped architecture of KRED and FRED-MD to study where U.S.--Korea dependence is concentrated across macroeconomic blocks. The motivating question is economic rather than purely methodological. Korea is a small open economy, and the previous FAVAR results already suggest that external monetary conditions, especially U.S. tightening, matter for Korean macroeconomic dynamics. The remaining question is whether this dependence is broad-based across real and financial sectors, or whether it is concentrated in a narrower set of transmission channels.
To address this question, we construct grouped monthly factors for Korea and the United States using a common eight-group classification. This grouped design allows us to compare the two countries within a unified framework and to distinguish broad real-side comovement from sharper macro-financial spillovers. In particular, it allows us to assess whether cross-country dependence is strongest in real activity, housing, money and credit, financial conditions, or equity-market blocks.
Our empirical findings point to a layered dependence structure. There is some broad linkage across real-side blocks, but the clearest and most persistent cross-country transmission is concentrated in financially oriented blocks. In particular, the financial-conditions block plays a central role: shocks from the U.S. financial block generate a sizable and persistent response in the Korean financial block, whereas the reverse direction is much weaker. By contrast, spillovers in real activity and housing are more limited and less directional. The contribution of the grouped tensor analysis is therefore structural rather than predictive. Its value lies in clarifying where cross-country dependence is strongest and how its direction differs across blocks.
Let $X_t \in \mathbb{R}^{2 \times 8}$ denote the monthly grouped-factor matrix, where the two rows correspond to Korea and the United States and the eight columns correspond to the common macroeconomic groups. We model $\{X_t\}$ by a matrix-valued tensor autoregression, which is a second-order special case of the tensor autoregressive models studied by chen2021matrixar,hill2021tensor:
where $A_c^{(ir)} \in \mathbb{R}^{2 \times 2}$ governs dependence across countries, $A_g^{(ir)} \in \mathbb{R}^{8 \times 8}$ governs dependence across groups, and $E_t$ is a mean-zero innovation matrix. For stability analysis and impulse-response calculations, we use the vectorized representation
with $\varepsilon_t = \mathrm{vec}(E_t)$.
The sample runs from 2009:06 to 2025:12. Quarterly series and sporadic missing observations are linearly interpolated before the transformation codes are applied. For each country $c \in \{\mathrm{KR}, \mathrm{US}\}$ and each group $g \in \{1,\ldots,8\}$, we extract the first principal component $f_{cgt}$ from all transformed series in that block, yielding \[ X_t =
, \qquad t = 1,\ldots,T. \] This construction does not impose exact one-to-one matching with the 104-series single-country panel. For the present purpose, the relevant common object is the grouped architecture itself, and using all available series within each country--group block preserves more information.
We estimate (ref) by least squares using the tensorTS package chen2025tensorts, and use maximum likelihood as a robustness check. Tuning parameters are selected by BIC subject to stability of the companion form. We search over $P \in \{1,2,3,6,12,13\}$, with $R_i \equiv R \in \{1,2\}$ for $P \le 3$ and $R_i \equiv 1$ for $P \in \{6,12,13\}$. The selected specification has $P=2$ and $R=2$, with maximum companion-root modulus 0.958.
Two sign normalizations are important for interpretation. The financial factor G6 is oriented so that larger values correspond to higher rates and wider spreads, so a positive shock represents tighter financial conditions. The stock-market factor G8 is oriented so that larger values correspond to stronger equity prices, so a negative response represents stock-market weakness.
We report two recursive expanding-window forecast exercises as secondary diagnostics. The first evaluates one-step-ahead forecasts over 2024:01--2025:12, using 2009:06--2023:12 as the initial training sample. The second evaluates the shorter 2025:01--2025:12 window, using 2009:06--2024:12 as the initial training sample. The benchmarks are country-specific VARs whose lag orders are selected by BIC over the same grid; the selected lag is $1$ for both countries. Impulse responses are computed from the model refitted on the full sample. The main text reports orthogonalized responses under a U.S.-first financial ordering, and the Appendix reports a supplementary real-activity contrast.
Figure (ref) reports the lag-specific coefficient matrices $\widehat\Phi_1$ and $\widehat\Phi_2$ implied by the selected grouped tensor model. Two features stand out. First, the main dependence structure is concentrated in $\widehat\Phi_1$, while $\widehat\Phi_2$ is weaker and more diffuse. Second, the largest coefficients are still found within country blocks, so the dominant pattern remains one of strong domestic persistence rather than broad cross-country coupling. Within this overall structure, the consumption / orders / inventories block ($G4$) is notable. In both countries, $G4$ is visibly connected to the real-side blocks $G1$--$G4$, which suggests that this block acts as a broad real-side adjustment margin linking output, labor, and demand-related fluctuations. Some cross-country entries involving $G4$ are also more visible than those of several other nonfinancial blocks. This pattern is consistent with an international demand--orders--inventories channel, although the coefficient matrices alone do not indicate a strongly directional spillover.
Forecasting evidence remains secondary. Table (ref) shows that the grouped tensor specification ($P=2$, $R=2$) does not outperform the separate-country VAR(1) benchmarks in average one-step-ahead forecasting over either the 2024--2025 or the 2025-only evaluation window. Over 2024--2025, the average RMSE is $1.788$ for Korea under TenAR and $1.737$ under the Korean VAR benchmark, while the corresponding U.S.\ values are $1.586$ and $1.556$. Over 2025 only, the average RMSE is $1.873$ for Korea under TenAR and $1.728$ under the Korean VAR benchmark, while the corresponding U.S.\ values are $1.575$ and $1.540$. The grouped tensor model therefore does not dominate in point-forecast accuracy. Its main value is structural rather than predictive.
Figure (ref) reports orthogonalized impulse responses under the U.S.-first financial ordering and should be interpreted separately from the coefficient matrices in Figure (ref). The coefficient matrices summarize lag-specific linear dependence, whereas the impulse responses trace the propagation of orthogonalized shocks through the full dynamic system. Under this ordering, the strongest cross-country propagation operates through the financial-conditions block $G6$. A positive U.S.\ financial shock generates a sizable and persistent response in the Korean financial factor, whereas the response of the U.S.\ financial block to a Korean financial shock remains comparatively small. The money-and-credit block $G5$ moves in the same broad direction as $G6$ in the impulse responses, but with smaller magnitude and faster decay, so it is better interpreted as a secondary receiving margin than as a primary propagation channel. The stock-market block $G8$ displays a similar but weaker asymmetry. After a positive U.S.\ financial shock, the U.S.\ equity factor declines clearly, whereas the Korean equity response is smaller and less persistent. By contrast, a Korean financial shock generates little systematic response in the U.S.\ equity factor. The grouped tensor model therefore suggests that the main directional transmission runs from the U.S.\ financial block to the Korean financial block, with weaker spillovers to money and credit and then to equity markets.
This directional evidence also clarifies how to read the coefficient matrices in Figure (ref). Those matrices indicate that the consumption / orders / inventories block $G4$ is broadly connected to the real-side blocks $G1$--$G4$ in both countries, which is consistent with a real-side adjustment channel linking output, labor, and demand-related fluctuations through orders and inventories. However, the impulse responses show that these real-side links are less directional and less persistent across countries than the financial transmission documented for $G6$. The grouped tensor results therefore point to a layered dependence structure: a broad real-side linkage centered on $G4$, and a sharper, more asymmetric cross-country transmission mechanism centered on financial conditions.
This paper introduces KRED, a FRED-MD-compatible monthly macroeconomic database for Korea. KRED consolidates 125 monthly series from major public sources into a standardized and documented panel, and the empirical analysis is based on a balanced subset of 104 series over 2009:06--2025:12. The factor analysis shows that a relatively parsimonious representation already captures an important part of the common variation in the Korean macroeconomic panel. Principal-components analysis extracts four factors that explain about 30% of total variation. These factors correspond to financial conditions, real activity, housing and real-estate credit, and labor-market and price pressures. Their diffusion indices summarize major Korean macroeconomic episodes while separating financial, real, housing, and cost-pressure dynamics.
KRED is also useful for data-rich policy analysis. In FAVAR specifications that include the U.S. federal funds rate, U.S. monetary tightening transmits strongly to Korea, weakens real activity, and induces a delayed increase in the domestic policy rate. For domestic monetary shocks, the preferred factor-augmented specification yields a more coherent inflation response than the low-dimensional VAR benchmark. These results show that a large and standardized Korean macroeconomic panel improves the empirical analysis of monetary transmission in a small open economy.
The grouped U.S.--Korea tensor analysis adds a complementary structural perspective. Although the grouped tensor autoregression does not outperform separate-country VAR benchmarks in average one-step-ahead forecasting, it reveals a clear pattern in cross-country dependence. The strongest and most persistent international transmission is concentrated in financially oriented blocks, especially financial conditions, with smaller spillovers to money and credit and equity markets, and much weaker transmission in real activity and housing. This pattern suggests that U.S.--Korea dependence is not diffuse across all sectors but is organized around a narrower macro-financial transmission mechanism.
KRED therefore contributes along two margins. First, it provides a transparent public database for Korean macroeconomic research in a format that is closely aligned with FRED-MD. Second, it provides a practical building block for comparative work on external dependence and macro-financial transmission in Asia. These features make KRED useful not only for domestic business-cycle monitoring and policy analysis, but also for future cross-country research in standardized data-rich environments.
The KRED data repository is publicly available at \url{https://github.com/crbaek/KRED}. The repository contains the database files, documentation, and update scripts, and it will be updated regularly.
This research was initiated during the first author's visit to Professor M. C. D\"uker in February 2025 at the Friedrich-Alexander University of Erlangen-Nuremberg. The authors are grateful for her generous hospitality and the intellectual inspiration that helped shape the early development of this project. The authors also thank Professor Vladas Pipiras at the University of North Carolina at Chapel Hill for his valuable comments, which substantially improved the quality and clarity of the paper.
{ {4pt} }