EM GDP NOWCASTS

← All countriesKorea · 2026Q3Last run 2026-09-11 02:38:29 UTC

Korea · real GDP, tracking

Tracker firms toward 4.2%

4.20%↑ 0.83 since 02 Sep

YoY · in-sample fit dispersion ±0.68pp · 16 runs this quarter · Trim-25 ensemble

Nowcast history — ensemble aggregations

Trim-25MeanMedianBest modelGDP outturn

One mark per scheduled run, placed at its actual run time, so several runs in a day sit close together and a pause shows as a gap. The shaded band is ±1.96×SE of the Trim-25 ensemble, where SE is the dispersion of its in-sample one-quarter-ahead fits (≈pp). It is computed on the same final-vintage data the models were fitted to, so read it as fit dispersion, not as a 95% prediction interval. The honest uncertainty measure is how far the models disagree — the indicator-disagreement figures below. The dashed rule is the last GDP outturn on record, for reference. A flat stretch means the model indicators did not move; each quarter is its own tracking exercise, and earlier quarters stay in the database.

Latest run — all models

Availability pattern: ip: months 1–1, ip_all: months 1–1, retail: months 1–1, pmi: months 1–2, bsi: months 1–2, lead: months 1–1, cpi_core: months 1–2, trade_bal: months 1–2, elec: 1m before quarter

Model	Ragged edge	YoY %	±1.96×SE	RMSE ≈pp	#RHS	#Specs
Trim-25	Ensemble	4.20	±0.68	0.35	5	816
BIC trimmed median	Ensemble	3.73	±1.17	0.60	9	820
Mean	Ensemble	3.73	±0.93	0.48	19	1641
Median	Ensemble	3.66	±0.92	0.47	19	1641
BIC trimmed mean	Ensemble	3.67	±1.14	0.58	9	820
Best model (MIDAS, Bridged)	Ensemble	5.31	±0.58	0.30	40	8
MIDAS	Bridged	5.31	—	0.30	40	8
MIDAS	Real-time	3.84	—	0.34	40	8
LASSO	Bridged	4.12	—	0.43	13	100
ElasticNet	Bridged	4.11	—	0.43	13	600
LASSO	Real-time	3.64	—	0.43	58	100
ElasticNet	Real-time	3.64	—	0.43	58	600
Bridge	Bridged	4.24	—	0.44	13	1
U-MIDAS	Real-time	3.41	—	0.44	36	1
Ridge	Bridged	3.98	—	0.45	13	100
Ridge	Real-time	3.66	—	0.46	58	100
U-MIDAS	Bridged	3.83	—	0.46	31	1
bsi	Single indicator	3.65	—	0.60	5	1
3PRF	Bridged	3.39	—	0.60	27	1
MF-VAR	Bridged	3.73	—	0.62	10	1
csi	Single indicator	3.07	—	0.63	5	1
MF-VAR	Real-time	3.38	—	0.64	10	1
Indicator mean	Ensemble	3.37	±1.27	0.65	8	8
3PRF	Real-time	2.86	—	0.65	9	1
ip_all	Single indicator	3.43	—	0.66	6	1
Indicator median	Ensemble	3.39	±1.30	0.66	8	8
cpi_core	Single indicator	3.34	—	0.67	5	1
retail	Single indicator	3.41	—	0.68	6	1
ip	Single indicator	3.38	—	0.69	6	1
elec	Single indicator	2.89	—	0.69	6	1
trade_bal	Single indicator	3.64	—	0.70	5	1
pmi	Single indicator	3.06	—	0.71	5	1
U-MIDAS-PC	Real-time	3.49	—	0.71	8	1
lead	Single indicator	4.10	—	0.71	6	1
MIDAS-PC	Real-time	3.36	—	0.72	8	8
MIDAS-PC	Bridged	3.14	—	0.72	8	8
U-MIDAS-PC	Bridged	3.73	—	0.73	7	1
DFM	Bridged	3.86	—	—	8	1
DFM	Real-time	3.83	—	—	9	4

39 rows · scroll within the table. Ragged edge: Bridged imputes the quarter's missing months by ARIMA before estimation; Real-time builds the specification from the observed cells only. #RHS = regressors in the final specification; #Specs = specifications tested (lag lengths / regularisation grid; for an ensemble, the total across its members). RMSE in 100×log terms under this run's availability pattern; dynamic factor models excluded by design.

Dispersion — nowcast against accuracy

Each mark is one model: circles are Real-time, diamonds Bridged. Marks lower on the chart fit better under today's information set; the dashed rule is the Trim-25 ensemble. A tight cluster near the rule means the accurate models agree.

Indicator manifest & data on file

9 of 11 in the joint baseline · 10 with their own model

Indicator	Freq	Role	Category	Transform	Last period	Arrived (UTC)	Obs
bsi	M	joint + own model	surveys	`d`	2026-08	2026-09-02 17:04	284
cpi_core	M	joint + own model	prices	`dlog`	2026-08	2026-09-09 16:02	560
elec	M	joint + own model	activity	`dlog`	2026-06	2026-09-02 17:04	462
ip	M	joint + own model	activity	`dlog`	2026-07	2026-09-02 17:04	559
ip_all	M	joint + own model	activity	`dlog`	2026-07	2026-09-02 17:04	319
lead	M	joint + own model	activity	`dlog`	2026-07	2026-09-02 17:04	559
pmi	M	joint + own model	surveys	`d`	2026-08	2026-09-02 17:04	269
retail	M	joint + own model	consumption	`dlog`	2026-07	2026-09-02 17:04	379
trade_bal	M	joint + own model	trade	`d`	2026-08	2026-09-02 17:04	560
csi	M	own model	surveys	`d`	2026-08	2026-09-02 17:04	218
rgdp	Q	tracked	—	`—`	2026Q2	2026-09-02 17:04	186

The indicator manifest, and where the tracker stands whether or not anything printed today. joint + own model = in the multivariate baseline AND carrying its own single-indicator model; own model = single-indicator only (usually because its history is too short to join the baseline without truncating every joint model's sample); tracked = stored but not modelled. Transform: dlog = log difference, d = first difference (diffusion indices, rates and anything that can go negative). Last period is the newest observation held, Arrived when it first appeared in a download. Reference periods and dates only; indicator values are not redistributed.

New data releases

When (UTC)	Indicator	First ref	Last ref	# obs
Nothing yet. The log records only genuinely new prints, so it stays empty between releases — see the data-on-file table for what the nowcaster is currently holding.

One line per indicator per run: the range of reference periods that arrived.

Single-indicator models — who says what

10 indicators · median 3.41% · range 2.89–4.10%

Indicator	Category	YoY %	RMSE ≈pp
elec	activity	2.89	0.688
pmi	surveys	3.06	0.708
csi	surveys	3.07	0.627
cpi_core	prices	3.34	0.670
ip	activity	3.38	0.685
retail	consumption	3.41	0.676
ip_all	activity	3.43	0.662
trade_bal	trade	3.64	0.698
bsi	surveys	3.65	0.604
lead	activity	4.10	0.715

One U-MIDAS per indicator, each with a single autoregressive lag and its own estimation sample — so a series with a short history costs nothing to any other model. The spread across these rows is the interpretable uncertainty measure on this page: it says how much the answer depends on which indicator you believe. It is a measure of disagreement, not a probability distribution.

Out-of-sample scoreboard

archive vs first-release outturns

Model	Family	RMSE pp	MAE pp	Bias pp	N
Not enough scored quarters yet. A quarter is scored when its GDP prints: the last nowcast made before that release is compared with the first-release outturn. This is the only genuinely out-of-sample table on the page.

For every quarter whose GDP has since printed, the last nowcast recorded before that release is scored against the first-release outturn. Errors are percentage points of YoY growth. Unlike the in-sample RMSE table above, nothing here was seen by the models when they were fitted — which is why N matters as much as RMSE: until this table has a couple of years in it, treat it as descriptive and do not weight models by it.

Machine-learning cross-validation

Model	Ragged edge	Folds	α grid	Chosen α	L1 ratio	CV RMSE	Active
LASSO	Bridged	5	100	0.00011	—	0.0054	12
ElasticNet	Bridged	5	100	0.00118	0.10	0.0054	12
LASSO	Real-time	5	100	0.000265	—	0.0061	35
ElasticNet	Real-time	5	100	0.000279	0.95	0.0061	35
Ridge	Bridged	5	100	18.7	—	0.0053	13
Ridge	Real-time	5	100	47.5	—	0.0066	58

Regularisation strength chosen by k-fold cross-validation, re-tuned each run for the current availability pattern; Active = regressors with non-zero coefficients.

EM GDP nowcaster · model outputs onlyUnderlying indicator data not redistributed