thainowcast: pseudo-real-time nowcasting data for Thai GDP and inflation

Most of this dataset is synthetic. The panel in data/ and every result in results/ come from generated test data, because the environment the project was built in had no access to the Thai statistical agencies. They exercise the pipeline and say nothing about the Thai economy. The one real-data file is geo/province_night_lights.csv.

Data for the working paper ThaiNowcast: Pseudo-Real-Time Nowcasting of Thai GDP and Inflation with Official and Alternative Data by Thanatchon Chongmankhong (read online · PDF). The paper nowcasts quarterly real GDP growth and monthly headline CPI inflation from official monthly statistics and five families of alternative data (Google Trends, nighttime lights, Sentinel-5P NO2, online prices, mobility), with a model race, regime-split evaluation, Diebold-Mariano tests, conformal intervals, an ablation per alternative data family and a Kalman-filter news decomposition. Code: https://github.com/thanatchon36/thainowcast; archived on Zenodo: 10.5281/zenodo.23995547.

Independent personal research. Does not represent the views of any employer or institution.

Files

path	content
data/panel.csv	synthetic. The pseudo-real-time panel: 5,708 observations of 26 series (series, freq M or Q, period, value, release), January 2004 to 2026Q3. release is the date the value is assumed published; a model at origin t sees only rows with release <= t
data/indicators.csv	the registry: English and Thai name, group (target, hard, five alternative families), agency, frequency, assumed publication lag in days, and the last period released as of 2026-11-05. Lags are approximations
data/raw/synthetic_*.csv	synthetic. The same 26 series as one period,value file each, in the layout the pipeline's loader expects; synthetic_manifest.json records the seed and the data-generating process
results/backtest_gdp_yoy.csv	synthetic. Expanding-window backtest of GDP growth, 2015Q1 to 2026Q2: one row per period, origin (M1, M2, M3, PR = the day before the official release), model and feature set, with nowcast, first-release actual, error and the 80% conformal interval (3,680 rows)
results/backtest_cpi_yoy.csv	synthetic. The same for CPI inflation, 2015-01 to 2026-09, origins M0 (mid-month) and PR (6,204 rows)
results/summary_*.csv	RMSFE and MAE by model, feature set, origin and regime (full, pre-COVID, COVID, post-COVID), relative to the AR benchmark
results/dm_*.csv	Diebold-Mariano tests against the AR benchmark with the Harvey-Leybourne-Newbold correction
results/coverage_*.csv	empirical coverage and width of the 80% conformal intervals
results/ablation_*.csv	RMSFE of each alternative data family added to hard data alone, relative to hard data alone
results/live_*	the live nowcast as of 2026-11-05 (live_nowcast.json), its evolution across origins, and the news decomposition of the latest GDP revision
geo/province_night_lights.csv	real data. For each of Thailand's 77 provinces: share of land area, share of national night-light brightness in 2012 and 2016, and the share of its 2016 pixels saturated at 255
geo/geo_summary.json	the aggregates quoted in Section 3.3 of the paper (Bangkok Metropolitan Region, Eastern Economic Corridor, top ten provinces)
SHA256SUMS	checksums of every file above

Model codes: rw random walk, ar AR(p), bridge bridge equation on principal components, midas MIDAS with exponential Almon weights, dfm mixed-frequency dynamic factor model, enet elastic net, lgbm LightGBM. Feature sets: hard_only, hard_ plus one alternative family (trends, lights, no2, prices), all, and none for the univariate benchmarks.

The synthetic data

One AR(1) common factor drives all series, with event shocks placed where Thai history had them (the global financial crisis 2008-09, the 2011 floods, the 2014 political crisis, COVID-19 2020-21, the 2022 energy-price spike). GDP growth is a trend plus 1.1 times the quarterly mean of the factor plus noise; CPI inflation loads on oil prices, farm prices, a core component and the factor lagged six months; each indicator is a loading on the factor at its own lead or lag plus AR(1) noise, rebuilt into levels with seasonality. Seed 20261105. Which alternative data help, and by how much, is therefore set by the generator's loadings, not by Thailand. Replace the files with real ones named in docs/data_sources.md and run make all to regenerate every table here.

The night-light table

Computed by scripts/make_geo_figures.py from NASA Earth Observatory's Black Marble 2012 and 2016 global grayscale composites (3 km; Román et al., 2018) and geoBoundaries ADM1 province boundaries (Runfola et al., 2020; © OpenStreetMap contributors). Brightness is summed within each province with cos(latitude) pixel-area weights. The composites are 8-bit images scaled for display, not calibrated radiance, so the shares describe where light is rather than how much: dense urban cores saturate (40.6% of Bangkok's pixels in 2016), which understates their share, and the 2012 and 2016 images may differ in processing as well as in activity. No model uses this table.

Licence

LICENSE gives the terms. The synthetic data and all results are CC BY 4.0. geo/province_night_lights.csv and geo/geo_summary.json are derived from OpenStreetMap data through geoBoundaries and are released under the Open Database License (ODbL) 1.0, with attribution to © OpenStreetMap contributors and geoBoundaries; the Black Marble imagery is NASA public domain.

Citation

@misc{chongmankhong2026thainowcast,
  author = {Chongmankhong, Thanatchon},
  title  = {ThaiNowcast: Pseudo-Real-Time Nowcasting of Thai GDP and Inflation with Official and Alternative Data},
  year   = {2026},
  doi    = {10.5281/zenodo.23995547},
  url    = {https://thanatchon36.github.io/thainowcast/paper/},
  note   = {Working paper; results are from synthetic test data}
}