Paponpat Taveeapiradeecharoen

arXiv2025

Forecasting in small open emerging economies Evidence from Thailand

Forecasting inflation in small open economies is difficult because limited time series and strong external exposures create an imbalance between few observations and many potential predictors. We study this challenge using Thailand as a representative case, combining more than 450 domestic and international indicators. We evaluate modern Bayesian shrinkage and factor models, including Horseshoe regressions, factor-augmented autoregressions, factor-augmented VARs, dynamic factor models, and Bayesian additive regression trees. Our results show that factor models dominate at short horizons, when global shocks and exchange rate movements drive inflation, while shrinkage-based regressions perform best at longer horizons. These models not only improve point and density forecasts but also enhance tail-risk performance at the one-year horizon. Shrinkage diagnostics, on the other hand, additionally reveal that Google Trends variables, especially those related to food essential goods and housing costs, progressively rotate into predictive importance as the horizon lengthens. This underscores their role as forward-looking indicators of household inflation expectations in small open economies.

stat.AP↗

arXiv2025

Forecasting Thai inflation from univariate Bayesian regression perspective

This study investigates the forecasting performance of Bayesian shrinkage priors in predicting Thai inflation in a univariate setup, with a particular interest in comparing those more advance shrinkage prior to a likelihood dominated/noninformative prior. Our forecasting exercises are evaluated using Root Mean Squared Error (RMSE), Quantile-Weighted Continuous Ranked Probability Scores (qwCRPS), and Log Predictive Likelihood (LPL). The empirical results reveal several interesting findings: SV-augmented models consistently underperform compared to their non-SV counterparts, particularly in large predictor settings. Notably, HS, DL and LASSO in large-sized model setting without SV exhibit superior performance across multiple horizons. This indicates that a broader range of predictors captures economic dynamics more effectively than modeling time-varying volatility. Furthermore, while left-tail risks (deflationary pressures) are well-controlled by advanced priors (HS, HS+, and DL), right-tail risks (inflationary surges) remain challenging to forecast accurately. The results underscore the trade-off between model complexity and forecast accuracy, with simpler models delivering more reliable predictions in both normal and crisis periods (e.g., the COVID-19 pandemic). This study contributes to the literature by highlighting the limitations of SV models in high-dimensional environments and advocating for a balanced approach that combines advanced shrinkage techniques with broad predictor coverage. These insights are crucial for policymakers and researchers aiming to enhance the precision of inflation forecasts in emerging economies.

econ.EM↗