A Hybrid Early-Warning System for Inflation in an Emerging Market: Combining Econometric Models, an Agent-Based Decomposition with Heterogeneous Expectations, a Large Language Model, and a Multi-Output Agent Architecture

Labastidas, Esteban (2026): A Hybrid Early-Warning System for Inflation in an Emerging Market: Combining Econometric Models, an Agent-Based Decomposition with Heterogeneous Expectations, a Large Language Model, and a Multi-Output Agent Architecture. Forthcoming in:

Abstract

We develop and evaluate a hybrid early-warning system for year-over-year (YoY) inflation in Colombia that combines four econometric models (ARIMA, LASSO, ElasticNet, and a weighted ensemble), a reduced-form VAR, an agent-based model with heterogeneous expectations and tradable/non-tradable pass-through (ABM v2), a large language model (LLM) forecaster, and a multi-output agent architecture, integrated through Dynamic Model Averaging (DMA). We evaluate the system on a rolling out-of-sample backtest from February 2010 to March 2026 (n ≈ 194 months) spanning the 2021–2023 inflation surge and its ongoing disinflation. Five contributions emerge. First, an identity-based monthly-to-YoY decomposition applied uniformly across reduced-form models reduces MAE by 15–20% relative to direct YoY forecasting without adding variables. Second, regime-conditional analysis shows that MAE is 2.0–3.0 times larger in surge regimes than in stable regimes across all models. Third, an ABM with regime-dependent heterogeneous expectations reduces full-sample MAE from 0.337 pp (v1) to 0.268 pp (v2, −20.4%) and surge-regime MAE from 0.585 pp to 0.404 pp (−31.0%), with Diebold-Mariano statistic 6.21 (p < 0.001). The accompanying four-channel shock decomposition attributes the 2021–2023 surge primarily to amplification of wage (1.96×) and expectations (1.82×) channels rather than direct import pass-through (1.36×). Fourth, a structured-reasoning LLM forecaster is validated via live-API replication and a 60-call fake-date test; the non-monotonic MAE pattern is shown to reflect target-volatility confounding rather than look-ahead bias. Fifth, a multi-output agent architecture predicts BanRep monetary policy decisions with 77.3% direction accuracy (17/22) and 63.6% exact-magnitude match (14/22) on 2022–2026 meetings, with all prediction errors concentrated at cycle inflection points. The full DMA ensemble achieves MAE 0.26 pp and 94.8% coverage at horizon h = 1.

• Item Type: MPRA Paper
• Language: English
• Keywords: Inflation forecasting, Emerging markets, Agent-based models, Heterogeneous expectations, Large language models, Dynamic model averaging, Diebold-Mariano test, Colombia
• Subjects: C - Mathematical and Quantitative Methods > C4 - Econometric and Statistical Methods: Special Topics > C45 - Neural Networks and Related Topics; C - Mathematical and Quantitative Methods > C5 - Econometric Modeling > C53 - Forecasting and Prediction Methods ; Simulation Methods; E - Macroeconomics and Monetary Economics > E3 - Prices, Business Fluctuations, and Cycles > E31 - Price Level ; Inflation ; Deflation; E - Macroeconomics and Monetary Economics > E3 - Prices, Business Fluctuations, and Cycles > E37 - Forecasting and Simulation: Models and Applications

References

Di Domenico, G., Catalano, M., & Riccetti, L. (2025). Economic Modelling, 147. Zárate-Solano, H., & Manrique-Rodriguez, C. (2025). Borradores de Economía, Banco de la República. Harvey, D., Leybourne, S., & Newbold, P. (1997). International Journal of Forecasting, 13(2).