Research
Research on forecasting, machine learning, and applied modeling — most of it with the Banco Central de Reserva del Perú. Working papers and the talks where the work was presented.
Prepared for XLIV Encuentro de Economistas del BCRP · Draft
Synthetic economies with known regime-switching dynamics give exact ground-truth counterfactuals. Across five models and five nonlinearity levels, simpler VAR(1) beats more flexible specifications — a curse of dimensionality in counterfactual estimation.
ReadBCRP · DT N° 012-2026 · Working paper
Whether structural metadata from an institutional registry carries enough economic signal to drive variable selection. A metadata-only Bayesian method matches state-of-the-art baselines on Peruvian CPI without ever evaluating a forecast during selection.
ReadBCRP · DT N° 025-2025 · Working paper
Kolmogorov–Arnold Networks recover the functional form of the Phillips curve, locating the nonlinearity in exchange-rate pass-through rather than the usual output-gap or unemployment channel.
ReadBCRP · DT N° 009-2024 · Working paper
An AutoML framework that searches neural architectures and preprocessing pipelines over tree-structured spaces to forecast Peruvian inflation, removing manual trial-and-error from model design.
ReadXLIV Encuentro de Economistas del BCRP · Accepted - Session TBD
What happens when counterfactual models can finally be checked against the truth? MacroGym reveals a persistent oracle gap, unexpected advantages for simple VARs, and a tradeoff between forecasting performance and counterfactual validity.
ViewXLIII Encuentro de Economistas del BCRP · Sesión metodológica
How explainable meta-learning can decide which series to feed a forecasting model, and why that choice should stay interpretable.
ViewXLII Encuentro de Economistas del BCRP · Sesión 4 · Redes Neuronales
Using Tree-structured Parzen Estimators to optimize neural forecasting architectures and hyperparameters for time series.
ViewXLI Encuentro de Economistas del BCRP · Sesión 20
A survey comparing artificial-intelligence model topologies for macroeconomic forecasting, with Juan Aquino (BCRP).
ViewAvailable for AI, software & systems consulting
AI, software, research, operations, or something that doesn't fit neatly into a category.
I typically work on
How I engage: I respond to serious inquiries within 24h. First conversation is a scoping call — no proposal until I understand the problem.