Inflation Unpacked: Breaking Down the Key Components Using a Neural Phillips Curve

Abstract

Background and Objectives: Effective monetary policymaking requires a clear understanding of the underlying drivers of inflation. For central banks such as the Bangko Sentral ng Pilipinas (BSP), this task is challenging because inflation reflects the interaction of demand-side pressures, supply-side shocks, and expectations. Key components of inflation dynamics, including the output gap and inflation expectations, are unobserved and are often difficult to measure accurately. Traditional Phillips curve models typically rely on strong assumptions, filtering methods, or survey-based proxies, which may be sensitive to misspecification, measurement error, and nonlinearities. This study applies a deep learning approach to examine inflation dynamics in the Philippines and disentangle the contributions of real activity, inflation expectations, and commodity prices within a New Keynesian Phillips Curve framework.

Methodology: The study employs a Hemisphere Neural Network (HNN), a deep neural network architecture designed to improve interpretability by grouping input variables into separate “hemispheres.” Each hemisphere corresponds to a key component of inflation: real activity, inflation expectations, and commodity prices. This structure allows the model to extract latent indicators that can be interpreted as macroeconomic states and to decompose realized inflation into component-specific contributions. Unlike standard neural networks, which often operate as black boxes, the HNN imposes an economically meaningful structure on the final layer, making the estimated components more transparent. The model is estimated using Philippine quarterly macroeconomic data and is assessed based on its ability to generate meaningful inflation forecasts, align with historical macroeconomic events, and produce interpretable indicators of the output gap and inflation expectations.

Key Findings: The results show that Philippine inflation can be decomposed into distinct contributions from real activity, inflation expectations, and commodity prices. Long-run inflation expectations remained relatively stable at around 3.5–4.5 percent over the sample period, broadly consistent with the BSP’s inflation target range of 2–4 percent. This stability suggests that expectations have remained relatively well anchored. Commodity prices account for much of the short-term volatility in inflation, particularly during episodes such as the 2014–2015 oil supply glut and the 2021–2022 global supply chain disruptions. By contrast, real activity appears to play a more important role in medium-term inflation movements, especially during the pandemic and post-pandemic recovery periods. The results also indicate that the relationship between inflation and economic activity in the Philippines remains relevant, contrary to claims in some strands of the literature that the Phillips curve has weakened or disappeared. The HNN-derived real activity gap broadly reflects domestic economic conditions and can be interpreted as an additional measure of inflationary pressure. Similarly, the model-based measure of inflation expectations aligns with short- to medium-term expectations from businesses and professional forecasters, suggesting that it can serve as a useful supplementary indicator when survey-based measures are limited.

Policy Implications: The findings highlight the usefulness of interpretable deep learning methods for monetary policy analysis. By distinguishing between inflation driven by demand-side pressures and inflation arising from temporary supply shocks, the HNN can help policymakers assess whether inflation movements warrant monetary tightening, easing, or a more cautious policy response. Inflation driven by real activity or expectations may call for a stronger monetary policy response, while inflation caused mainly by temporary commodity price shocks may justify a more measured approach. The model’s estimates of the output gap and inflation expectations can also complement existing indicators used by the BSP to assess domestic demand conditions, expectation formation, and the credibility of the inflation target. Overall, the study demonstrates that theory-guided machine learning can strengthen macroeconomic monitoring and support more targeted, evidence-based monetary policy decisions.