Andrew Martinez on the Art of Forecasting

Episode Summary

Executive Summary: Andrew Martinez discusses his path into economics and forecasting, the evolution of macro forecasting methods, the strengths and limits of Bayesian and AI approaches, and two research projects on monetary policy surprises and nominal GDP expectations gaps. A core theme is that policymakers need interpretable models and real-time, not revised, information to make decisions and evaluate shocks accurately.

Main Topics: Martinez’s path into economics and forecasting (Priority: 4/5): Martinez explains how international experience, an interest in diplomacy, and exposure to economics during college led him toward macroeconomics, time series, and eventually forecasting as a career focus. The state of macroeconomic forecasting (Priority: 5/5): He describes forecasting’s evolution from simple models and random walks toward big data, factor models, machine learning, and methods that handle structural breaks, while emphasizing persistent trade-offs between fit, simplicity, and interpretability. Bayesian vs. frequentist thinking (Priority: 4/5): The conversation contrasts Bayesian priors and shrinkage with frequentist approaches that rely more heavily on the data, with Martinez stressing that the choice depends on context and the danger of over-relying on priors in changing environments. AI and forecasting under real-time constraints (Priority: 5/5): Martinez sees promise in AI for processing large information sets, but warns that real-time forecasting remains difficult because AI can be distorted by hindsight, data revisions, and lack of clean vintage information. Paper on SEP releases and monetary policy surprises (Priority: 5/5): Martinez and Tara Sinclair study how FOMC Summary of Economic Projections releases affect market reactions and show that SEP meetings generate larger monetary policy surprises because they reveal additional information beyond the rate decision. Expectations gap / nominal GDP gap research (Priority: 5/5): He discusses a project with Beckworth and Alex Shabola showing that forecast-based expectations gaps for nominal GDP can be more stable and less revised than traditional output-gap measures, making them potentially useful policy indicators.

Key Arguments: Forecasting in macroeconomics is as much about interpretability and economic mechanism as it is about statistical accuracy; policymakers need a story they can explain and defend. Simple models can outperform complex ones in stable periods, but turning points and crises require richer information sets and methods that can incorporate new or higher-frequency data. Machine learning and big data improve forecasting, but they do not eliminate the need for judgment, especially because policymakers must understand why a forecast changes. Bayesian methods are valuable for regularization and small-sample settings, but they can impose too much structure if priors are poorly matched to a changing world. AI can help sift through massive amounts of information, but real-time evaluation is hard because models may implicitly use hindsight, data revisions, or non-vintage information. SEP releases matter because they provide additional central-bank outlook information, not just a policy-rate decision, and that extra information helps explain larger financial-market surprises. Monetary policy shocks may be contaminated by information effects, so researchers should be careful when using high-frequency surprises as pure exogenous shocks. Forecast-based expectations gaps offer a practical alternative to traditional output-gap measures because they are less volatile, less revised, and naturally adapt with forecasters’ changing views.

Data Points: FOMC SEP release frequency: 4 times per year - SEP releases occur every other FOMC meeting, enabling comparison with non-SEP meetings. High-frequency event window: 30-minute window - Academic work on monetary policy surprises often studies financial-market reactions around FOMC decisions in a narrow 30-minute window. Bloomberg survey timing: about 10 days before the meeting; closes about 7 days before - Used to measure market expectations of rate decisions and SEP projections before the FOMC meeting. Relative size of surprises on SEP meetings: about 2 times larger - Martinez and Sinclair find absolute monetary policy surprises are significantly larger during SEP-release meetings than non-SEP meetings. Forecast horizon for expectations gap: about 5 years - Their measure aligns with forecasts over roughly a five-year window, matching some policy-institution forecast horizons. Treasury start date: fall of 2019 - Martinez began working at the Treasury Department in the macro analysis group. Podcast publication context: December 2024 meeting - Martinez notes market reaction to the FOMC meeting and SEP even when the rate decision itself was widely expected. Sample scope implied for early expectations-gap work: back to 1990 (later extended further with Blue Chip data) - Martinez references that the initial measure went back to 1990 before being extended with broader forecast datasets.

Pivotal Quotes: "the economic affairs track is almost always undersubscribed" — Andrew Martinez: Explaining what first pushed him toward economics and away from a planned diplomatic career. "forecasting, I think, is really honest about what it's doing" — Andrew Martinez: Describing why forecasting appealed to him as a discipline with direct feedback from reality. "you have to have a story" — Andrew Martinez: On why policymakers need interpretable forecasts rather than only black-box outputs.

Implications: The episode suggests future forecasting will blend AI and big data with economic judgment, but real-time transparency remains crucial. For policy researchers, SEP information effects and forecast-based expectations gaps highlight the need for careful shock measurement and robust, interpretable indicators.