Forecasting

Project Prism -- Probabilistic Regional Inference for Sub-national Macroeconomics

This page demonstrates a national macroeconomic forecasting model that is part of a larger agenda to build an accurate, timely, sub-national forecasting framework for local governments and non-profit organizations. State and metro-area policymakers routinely make consequential decisions with little more than lagged national data and back-of-the-envelope extrapolations. The alternative is expensive price sector solutions. The goal of this project is to change that by delivering probabilistic, scenario-based forecasts at the geographic resolution where decisions are actually made.

The national model documented here is the first step. It demonstrates the core forecasting methodology: A large-scale Bayesian VAR, disciplined shrinkage priors, rigorous out-of-sample evaluation, and scenario analysis at the level where data is richest and benchmarks are most established. Achieving forecast accuracy competitive with the Survey of Professional Forecasters at the national level provides a credible foundation before extending the framework to state and local economies, where data is sparser and the modeling challenges are more difficult. In short, this is a proof of concept: evidence that the statistical machinery works before deploying it where it is needed most. Forecasts and scenarios updated for the macroeconomy on a quarterly basis. Anyone looking to learn more should feel free to reach out and I will share the underlying forecasts with you.

Overview

This model produces probabilistic forecasts for the U.S. economy using a 50-variable Bayesian Vector Autoregression (BVAR) estimated on quarterly data from 1960 to the present. The methodology follows Crump, Eusepi, Giannone, Qian & Sbordone (2025) with a few personal modifications.

• 50 quarterly variables modeled jointly — GDP, employment, inflation, interest rates, credit, housing, industrial production, and financial conditions all interact in a single coherent system. 

•  Full probability distributions.

• AI-based macroeconomic scenarios for conditional forecasting.

Current Scenarios

Baseline. The model’s unconditional forecast with no external assumptions. Represents the most likely path given current data and historical relationships.

Goldilocks Soft Landing. The Fed threads the needle. Inflation gradually returns to target as the labor market cools just enough—unemployment drifts to 4.5%—without triggering a downturn. The Fed eases 150bp over three years in measured 25bp steps. Growth stays positive but below trend as the economy rebalances. This is the consensus “best case” scenario.

Consumer-Led Recession. Accumulated pressure from high rates and depleted excess savings triggers a consumer pullback. Spending contracts, layoffs accelerate, and unemployment surges to 6.5% within 18 months. The Fed responds with aggressive easing, cutting to near-zero by mid-2027. A classic demand-driven downturn reminiscent of 2001.

1970s-Style Stagflation. A second wave of inflation driven by supply-side disruptions—energy shocks, reshoring costs, and persistent services inflation—pushes core CPI back above 5% annualized. Simultaneously, growth stalls and unemployment rises to 5.5%. The Fed is paralyzed, holding rates steady as both mandates deteriorate. The worst of both worlds.

Credit Crunch. A sudden repricing of credit risk—triggered by commercial real estate losses, a major bank stress event, or emerging market contagion—sends Baa spreads 200bp wider. Lending freezes. The Fed pivots to emergency cuts and liquidity facilities. Echoes of March 2020 or the 2008 credit crisis, but without the housing collapse.

AI Productivity Boom. Generative AI and automation drive a sustained productivity surge. Firms hire aggressively, unemployment falls to 3.2%, but inflation stays contained because output per worker is rising. The Fed holds rates steady—strong growth doesn’t require easing, but low inflation doesn’t require tightening. A new economy narrative takes hold.

Term Premium Shock. A sudden 100bp rise in the 10-year Treasury yield driven by fiscal concerns—ballooning deficits, a failed Treasury auction, or a sovereign credit downgrade. Not a Fed action but a market repricing of duration risk. Mortgage rates spike, housing freezes, and equity valuations compress.

Global Supply Shock. A major geopolitical disruption—escalation in the Middle East, a Taiwan Strait crisis, or coordinated OPEC+ cuts—sends commodity prices surging. Metals jump 20%, energy and intermediate materials follow. Input costs spike, squeezing margins and passing through to consumer prices. A supply-side inflation shock that monetary policy can’t easily address.

Deep Yield Curve Inversion. The yield curve inverts by 50bp as markets price in an imminent recession while the Fed holds short rates elevated. Historically, sustained inversions of this magnitude have preceded every recession since 1970 with a 12–18 month lead.

Mission Accomplished Disinflation. The Fed’s tightening campaign succeeds. Core PCE decelerates steadily back to the 2% target over the next two years as shelter inflation normalizes and goods deflation continues. With the inflation mandate satisfied, the Fed begins a methodical easing cycle starting in mid-2026, cutting 25bp per quarter. The soft landing is confirmed.

Scenario Probabilities

Each scenario is assigned a model-derived probability that measures how likely that path is relative to the others, given current economic conditions. The BVAR produces a full joint probability distribution over all 50 variables at every forecast horizon. Each scenario defines a specific path through that distribution. We measure the statistical distance (Mahalanobis distance) between each scenario’s path and the model’s unconditional forecast, using the model’s own estimate of forecast uncertainty and cross-variable correlations. Scenarios closer to the baseline receive higher probability; extreme scenarios that require large, unusual shocks receive lower probability.