Neural model detected US inflation pressure before forecasters
Deep-π* forecast US inflation of 3.8–3.9 percent over the following year in early 2021, versus professional forecasters' 2.0 percent; the outcome was 7.3 percent. Carlos Montes-Galdón's model links persistent inflation to observable economic conditions.
Two trends reveal different drivers
Carlos Montes-Galdón's European Central Bank (ECB) Working Paper 3307 estimates Deep-π* using 122 US indicators from February 1970 to December 2025.
Neural networks produce a current inflation trend and a forward-looking trend optimized for twelve-month forecasts.
Shapley values attribute movements to nine indicator groups, not causal shocks.
Employment explains 41–52 percent of the current trend's rise from January 2019 to its post-pandemic peak.
Real activity, employment and money and credit drive the forward-looking rise, with weights varying by network design.
With time-varying volatility, trends peaked at 6.7–7.5 percent after the pandemic; December 2025 estimates were 2.7 percent for the current trend and 2.2–2.4 percent forward-looking.
Earlier warning, slower retreat
Monthly re-estimation from December 2020 to November 2023 yielded twelve-month forecast RMSEs 15–18 percent below the unobserved-components model with stochastic volatility and 24–26 percent below a random walk.
Point forecasts were about as accurate as a twelve-lag autoregression; forward-looking density forecasts performed best.
Professional forecasters anticipated disinflation better: their RMSE was 0.33 percentage points over 2022Q3–2023Q4, versus 2.13–2.25 for the forward-looking model.
The paper does not represent ECB views.
Hindsight is not foresight
The eight-to-twelve-month lead of the forward-looking peak belongs to full-sample estimates, not real-time results.
The evaluation uses September 2026 data and excludes historical data revisions.
Different driver weights across similarly accurate networks leave attribution subject to model uncertainty.