Key Details
European Central Bank (ECB) paper, Deep-π: A Neural Network Estimate of Trend Inflation, studies US consumer-price inflation. It tests both what the model can explain with hindsight and what it could have forecast using data available during the 2021–23 surge.
Finding | What the Paper Shows |
|---|---|
Two measures of trend inflation | A current trend tracks persistent inflation today; a forward-looking trend is designed to predict inflation over the next year |
Early signal | In early 2021, the forward-looking model forecast about 3.8–3.9% inflation over the following year, versus 2.0% from professional forecasters; actual inflation was 7.3% |
Forecast accuracy during the rise | Across six quarterly forecasts from early 2021 to mid-2022, the model’s forecast-error measure was 2.15–2.25 percentage points, versus 4.86 for professional forecasters |
Important reversal | During the subsequent decline, professional forecasters performed substantially better than the model |
Two Inflation Measures Answer Different Questions
Conventional statistical models can estimate an underlying inflation trend but often cannot explain why it moved. The paper’s Deep-π* model uses 122 US indicators—including employment, economic activity, money and credit, interest rates and prices—to link movements in the trend to observed economic conditions.
Its two measures serve different purposes. The current trend estimates the persistent component of inflation at a given time. The forward-looking trend looks for the underlying rate that best predicts inflation over the following twelve months. In estimates using the full dataset, the forward-looking trend peaked eight to twelve months before the current trend after the pandemic. That hindsight result should be kept distinct from the paper’s separate forecasting exercise.
Labour and Monetary Indicators Carried the Early Signal
The model attributes 41–52% of the post-pandemic rise in the current trend to employment-related indicators. The two versions of its forward-looking measure disagree more: one gives greater weight to real activity and employment, while the other gives greater weight to money and credit. This difference is model uncertainty, not proof that one set of indicators caused the inflation surge.
Both versions largely looked past movements in energy and commodity prices when estimating the persistent trend. Their advantage was most visible while inflation was rising. When inflation began to fall, however, professional forecasters adjusted faster. Across the full 2021–23 evaluation, the model beat two standard statistical benchmarks but performed broadly comparably to a conventional autoregressive model. The exercise also uses a later vintage of revised economic data, so it is quasi-real-time, not a test using every original data release exactly as first published.
Policy Relevance
For the RBI, the paper offers a method to test. Comparing a measure of persistent inflation now with one designed to forecast the next year could help identify when price pressures are becoming more durable. Showing which indicators accompany a change may also make forecasts easier to scrutinise.
An Indian application would need to be estimated and evaluated on Indian data, against existing RBI forecasting approaches and across more than one inflation episode. The US result itself cautions against relying on a single model: an early warning during a surge was followed by a slower reading of the decline.
Follow the Full Working Paper Here: Deep-π*: A Neural Network Estimate of Trend Inflation