One of the questions we receive most frequently concerns the performance of our models. In this note, we assess the performance of our forecasting framework in 2025 for the United States, the euro area, and Japan. In addition, in the final section of the note we discuss the drivers behind this performance—which we view as particularly strong—and how the signals generated by the models can be effectively interpreted and used (including new distributions of probabilities).
United States
Figure 1 shows the evolution of our “main” model’s forecast for 2025 core PCE inflation (Q4/Q4) alongside the realized year-end inflation outcome. As the chart illustrates, 2025 was a particularly volatile and challenging year. This is hardly surprising given the sequence of announcements by the U.S. administration.
At the beginning of the year, the model projected core inflation close to 3 percent. Following “Liberation Day,” the forecast rose sharply, reaching nearly 3.9 percent. As new announcements were subsequently made, the forecast declined, and by June it had already stabilized at a level very close to the final outcome—aside from revisions in November and December related to the government shutdown.
We view this evidence as a clear success of the model. Once the information set stabilized and broadly reflected the administration’s final policy choices, the model was able to accurately project annual core inflation.
Fgure 1. “Main” model 2025 forecast evolution and realized inflation in the US.
Note: The Figure shows the evoluation of our “main” model forecast for core PCE (Q4/Q4). The dashed line shows the realized end-of-year inflation.
Euro Area
Figure 2 shows the evolution of the forecast for core HICP inflation in the euro area (average year-on-year). The evidence here is particularly strong: the model correctly anticipated the 2025 outcome as early as mid-2024 and has barely revised that forecast since, by at most one tenth of a percentage point. It is worth recalling that, in 2024, the model was entirely out of consensus.
Figure 2. “Main” model 2025 forecast evolution and realized inflation in the EA.
Note: The Figure shows the evoluation of our “main” model forecast for core HICP (average YoY in 2025). The dashed line shows the realized end-of-year YoY average.
Japan
Figure 3 presents the same exercise for Japan, distinguishing between core-core inflation (red lines) and core inflation in a Western-style definition (grey lines). While FY2025 has not yet formally concluded, the final outcome is by now largely clear. Here again, our models were out of consensus at the beginning of 2025 and significantly above the Bank of Japan’s forecasts. Despite this, revisions over the year were limited, and by mid-year the models were already pointing to the correct final outcome.
Figure 3. “Main” model FY2025 forecast evolution and realized inflation in Japan.
Note: The Figure shows the evoluation of our “main” model forecast for core CPI (average YoY in FY2025).
The drivers of the models performance and how to trade them
The performance of our models reflects a combination of econometric specification and accumulated experience. In particular, one key lesson we have learned over time is to react far less aggressively to news than market participants typically do. In the vast majority of cases, the impact of shocks on firms’ marginal costs is relatively modest—the natural gas shock of 2022 being a notable exception.
In our experience, there are two primary ways to use our medium-term models, which we consider one of our core assets. First, they can be used to assess the risks surrounding central bank staff forecasts and to anticipate future revisions. Second, they are particularly valuable when market consensus reacts strongly to incoming news and deviates significantly from model-based signals. Both approaches have delivered solid results in recent years, including in 2025.
(Finally, we would like to note that we have recently put into production a new methodology for computing confidence intervals for our models. This new approach provides, in real time, the full (non-symmetric) probability distribution around the forecast. Further details will follow in a forthcoming note.)