Still fit for purpose? Upgrading our economic models for an age of uncertainty

10 July 2026

By Matteo Ciccarelli and Antoine Kornprobst

Galileo famously wrote that the book of nature is written in the language of mathematics. And modern physics has gone on to describe the motion of planets and particles with ever greater precision. The economy, by contrast, is made up of people. People who make their own decisions – “particles” that feel, think and anticipate, and that may not always act entirely rationally. Economists nevertheless try to find patterns in the ways households and firms make decisions and interact, and in how this all adds up to what we call “the economy”. This comes with its own very specific empirical challenges, which also matter for central banks and monetary policy.

Our models were showing their limitations

In macroeconomics in particular, large-scale experiments are generally not an option: you cannot run a recession twice with two different sets of interest rates. Instead, economists build macroeconomic models in which equations stand in for relationships in the economy. These models are like small laboratories in which a simulated economy can be poked, prodded and shocked, and forecasts can be confronted with the data. Which is why models are crucial in central banking: economists at the ECB maintain and use macroeconomic models to prepare forecasts, scenario analyses and policy prescriptions. But, with large shocks becoming increasingly frequent and uncertainty more prevalent, our models had begun to show their limitations. The time had come to review our modelling toolkit and practices.

Economic forecasts can fail for all sorts of reasons. The culprit can usually be found hiding in the technical assumptions. For instance, some of the basic conditions we assume (like the price of energy) do not in fact remain constant. Other errors can arise from the way we estimate the relationships within an economy, i.e. how exactly a change in one particular variable propagates through a model, affecting other variables. The first type is the easiest to identify and the least interesting to dwell on. Unexpected events – the pandemic, the war in Ukraine – mean that assumptions made before a shock can quickly become outdated. No model, no matter how carefully specified, can be expected to forecast accurately following a structural break (i.e. a fundamental change) no one saw coming.

When old tools fail to capture new shocks

But mistaken assumptions can only go so far in explaining a model’s past shortcomings. When we reran our projections, this time with what really happened in terms of commodity prices, interest rates, exchange rates and other conditioning assumptions, our models proved to be valid. However, they still underestimated the inflation surge that followed.^[1] In other words, the shortfall was not only a question of bad luck with assumptions. Something was also missing in the transmission. Two things, in fact.

First, linear models struggle to capture how commodity price shocks as large as those we have seen feed into headline inflation. They fail because the relationship is non-linear: the larger the initial shock, the stronger the pass-through of higher energy prices to inflation. The effect is also bigger if rising energy prices hit the economy when inflation is already high. Neither of these factors can be easily accommodated by a linearised model.

Second, the indirect and second-round effects were stronger than the models implied. Firms passed higher input costs through to consumer prices more aggressively than historical patterns might have suggested. Meanwhile, wages caught up with higher prices unusually rapidly.^[2] As a result, the equations of our models were no longer able to capture the key parameters of the post-pandemic economy because they were estimated largely on the basis of data drawn from a low-inflation environment.

In times of war standard correlations can break down

We are learning from these lessons. The ECB is developing and deploying models that are more detailed and better equipped to deal with unusual economic fluctuations. Because energy and food are priced on global markets, geopolitical shocks transmit rapidly into higher consumer prices. In times of war, the standard economic correlations and transmission mechanisms embedded in models can break down entirely. As was the case, for example, when oil, gas and electricity prices decoupled unexpectedly following Russia’s invasion of Ukraine.

Add to this more frequent and forceful fiscal and regulatory interventions in energy markets, and the modelling of aggregate energy prices hits its limits. With this in mind, our novel suite of models now follows a more disaggregated approach. The models provide a more granular, bottom-up assessment of the risks surrounding short-term inflation projections. The new models split pre-tax energy prices into subcomponents: fuel, gas and electricity. They also incorporate modelling techniques to deal with outliers and extreme volatility and boost their explanatory power and forecasting accuracy.^[3] This granular modelling allows us to better gauge the effects of structural changes in energy markets, extreme weather events and climate policies, such as the forthcoming EU Emissions Trading System 2 (ETS 2).

We have also introduced new data sources as inputs for our models to help avoid blind spots. Thus, our models can now estimate and disentangle a greater variety of macroeconomic shocks. During the pandemic, for example, global supply chain bottlenecks led to severely disrupted trade in intermediate goods, longer delivery times and shortages of key production inputs. We have therefore incorporated supply constraint indicators into our satellite models, i.e. models that focus exclusively – and in greater depth – on specific sectors or relationships in the economy. For instance, our large core models do not usually factor in supply chain frictions or quantity rationing. However, more focused satellite models can offer a complementary view when such mechanisms become a first-order concern, as was the case after 2021.^[4] The same framework enables us to assess the inflation risks associated with supply constraints from new shocks. Examples include the tariff volatility following the blanket imposition of tariffs on US imports in 2025 (on so-called “Liberation Day”) or the Gulf energy export disruptions amid the closure of the Strait of Hormuz in 2026. We also complement and cross-check projection models with higher-frequency indicators, drawing on surveys, market-based measures, news-based attention indices, Google search activity and mobility data. Aggregated into harmonised weekly activity indices, these sources provide timely signals on near-term economic developments.

Focus on core projection models

We have also worked to make our core projection models more accurate, particularly over medium-term horizons, and better able to capture what drives key macroeconomic variables. We have done this by revisiting some of these models’ behavioural equations. We re-estimated parameters using a longer sample of historical data and looked at how shocks propagate compared with empirical estimates from the literature.^[5] Here, based on very recent evidence, two channels have garnered most of the attention: monetary policy transmission and the pass-through of commodity price shocks. We have also rethought how prices are modelled, so that foreign shocks are reflected in our models in a way that resembles what is actually happening in the economy. In particular, the re-estimated version of ECB-BASE aims to better capture the pass-through of energy price shocks to core and food prices.^[6] Likewise, the New Area-Wide-Model (NAWM) II now includes a direct oil price propagation channel to separate the pass-through of commodity price shocks from other foreign price shocks.^[7] Recent evidence shows that firms change prices more often when inflation is high. Rather than absorbing higher input prices, as they did in calmer periods, firms actually tend to react to shocks when inflation is already high.^[8] The Phillips curve, in other words, is not a fixed object.

To capture this, the ECB has developed complementary models using state-dependent pricing in which the way such shocks are transmitted is allowed to change over time and across different states of the world. These can then be used to cross-check the core projections. Also included are new methods for producing density forecasts of inflation, in particular machine learning approaches that do not impose a specific predetermined relationship between inflation and its drivers.^[9] Unlike models grounded in economic theory, these machine learning methods are more data driven and allow for flexible, non-linear relationships between inflation and its determinants. They are particularly useful for gauging the risks around the baseline, which, after recent years, is exactly where the ECB wants to focus its attention.

Beside foreign shocks, public policy interventions can also affect the euro area economy business cycle. The fiscal and monetary policy responses to the pandemic and the subsequent energy crisis were particularly significant in the region. During the pandemic, government spending and transfers helped cushion the impact on aggregate demand, as did the ECB’s long-term refinancing operations and quantitative easing programmes. As inflation surged in 2022, monetary policy tightened, and specific government measures shielded households and firms from higher energy prices. Thanks to the new models, the macroeconomic implications and distributional impact of these complex interactions between monetary and fiscal policy can now be assessed. They also allow for integrated, structural policy assessment in an open-economy heterogeneous monetary union.^[10][11]

Wealth and income matter for how households react to shocks

This combination of shocks and an active policy mix has also highlighted how much it matters that households and firms are not all alike.^[12] Models with heterogenous agents allow for a richer analysis of inequality and distributional considerations than models based on a single representative household. For instance, they make it easier to analyse how differences across balance sheets and marginal propensities to consume can shape aggregate dynamics.^[13] The ECB can use these models to study how monetary policy affects households differently depending on income, wealth and asset composition. These models can also be used to analyse how monetary and fiscal policies can be jointly designed, e.g. in response to rising energy prices. Lastly, they can help capture the way in which different households and firms adjust their consumption and investment, and what this might mean for goods, asset and labour markets, in equilibrium.

As the macroeconomic environment grows increasingly complex and uncertain, the ECB’s modelling toolkit is being reshaped. Earlier frameworks treated the economy in aggregate terms. The current suite adds granularity and heterogeneity. Structural, sectoral and heterogeneous agent models are now used side by side, so that the same shock can be viewed through several lenses. Thus, not only can policies be assessed by their aggregate effects, it is now also possible to identify who actually bears their ultimate consequences. The pay-off comes from the interplay: each individual model has its blind spots, but together, through a process of triangulation, a clearer picture emerges. The latest statistical and computational techniques lie at the heart of all this, ensuring that our framework remains at the cutting edge of methodological progress. The result is a richer reading of the issues that dominate today's policy debates: energy price volatility, supply chain disruptions, monetary policy transmission and the macroeconomics of the climate transition.

To conclude

All of these investments come together in the projections and scenario analyses that inform policy decisions. The lessons of the ECB’s 2025 strategy assessment have been put into practice: scenario analysis is no longer a linear extrapolation around the baseline, but rather a structured exploration of how large shocks actually propagate, including the non-linearities and second-round effects that become most consequential when price stability is at risk.^[14] The March 2026 projections and scenarios show what this looks like in practice. Faced with the prospect of escalation in the Middle East conflict, the exercise went beyond simply scaling up a baseline pass-through coefficient. Instead, the core model elasticities were adjusted in line with the severity of the scenario. That benefits from empirical evidence that pass-through is state-dependent and stronger when shocks are large, and when inflation is already high and the economy is running hot (Chart 1).

Chart 1

Linear models may underestimate the pass-through of commodity price shocks

The response of core prices to a given shock to energy and food prices is conditional on the size and sign of the shock, the level of current inflation and the cyclical position of the economy.

However, some analytical gaps still persist. Current initiatives include the development of models to evaluate the effects of structural long-term issues in combination with macro-financial linkages. This will enable analyses of the macroeconomic effects of such key issues as defence spending, competitiveness policies, geopolitical developments, artificial intelligence, demographic change and long-term fiscal challenges. We will continue to refine our modelling techniques and their application, because models can help lay the groundwork for the preparation of monetary policy decision making.

The views expressed in each blog entry are those of the author(s) and do not necessarily represent the views of the European Central Bank and the Eurosystem.

Check out The ECB Blog and subscribe for future posts.

For topics relating to banking supervision, why not have a look at The Supervision Blog?