"A model is only as good as the model engineer"

Since this autumn the KOF Nowcasting Lab has been providing daily updated forecasts of growth in real gross domestic product (GDP) in Switzerland, the euro area and selected EU member states (Belgium, Germany, France, Italy, the Netherlands, Austria, Spain, Poland and Sweden) as well as the United Kingdom. KOF economic expert Heiner Mikosch explains how the KOF Nowcasting Lab works and talks about potential extensions to this platform.

The KOF Nowcasting Lab does not incorporate any expertise and the model learns independently. Does this mean that economists will become superfluous in the future and that we will only need clever programmers?

No, not at all. Econometric models are made by humans. All models are a simplification of reality to reduce complexity. Models’ approach cannot reflect the full complexity of reality. A model is only as good as the model engineer behind it.

How does collaboration between the model and the model engineer work ideally?

We modellers give the model a certain reality, and that is a necessary slice of reality. What doesn’t work is to feed a model with all of the data that exists in the world. The model relies on what we give it. The model then tries to make connections – in the case of the KOF Nowcasting Lab between certain leading indicators such as business surveys, interest rates, monthly industrial output and gross domestic product (GDP). The aim of the model is to be able to forecast GDP as accurately as possible.

What part of reality does the KOF Nowcasting Lab ignore?

We need to provide a framework for the model. Models such as those used in the KOF Nowcasting Lab are generally based on linear relationships. This means that we hide non-linear correlations. Every model represents parts of reality. Other models can capture non-linear relationships but have other weaknesses. This is why model pluralism, in which several models compete with and complement each other, is important.

The terms 'machine learning' and 'Big Data' are on everyone’s lips these days. Does this mean that researchers only have to feed their models with as much data as possible in order to arrive at the most accurate forecasts possible?

No. Machine learning is good at calculating relationships in a sample from a mass of past data, but it has its limitations. Finding a mathematical explanation of the past is only the first step from a forecaster’s perspective. Accurately forecasting the future is our ultimate objective. In doing so, we go beyond the available data and move on to a new sample. A key insight here is that models that are supposed to work have to be simple. The more complex we build the models, the worse they often perform when forecasting. It’s not just about Big Data, it’s about smart data. To explain it technically in a nutshell: the determination measure R² indicates the amount of variation in the data that the model can explain. Beyond a certain point, R² often correlates negatively with predictive ability. A certain minimum value of R² is needed for good forecasting. But a model with a relatively high R² usually produces poor forecasts. Complexity must not be allowed to introduce false signals.

What is the key difference between the KOF Nowcasting Lab and KOF’s quarterly economic forecast?

There is a fundamental difference between a quarterly economic forecast – such as the one we produce at KOF – and the KOF Nowcasting Lab. KOF’s economic forecast incorporates both mathematical model calculations and expert assessments. Model results – relating, for example, to the labour market, consumption or investment – are critically scrutinised and readjusted by our experts. The economic expertise in a forecast is at least as important as the pure mathematical calculations behind our models. We would never accept an unrealistic result from a model. This is not the case with the KOF Nowcasting Lab. It is quite possible for it to produce an unrealistic or even completely crazy result without us taking corrective action. That’s why it’s a lab – i.e. a laboratory of an experimental nature – which sometimes gets it wrong but occasionally delivers astonishingly good results.

What does the future hold for the KOF Nowcasting Lab? In what directions could it evolve?

The KOF Nowcasting Lab is not something ready-made; we have intentionally designed it to be technically capable of incorporating various extensions. The more alternative models that run simultaneously on this platform, the livelier the Lab is. That’s why we would be happy to conduct research collaborations with other institutes.