Figures and method

Data and topics

The analysis is based on speeches by European central bankers only. First, I use a topic model (LDA) to examine which topics each central bank discusses in its speeches and how this changes over time. The topic figures show topic shares over time, a topic-by-month heatmap, small multiples for individual institutions, overall topic prevalence and the top words that define each topic.

Demand and supply uncertainty

Second, I analyse the text in more depth. Using a trained machine-learning model, I identify passages that express uncertainty about supply (for example, oil supply) and uncertainty about demand. Aggregating these passages over time gives two indices: one for demand uncertainty and one for supply uncertainty.

Outlook and disagreement indices

Third, I build indices of the economic outlook and of disagreement among policymakers in the euro area. The approach is inspired by Ahrens and McMahon (2021), who turn central bank speeches into quantitative signals about the economic outlook. Here I focus on two variables:

• GDP growth
• inflation

The challenge is that a speech never comes with an explicit statement such as “my GDP forecast has increased by 0.4 percentage points”. Training such a model therefore requires a dataset that contains both text and an actual numerical forecast. For this, I use the ECB Economic Bulletin, which combines a written assessment of the economy with the ECB's macroeconomic projections.

Suppose a bulletin says something like: “demand has weakened, labour market conditions have deteriorated, and business investment is expected to remain subdued”, and the GDP forecast in the same bulletin is revised down. Across many bulletins, the model learns that language of this type tends to go together with downward revisions to the GDP outlook.

The dependent variable is the revision to the forecast for inflation (or GDP growth) over the coming year at ECB meeting \(m\). The model is

\[ \text{Forecast revision}_m = f\big(\text{current macro conditions}_m,\; \text{previous forecast}_m,\; \text{text}_m\big) + \varepsilon_m . \]

It controls for current GDP growth, inflation and unemployment, and for the forecasts published in the previous Economic Bulletin. The text then adds information about how the ECB interprets those conditions. Once trained, the model is applied to individual speeches to estimate the outlook revision implied by each speaker's language.

The indices

Rather than a generic hawkish/dovish score, the dashboard shows three economically defined indices:

• Euro area inflation outlook index: whether the language in speeches implies higher or lower future inflation.
• Euro area growth outlook index: whether speeches imply a more optimistic or more pessimistic outlook for the real economy.
• Euro area policymaker disagreement index: the dispersion of the outlooks implied by individual policymakers' speeches.

Interpretation

The indices are standardised, which gives them a natural interpretation. For a speech \(s\),

\[ I^{\pi}_{s} = +1 \]

means that the speech uses language which has historically corresponded to a one-standard-deviation upward revision of the ECB's inflation outlook. This is much easier to interpret economically than a standard hawkish/dovish sentiment score. The growth outlook index \(I^{y}_{s}\) is interpreted in the same way, in terms of revisions to the ECB's real GDP growth outlook.

Reference: Ahrens, M. and McMahon, M. (2021), Extracting economic signals from central bank speeches, Proceedings of the Third Workshop on Economics and Natural Language Processing.