ECB communication and money market rates

Aggregate tone of ECB press conferences from 1998 to 2016, hawkish periods shaded red and dovish periods blue, plotted with the MRO rate and EONIA.

A replication of Czudaj and Nguyen (2025). I score the tone of every ECB press conference with a hawkish/dovish dictionary, then test whether that tone predicts the overnight interbank rate beyond the policy rate itself.

Key result: 1.15 pass-through of the policy rate to EONIA, against about 1.14 in the paper
Context: Personal project, replication of a paper published in Macroeconomic Dynamics

Question

Central banks do not only move rates, they also talk. Czudaj and Nguyen (2025) build sentiment indices from the ECB’s press conferences and show that their tone predicts the interbank rate beyond the information in the policy rate decision. I reimplemented their method from the paper to check whether the result holds.

Method

A score for each press conference. Each transcript is split into sentences. Around every occurrence of inflation, interest rate(s) and unemployment, I keep the 10 words before and after and remove stop words. These words are matched against the paper’s dictionary of hawkish and dovish terms, including truncated patterns such as ris* or increas*. A window scores its hawkish matches minus its dovish matches, divided by its length, and the window scores are summed into three topic indices and an aggregate one.

Predictive regressions. For each index X and each lag k from 0 to 12 press conferences, I estimate

EONIA_t = β0 + β1 · X_{t−k} + β2 · MRO_t + ε_t

by OLS with HAC standard errors (4 lags), where EONIA is taken on the day of the press conference and MRO is the ECB’s main refinancing rate. β1 measures what the past tone adds to the current policy rate.

Validation

Check	Replication	Paper
Example sentence of the paper, inflation score	−0.333	−0.33
Inflation index, mean and std	0.13, 0.54	0.06, 0.51
Aggregate index, mean and std	0.00, 0.73	−0.16, 0.70
Correlation with the paper’s Figure 2, 1998-2016	0.69
Pass-through coefficient β2	1.15	about 1.14

The transcript corpus I used stops in December 2016, so the sample has 205 press conferences instead of the paper’s 266.

Results

Sentiment coefficient by lag, with 95% HAC confidence intervals, for the inflation, interest rate, unemployment and aggregate indices

Inflation and interest rate sentiment have a negative and significant coefficient from a lag of 4 and 1 press conferences, and it grows with the lag.
Aggregate sentiment is negative and significant from lag 1.
Unemployment sentiment is mostly not significant.

These signs are the paper’s. Its reading: the policy rate passes through to EONIA slightly above one, so the market overreacts a little to rate decisions, and the negative sentiment coefficients act as a correction.

Limitations

A bag-of-words score misreads context. “Inflation is not expected to rise” scores as hawkish, because “not” is removed as a stop word, and “we see downside risks to inflation” also scores as hawkish, because the pattern ris* matches “risks”. This may explain part of the rise of the index after 2014, when rates stay at zero and the ECB talks about future normalization instead of moving rates. The sample ends in 2016, and the paper’s robustness check with the deposit facility rate is not reproduced.

View the code