Central banks use press conferences and other public statements to communicate their monetary policy decisions and their assessments of economic developments. This influences market participants’ expectations regarding the future monetary policy stance. Communication, like conventional monetary policy instruments, can thus have an impact on aggregate demand and inflation dynamics. A significant factor in the signal effect of communication is whether the statements are indicative of restrictive or accommodative monetary policy, commonly referred to as “hawkish” and “dovish”, respectively. An optimistic or pessimistic tone can also change the effect.

At present, state-of-the-art technologies such as artificial intelligence (AI) are increasingly being used to analyse monetary policy communication in an efficient way. The Bundesbank has developed a novel AI model named the Monetary-Intelligent Language Agent (MILA) that can produce detailed, consistent and transparent evaluations of monetary policy texts. MILA classifies individual sentences, taking into account the macroeconomic context, and justifies its own assessment. On the basis of this sentence-by-sentence classification, the model calculates a comprehensible evaluation of the overall text.

According to MILA, the communication by the ECB Governing Council between 2011 and 2024 was broadly in line with the macroeconomic environment in the euro area. Monetary policy communication was predominantly dovish, especially at the peak of the coronavirus pandemic in 2020. In 2021, the Governing Council increasingly balanced its inflation narrative, but communication on key interest rates initially remained dovish. For the period of monetary policy tightening between 2022 and 2023, MILA sees a marked shift towards hawkish press conferences and speeches. In 2024, communication became more balanced.

AI analysis does not replace human expertise, but improves the understanding of monetary policy communication and its impact. However, the increasing prevalence of AI analyses could also reduce the incentives for different market participants to obtain and evaluate their own information. This could lead to lesser diversity of opinion and make central bank communication more challenging. Critical examination of AI-assisted text analysis and the associated risks therefore remains essential.

1 The role of communication in monetary policy

In recent decades, central bank communication has taken on an increasingly important role in monetary policy. It is now a key instrument in the established toolkit that central banks use to steer their monetary policy stance (see the supplementary information entitled “Measuring the monetary policy stance”). In fact, some economists even argue that communication is the single most important monetary policy instrument.

economist Michael Woodford writes: “Because the key decision-makers in an economy are forward-looking, central banks affect the economy as much through their influence on expectations as through any direct, mechanical effects of central bank trading in the market for overnight cash.” In the view of Ben Bernanke, former Chairman of the Board of Governors of the

Federal Reserve, communication is actually more important than key interest rate decisions themselves: “Monetary policy is 98% talk and 2% action”. See Woodford (2005) and Bernanke (2022).

Monetary policy communication through press conferences, speeches and other public statements conveys messages about monetary policy decisions and the state of the economy.

In addition, interviews, publication of data, economic forecasts, regular reports, or minutes of meetings could also be interpreted as communication by central banks. See, for example, Deutsche Bundesbank (2000), Blinder et al. (2008), Apel and Blix Grimaldi (2014), and Istrefi et al. (2024).

These messages, in turn, influence market participants’ expectations and thus also market interest rates, which are of critical importance for financing conditions and the transmission of monetary policy. In this way, through additional signal effects, communication complements and supports the direct impact of changes to key interest rates or asset purchase programmes.

Monetary policy communication sends direct and indirect signals about the monetary policy stance, particularly regarding the future path of key interest rates. Direct information is represented by statements by the central bank concerning monetary policy instruments, primarily including statements on its current decisions and signals regarding the future path of key interest rates.

An especially strong form of direct information is “forward guidance”, whereby the central bank makes explicit statements on future key interest rates or its monetary policy reaction function. The

ECB

issued forward guidance for the first time in July 2013. At that time, it communicated that it “expect[ed] the key

ECB

interest rates to remain at present or lower levels for an extended period of time”. See Deutsche Bundesbank (2013), Hartmann and Smets (2018), Rostagno et al. (2019), Weidmann (2019), and Taskforce on Rate Forward Guidance and Reinvestment (2022).

Indirect information is transmitted by the central bank via the economic narrative, which consists, in particular, of its assessment of the current and future path of inflation and the real economy.

Direct and indirect information can be “hawkish” or “dovish”, indicating rising or high key interest rates and falling or low key interest rates, respectively. “Hawkish” describes a position advocating restrictive monetary policy to safeguard price stability. Such communication emphasises the need to raise key interest rates, or maintain them at an elevated level, in order to combat high rates of inflation. A “dovish” stance, by contrast, is in favour of accommodative monetary policy. Dovish statements argue for lowering key interest rates, or maintaining them at a low level, in order to stimulate aggregate demand and thereby bring up excessively low rates of inflation. If market participants perceive communication to be unexpectedly hawkish or dovish, they will adjust their expectations about the future monetary policy stance. Surprising communication can thus have an impact on growth and inflation.

This means that communication can have a similar effect to a monetary policy shock, i.e. a change in key interest rates that was not expected by market participants. See Bernanke and Blinder (1992) and Christiano et al. (1999).

Furthermore, a role in the monetary policy signal effect is also played by whether communication exhibits positive or negative sentiment. Positive statements by the central bank about the economic situation or inflation dynamics can boost optimism in markets and among the general public.

See Picault and Renault (2017).

By contrast, negative statements – e.g. about undesirable economic developments, uncertainties or risks – can lead to pessimism.

Not only the economic narratives of central banks but also their monetary policy decisions themselves can send signals about the economic situation. The literature describes this using the term “central bank information effect” (or “information shock”). See Cieslak and Schrimpf (2019), Deutsche Bundesbank (2020, 2021b), Jarociński and Karadi (2020), and Kerssenfischer (2022).

For this reason, the sentiment of the economic narrative also shapes market participants’ expectations about future monetary policy measures and economic developments.

In addition to verbal communication, sentiment can also be conveyed through non-verbal communication, such as the speaker’s facial expression or tone of voice. See Curti and Kazinnik (2023), Gorodnichenko et al. (2023), and Kanelis and Siklos (2023).

This can likewise affect actual consumption and investment decisions. Sentiment depends largely on the inflation context: a hypothetical statement of “price pressures are rising” would be considered positive if inflation were below target, but negative if inflation were above target. In both instances, however, this statement would be classified as hawkish. The aspects of positivity and negativity are therefore not equivalent to the aspects of hawkishness and dovishness. Moreover, normative statements using the two scales should only be derived with consideration of the overall macroeconomic context. The assessment of whether specific communication is consistent with the targeted objectives from the perspective of monetary policy decision-makers must always take account of the economic situation and the correspondingly appropriate monetary policy stance.

In addition to the aspects mentioned here, the comprehensibility or complexity of the language used is also relevant in the central bank context. See Hayo et al. (2022).

In the Eurosystem, press conferences following the meetings of the ECB

Governing Council play a key role in communication with the public.

Following monetary policy meetings of the

ECB

Governing Council, the monetary policy decisions taken are published in a “Monetary policy decisions” press release. This is followed by a press conference opening with a speech by the

ECB

President, which, since the rollout of the new monetary policy strategy in July 2021, has been known as the “Monetary policy statement”.

Until June 2021, this speech was known as the “Introductory statement”. The first press conference took place on 9 June 1998, just a few days after the

ECB

was established. The first press release on monetary policy decisions was published on 4 March 1999, two months after the introduction of the euro.

In addition to an explanation of the monetary policy decisions, this also includes the Governing Council’s assessment of current economic developments. The analysis presented in this article focuses on texts from this first part of the press conferences during the period from November 2011 (the beginning of Mario Draghi’s term as

ECB

President) to December 2024. It abstracts from the second part of the press conference, in which the

ECB

President answers questions from members of the press.

In rare cases, questions are also answered by the

ECB

Vice-President.

Furthermore, monetary policy speeches by the

ECB

Executive Board during the period from November 2011 to August 2024 are also analysed. Public statements by central bank representatives can explain past decisions or provide individual viewpoints on economic developments. Speeches thus transmit relevant monetary policy signals in the periods between the official meetings of the

ECB

Governing Council. The analysis comprises a total of 119 monetary policy statements and 377 speeches.

The corresponding texts are available on the

ECB

’s website. See European Central Bank (2019).

Supplementary information

Measuring the monetary policy stance

The monetary policy stance is multidimensional due to the use of multiple monetary policy instruments. The various monetary policy instruments have differing effects on the maturity segments of the yield curve, on asset prices and on exchange rates.

See Altavilla et al. (2019) and Deutsche Bundesbank (2023).

While the level of key interest rates is crucial, especially for very short-term market rates, communication plays a key role for medium-term market rates. Central banks’ asset purchases affect longer-term maturities, in particular. This is relevant because the entire yield curve is crucial for monetary policy transmission, i.e. the impact of monetary policy on households’ and firms’ consumption and investment decisions.

See Deutsche Bundesbank (2024a, 2024b) and Odendahl et al. (2024).

In addition, the use of monetary policy instruments affects risk premia in a broad spectrum of different financial market segments, especially through investors’ varying risk appetite.

See Deutsche Bundesbank (2025) and Caballero et al. (2024).

It may therefore be insufficient to assess the monetary policy stance solely on the basis of the level of key interest rates. For example, the

ECB

key interest rate stood at 4% in both March 2024 and August 2007, while the ten-year

OIS

rate stood at 2.5% and 4.6%, respectively. Measured by the level of this long-term interest rate, the monetary policy stance may therefore have been less restrictive in March 2024 than in August 2007.

A comprehensive measure of the monetary policy stance is the Bundesbank’s proxy monetary policy rate (PMPR

which combines information from the risk-free yield curve and risky asset prices.

This indicator is based on 11 financial market variables: risk-free interest rates over various maturities, yields on government and corporate bonds, spreads and the effective exchange rate of the euro. These financial market prices are available on a daily basis, which means that the indicator of the monetary policy stance can also be calculated at the same frequency. To this end, a statistical approach in the form of a principal component analysis is used to identify the typical co-movement among the financial variables. First, two common factors are extracted.

A similar methodological approach is used by Doh and Choi (2016), Choi et al. (2022) and Jude and Levieuge (2024).

These are then used in a regression as explanatory variables for the short-term interest rate (the

OIS

interest rate with a maturity of seven days) for the period from 2005 to 2012 (before the key interest rates reached the effective lower bound). The estimated values from this regression over the entire observation period yield the

PMPR

The PMPR

suggests that the monetary policy stance between 2012 and 2021 was considerably more accommodative than signalled by key interest rates in isolation

(see Chart 4.1)

Before 2012, the dynamics of the

PMPR

and the short-term interest rate were, by construction, largely similar – the regression serves to ensure that the indicator largely corresponds to the short-term interest rate during this period, as monetary policy was steered almost exclusively by the policy rate at that time. After 2012, however, the indicator reflects the impact of other monetary policy instruments that were introduced after the effective lower bound had been reached. The

PMPR

suggests that forward guidance and central bank asset purchase programmes made the monetary policy stance considerably more accommodative at the lower bound.

This is consistent with other indicators of the monetary policy stance at the lower bound, such as shadow rates. See Deutsche Bundesbank (2017) and Geiger and Schupp (2018).

According to the indicator, this effect was equivalent to a further reduction in key interest rates of around 150 basis points.

According to the PMPR

between 2022 and 2023, the extent of monetary policy tightening was comparable to the actual rise in key interest rates.

Based on the

PMPR

however, the monetary policy tightening began as early as the start of 2022, i.e. six months before the first interest rate hike. This reflects the fact that the anticipation effects of future key interest rate changes are captured at an early stage by the

PMPR

via their impact on medium-term and long-term interest rates.

The gradual reduction in the balance sheet, which began in 2022, is also evident in the

PMPR

albeit to a lesser extent. The primary determinant of yield effects is the level of securities holdings.

The indicator shows that the maximum tightening impulse was in September 2023, when key interest rates were last raised. The stance has recently been more accommodative, in some cases considerably so, than the short-term interest rates suggest. This is a reflection of lower medium-term and long-term interest rates as well as narrower spreads on government and corporate bonds by historical standards.

For more information on the impact of monetary policy measures on financial markets and risk premia since the tightening cycle began in 2022, see also Deutsche Bundesbank (2025).

Footnotes

See Altavilla et al. (2019) and Deutsche Bundesbank (2023).
See Deutsche Bundesbank (2024a, 2024b) and Odendahl et al. (2024).
See Deutsche Bundesbank (2025) and Caballero et al. (2024).
A similar methodological approach is used by Doh and Choi (2016), Choi et al. (2022) and Jude and Levieuge (2024).
This is consistent with other indicators of the monetary policy stance at the lower bound, such as shadow rates. See Deutsche Bundesbank (2017) and Geiger and Schupp (2018).
The gradual reduction in the balance sheet, which began in 2022, is also evident in the PMPR
, albeit to a lesser extent. The primary determinant of yield effects is the level of securities holdings.
For more information on the impact of monetary policy measures on financial markets and risk premia since the tightening cycle began in 2022, see also Deutsche Bundesbank (2025).

2 History of monetary policy communication analysis

In light of the increasingly transparent communication strategies of central banks since the 1990s, market participants have been focusing greater attention on analysing monetary policy statements. Traditionally, central bank circles believed that monetary policy decision-makers should communicate with the public as little as possible and only cryptically. Over the past three decades, however, the predominant view has changed: now, the prevailing standpoint is that transparent communication can directly affect market participants’ expectations and thereby improve the efficiency of monetary policy.

Transparent communication steers market participants’ expectations, promotes a clear understanding of the monetary policy stance, and strengthens public confidence in the central bank. This helps to anchor long-term inflation expectations and thereby safeguard price stability. See Clarida et al. (1999), Deutsche Bundesbank (2000, 2004), Blinder et al. (2008), Weidmann (2018, 2021), and Blinder et al. (2024).

With this paradigm shift in the style of central bank communication, market participants began to intensively examine monetary policy statements using their human expertise. In particular, the ability to correctly interpret and identify nuances in communication became a valuable asset. This is because, from the perspective of financial markets, even minor signals from central bank representatives can contain meaningful information and thus lead to strong market reactions.

Even minimal changes in communication compared with previous statements can be significant and trigger considerable market movements. See Ehrmann and Talmi (2020).

For example, the euro depreciated significantly in October 2000 when

ECB

President Wim Duisenberg indicated that the

ECB

would no longer support the currency. The statement by

ECB

President Mario Draghi in July 2012 that “the

ECB

[was] ready to do whatever it [took]

to preserve the euro” also had a profound impact on markets.

Economic research has traditionally made use of human expertise to assess and quantify the characteristics of central bank communication. Economists assess monetary policy texts manually, analysing the monetary policy stance and sentiment, in particular. Here, based on the human assessment, texts or individual text excerpts are assigned a numerical value: -1 for dovish statements, 0 for neutral statements, and +1 for hawkish statements; and -1 for negative sentiment, 0 for neutral sentiment, and +1 for positive sentiment.

This approach is also referred to as “content analysis”. See Jansen and De Haan (2005) and Ehrmann and Fratzscher (2007).

This human classification offers numerous benefits. The specialist expertise utilised here enables contextual factors that are vital for interpreting monetary policy communication to be comprehensively incorporated and quantitatively condensed. In addition, it allows for a flexible response to differing or nuanced formulations as well as highly detailed evaluations of the texts. However, the advantages of human classification are offset by some key disadvantages. These assessments are, by their very nature, subjective and may be inconsistent over time, which can lead to misclassifications. Furthermore, this type of analysis takes a considerable amount of time.

To address the limitations of human classification, computer-assisted algorithms based on word frequencies have been developed. This methodology creates a lexicon of terms or word combinations that typically appear in monetary policy texts.

See Apel and Blix Grimaldi (2014), Hansen and McMahon (2016), Picault and Renault (2017), Lombardi et al. (2019), and Aruoba and Drechsel (2024).

Each entry in this lexicon is assigned a specific classification: for example, the combination of “inflation” and “rising” is classified as hawkish, while “growth” in conjunction with “falling” is classified as dovish. Using these categorisations, it is possible to determine the frequency of dovish or hawkish (or, alternatively, negative and positive) word combinations within a given text. On this basis, a score on a scale from -1 to +1 can be calculated for the entire text. Sentiment is usually determined by calculating the difference between the numbers of positive and negative sentences and then dividing this by the total number of non-neutral sentences:

$\begin{matrix} (1) & S e n t i m e n t = \frac{# p o s i t i v e - # n e g a t i v e}{# p o s i t i v e + # n e g a t i v e} \end{matrix}$

An overall value for the monetary policy stance can be calculated in a similar way using the numbers of hawkish and dovish sentences instead. The methodology using word frequencies has the major advantage of formal transparency and allows for consistent, detailed and comparatively fast evaluation. However, it is fairly mechanical in its application and cannot take the context of the text into consideration. As a result, this method responds sensitively to linguistic nuances and minor variations in terminology, and is therefore susceptible to misclassification.

In more recent times, analysis using artificial intelligence – especially large language models – has proven promising. Large language models (LLMs

)

are trained on extensive text data. This enables them to process and interpret natural language (natural language processing, or

NLP

This includes computer-assisted text preparation (“text mining”) as well as the identification and evaluation of statistical language patterns (“machine learning”). At first, specialised language models were used in economic research to evaluate monetary policy communication.

See Gorodnichenko et al. (2023), Nitoi et al. (2023), Pfeifer and Marohl (2023), Gambacorta et al. (2024), and Kanelis and Siklos (2025). This research is based on further developments of the

BERT

language model (Devlin et al. (2019)).

These models were trained using large volumes of text data and optimised using manually classified sentences from financial market or central bank contexts. Unlike the word frequency approach, this first generation of language models is, to a limited extent, able to take the immediate context of a text into consideration. However, their analysis is less transparent and still neglects information beyond the actual text excerpt itself. By contrast, the latest

LLMs

(such as

GPT

Llama and Gemini) are able to factor in additional information provided by the user and justify their assessments (see the supplementary information entitled “

How large language models work

”). With regard to the analysis of central bank communication, they thus offer the potential to combine the advantages of previous approaches, in particular the flexibility of human analysis with the automation of the word frequency methodology.

See Hansen and Kazinnik (2024).

However, a key challenge when analysing texts with

LLMs

is obtaining a granular and transparent assessment. The next section presents a novel

model that can carry out such an analysis.

Supplementary information

How large language models work

Large language models (LLMs

)

are a form of artificial intelligence

(

)

specifically designed to understand and generate natural language.

These models are based on neural network models consisting of multiple interlinked layers of neurons and are capable of recognising complex patterns in data. Each layer processes the input data and transmits the results to the next layer, through which the model learns increasingly abstract attributes of the data. Using activation functions between the layers enables non-linearities to be modelled.

The Transformer architecture is one innovative type of these neural networks. Transformers were popularised by models such as Google AI’s BERT

(Bidirectional Encoder Representations from Transformers) and Open AI’s Generative Pre-trained Transformer

(

GPT

)

models and are based on the attention mechanism.

They can capture words within a section of text, irrespective of their position, by using selective weights to understand the context. This enables them to recognise relationships between words that are far apart, which is a decisive factor in the processing of natural language. This represents a significant advantage over models from the previous generation, which were specifically designed for sequential data processing (recurrent neural networks

(

RNN

)).

LLMs

are trained using enormous amounts of text data from various sources such as books, articles and websites, and their response accuracy is optimised by extensive evaluation measures.

By being trained on these extensive datasets, the model recognises complex patterns and semantic relationships in the language, allowing it to create contextual content on a variety of topics that is often barely distinguishable from human-generated text.

LLMs

use techniques such as self-supervised learning, where the next token (word or character) is predicted based on the context of the previous tokens. This results in a deep understanding of syntax, grammar, and stylistic nuances, allowing

LLMs

to be used in a wide range of applications. In addition, the performance of the models is evaluated and improved through reinforcement learning from human feedback

(

RLHF

In this approach, humans provide feedback on the generated texts, which is then used to further optimise the model and improve its ability to create high-quality and relevant content.

Although LLMs

are often referred to as

they do not possess any real awareness or intelligence.

They generate responses based on the statistical patterns learnt from their training data and can therefore occasionally produce errors or inaccurate responses (hallucination). Nevertheless, the development and application of

LLMs

represent significant technological progress with the potential to transform and improve many aspects of our everyday lives.

3 Monetary-Intelligent Language Agent (MILA
)

The Bundesbank has developed a new model called the Monetary-Intelligent Language Agent (MILA

)

for the purposes of analysing central bank communication.

See Geiger et al. (2025).

MILA

is an

model that analyses monetary policy texts in an automated and rules-based manner. The model approaches the task in a granular, transparent and consistent way: it assesses individual sentences taking into account contextual information, justifies its classification and provides an assessment for the overall text using mathematical formulae. Decomposing the text into individual sentences makes for greater transparency in the analysis, as this unit of text can be understood and assessed on a standalone basis by humans. This means that the human assessment of a sentence can be compared directly with the classification and associated justification produced by

and checked for plausibility. Human readers can therefore easily interpret and evaluate the results.

MILA

is based on state-of-the-art language models, chained instructions and manual sample responses.

At its heart is a very powerful language model called Llama 3.1 70B.

See Grattafiori et al. (2024).

The methodology for analysing central bank communication is based on established concepts applied when interacting with

based language models (prompt engineering). The language model is assigned the identity of an economist specialised in monetary policy analysis at a central bank (role-based prompting). Prompt chaining is employed, with the model being given a set of multi-step instructions to follow as it analyses a monetary policy text.

Wu et al. (2022) and Sun et al. (2024) show that breaking tasks into a number of steps improves the accuracy of results and increases transparency for the user.

In addition to the specific task, each work step includes sample texts demonstrating a correct assessment in the eyes of an expert (few-shot prompting).

See Brown et al. (2020).

The examples also specify that a brief explanation should be provided

before

the actual classification in order to reduce the risk of the language model producing inaccurate assessments.

Language models generate text by determining the most likely next word based on the foregoing context (see the supplementary information entitled “

How large language models work

”).

The following section describes the way in which

MILA

approaches the task of analysing

ECB

press conferences (see Chart 4.2).

This article uses the terms “press conference” and “

MPS

” synonymously. However, as explained in the section entitled "

The role of communication in monetary policy

", press conferences also include a second part, which is disregarded here.

MILA

’s first step when presented with an

ECB

press conference is to derive the inflation context; this inflation context serves as background information for the actual classification.

Information on current inflation dynamics is a crucial component when analysing many statements concerning monetary policy. For example, the hypothetical remark “inflation rose last month” can only be accurately classified as positive or negative if one knows where inflation stands at the time.

The first generation of language models is not capable of factoring in context beyond the text that it is processing. As a result, producing a state-dependent evaluation of inflation-related communication is challenging for these models.

If inflation were above the 2% target, the statement would be negative; if inflation were below target, the statement would be positive. Furthermore, the fact that the

ECB

Governing Council aims for an inflation rate of 2% over the medium term means that medium-term inflation expectations are a relevant part of examining many other statements.

MILA

extracts the inflation context from the

MPSs

made when new

ECB

/Eurosystem staff macroeconomic projections are released (March, June, September and December).

In principle, it would also be possible at this point in the process to have experts provide additional direct information (e.g. on the output gap) or to extract additional information.

MILA

then proceeds by analysing the communication in

ECB

press conferences on monetary policy instruments and assessing whether it is hawkish or dovish.

The opening paragraph of the

MPS

typically lays out the decisions taken by the

ECB

Governing Council along with explanations and comments on the planned trajectory going forward.

MILA

analyses this section of the text along four dimensions: interest rate decision, interest rate outlook, inflation, and general tone. A pre-defined scoring system is used to formulate a “Decision Hawk-O-Meter” from the assessment of these dimensions.

For a more detailed description of the scoring system, see Geiger et al. (2025).

This indicator reflects the degree of hawkishness or dovishness of direct monetary policy communication in the eyes of the digital economist. The scale ranges from -1 (very dovish) to +1 (very hawkish).

In a third step, MILA

examines the hawkishness or dovishness of the economic narrative by classifying individual sentences from the press conference.

Once the monetary policy decisions have been presented,

ECB

press conferences feature a detailed economic narrative.

This section is always introduced with phrases such as “Let me now explain our assessment in greater detail …” or “I will now outline in more detail …”.

This section covers the Governing Council’s assessment of real economic developments, inflation dynamics, and the financial and monetary landscape. Each sentence is classified into one of five categories: dovish, moderately dovish, neutral, moderately hawkish, hawkish (see the supplementary information entitled “

Examples of classifications by

MILA

”). When performing its assessment,

MILA

draws on additional context by taking account of the inflation context for the date in question that was ascertained in the first step plus the monetary policy decisions as well as previous sentences. On the basis of the sentence-level classification, a Hawk-O-Meter for the entire economic narrative in the text at hand is calculated outside of the

LLM

by applying the following formula:

See Apel and Blix Grimaldi (2014), Picault and Renault (2017), Kanelis and Siklos (2025), and Geiger et al. (2025).

$H a w k - O - M e t e r = \frac{# h a w k i s h + 0.5 * (# m o d e r a t e l y h a w k i s h - # m o d e r a t e l y d o v i s h) - # d o v i s h}{# h a w k i s h + # m o d e r a t e l y h a w k i s h + # m o d e r a t e l y d o v i s h + # d o v i s h}$

The resulting indicator can be interpreted in economic terms: a text’s Hawk-O-Meter takes a value of -1 if MILA

assesses all statements as being dovish and a value of +1 if it assesses all statements as being hawkish. In addition to looking at the narrative as a whole, it is also possible to make a differentiated analysis, focusing on specific thematic aspects.

To accomplish this,

MILA

assigns the individual sentences to various topics (topic modelling). See Deutsche Bundesbank (2020) and Bohl et al. (2023).

For example,

MILA

is also capable of calculating the Hawk-O-Meter just for statements directly related to inflation or the real economy.

MILA

also conducts a sentence-level analysis of the sentiment of an

MPS

as a whole and determines whether it is positive, neutral or negative.

As discussed in the section entitled “

The role of communication in monetary policy

”, the sentiment dimension is not to be conflated with the hawkishness or dovishness of monetary policy. For instance, a press conference might see the

ECB

Governing Council communicating in a hawkish and positive manner or in a hawkish and negative manner (for illustrations, see the supplementary information entitled “

Examples of classifications by

MILA

”). The sentiment metric therefore differs from the Hawk-O-Meter for the economic narrative described above. At this stage in the procedure,

MILA

again examines individual sentences, but this time they are classified as positive, neutral or negative. In order to ensure that the output focuses on the economic narrative, the entire

MPS

is taken into account, but statements without a direct reference to economic developments or monetary policy decisions are classed as neutral. Formula (1) is used to calculate the sentiment for the text as a whole: the indicator takes a value of -1 if the text contains only negative sentences and a value of +1 if the text contains only positive sentences.

MILA

is also capable of analysing speeches given by members of the

ECB

Governing Council.

When doing so,

MILA

starts by examining whether the speech is related to monetary policy and thus whether it makes sense to attempt to characterise it in terms of hawkishness or dovishness.

Speeches of a non-monetary policy nature can generate misclassifications. See Bohl et al. (2023).

Provided that a speech is deemed to be of a monetary policy nature,

MILA

then proceeds to generate the Hawk-O-Meter and the sentiment metric for the economic narrative according to the procedure described above. The classification process draws on the inflation context derived from

ECB

press conferences. A Decision Hawk-O-Meter can also be calculated, factoring in all sentences related directly to monetary policy instruments.

Evaluating speeches is a more complex and time-consuming exercise than evaluating press conferences. This is mainly due to the less standardised language, the large number of topic areas and varying text formats. Automated analysis is thus more difficult. In principle, however, the methodology described here is also suitable for handling other monetary policy texts, such as interviews or minutes of meetings (provided due consideration is given to their specific features).

Supplementary information

Examples of classifications by MILA

Table 4.1 lists various examples of how MILA

classifies monetary policy sentences and justifies its own assessments.

Table 4.1: Examples of classifications by MILA
¹
Sentence	Source	Hawk-O-Meter	Justification for Hawk-O-Meter	Sentiment	Justification for sentiment
Domestic inflation is still elevated, as wage pressures in the euro area remain strong.	ECB press conference, 17 October 2024	Hawkish	The sentence is classified as hawkish because it explicitly refers to upside risks to price stability, particularly to strong wage pressures, which could raise the inflation rate above target.	Negative	The sentence is classified as negative because it indicates that domestic inflation remains high due to rising wages, signalling that inflation, despite expectations of a decline, will not fall back down to target over the course of the following year.
Annual inflation fell further to 1.7 % in September, its lowest level since April 2021.	ECB press conference, 17 October 2024	Dovish	The sentence is classified as dovish because inflation is below target, which could indicate the need for monetary policy accommodation to support price stability.	Negative	The sentence is classified as negative because it indicates that the current rate of inflation is below the ECB ’s target and has continued to fall, moving it further away from the target of 2 %.
The Governing Council expresses its full support to the people of Ukraine.	ECB press conference, 10 March 2024	Neutral	The sentence is classified as neutral as it has no impact on monetary policy or the economic context; it therefore has no relation to inflation or the ECB ’s mandate.	Neutral	The sentence is classified as neutral because it provides a declaration of support and solidarity but does not convey economic information or sentiment regarding inflation or economic conditions.
The Russian invasion of Ukraine is a watershed for Europe.	ECB press conference, 10 March 2024	Neutral	The sentence is classified as neutral as it contains no information on inflation, monetary policy or economic developments.	Negative	The sentence is classified as negative because it refers to a significant and potentially destabilising event – Russia’s war of aggression against Ukraine – which could have far-reaching and adverse consequences for stability and the economy in Europe.
The current increase in inflation is expected to be largely temporary and underlying price pressures are building up only slowly.	ECB press conference, 9 September 2021	Moderately dovish	The sentence is classified as moderately dovish because it suggests that inflationary pressures are temporary and will build up only gradually, reducing the need for immediate monetary policy tightening, but not necessarily implying easing.	Positive	The sentence is classified as positive as it suggests that the inflation rate, which is currently above target, is likely to be temporary and that underlying price pressures will increase only gradually, indicating that inflation will go back down towards target.
Looking ahead, we will continue to closely monitor the situation in financial markets, as well as the potential implications for the monetary policy stance and for the outlook for price stability.	ECB press conference, 16 July 2015	Neutral	The sentence is classified as neutral because it merely expresses the ECB ’s intention to monitor the situation without providing specific information on inflation or monetary policy implications.	Neutral	The sentence is classified as neutral as it contains no information on inflation risks or economic developments that would justify tighter or looser monetary policy.
1 Classification by MILA , an AI model based on Llama 3.1. The Hawk-O-Meter measures whether communication is indicative of restrictive (hawkish) or accommodative (dovish) monetary policy. Sentiment measures the degree of positivity or negativity of communication.

4 Hawkish or dovish? A Hawk-O-Meter for monetary policy communication

According to MILA

the economic narrative of the

ECB

Governing Council has been largely dovish since the end of 2011

(see Chart 4.3).

The press conferences under Mario Draghi’s presidency (November 2011 to October 2019) are predominantly rated as particularly dovish. This period includes the European sovereign debt crisis, during which the Governing Council judged monetary policy transmission to be impaired and assessed the economic outlook as weak.

MILA

classifies the economic narrative as especially dovish around the time of Draghi’s “Whatever it takes” statement in 2012. The press conferences during the era of negative key interest rates and other unconventional monetary policy measures as of 2014 are also considered to be very dovish.

The deposit facility rate was lowered into negative territory in June 2014. In January 2015, the

ECB

Governing Council launched the asset purchase programme

(

APP

During this period, the Governing Council stressed the very low and, in some cases, negative inflation rates as well as the elevated risk of these dynamics becoming entrenched. The Governing Council’s press conferences at the start of the

COVID

19 pandemic in early 2020, when Christine Lagarde was

ECB

President, are highly dovish according to

MILA

At the time, the

ECB

Governing Council stressed that the euro area was facing an economic contraction of a magnitude and speed unprecedented in peacetime.

Against this backdrop, the

ECB

Governing Council launched the pandemic emergency purchase programme

(

PEPP

)

in March 2020.

Overall, looking at the period from 2011 to 2021, it is particularly striking that the Hawk-O-Meters for inflation and the real economy exhibit strong correlation over large parts. This is consistent with a narrative of prolonged weakness in aggregate demand, associated downward pressure on inflation dynamics, and the need for highly accommodative monetary policy.

Hawk-O-Meter of economic narratives in ECB press conferences

The economic narrative turned less dovish in 2021 and was hawkish during the period of monetary policy tightening between 2022 and 2023. MILA

finds that the inflation narrative grew less dovish over the course of 2021. At the time, the

ECB

Governing Council highlighted the elevated rate of inflation, but considered this increase to be transitory. At the beginning of 2022, it revised its assessment in the light of unexpectedly strong rises in inflation rates; as a result, for the first time in around ten years, communication around inflation turned markedly and more persistently hawkish. The Hawk-O-Meter peaked in June 2022, when the Governing Council announced that the next meeting would see a first rise in key interest rates. The inflation narrative remained highly hawkish until the end of 2022, when inflation in the euro area had surpassed the 10% mark. In 2023 and 2024, the inflation narrative gradually turned less hawkish. Most recently,

MILA

has characterised it as balanced. In contrast to the inflation narrative, the Governing Council’s communication concerning real economic developments remained predominantly dovish or – at most – balanced during this period. This kind of divergence in the narrative between inflation and the real economy indicates that, in the wake of the start of Russia’s war of aggression against Ukraine in February 2022, the

ECB

Governing Council primarily emphasised supply-side disruptions – with inflationary but growth-damaging effects. The rather dovish communication about the real economy since mid-2023 also reflects the structural weakness in growth in the euro area.

Ahead of the period of monetary policy tightening, the ECB

Governing Council’s communication on monetary policy decisions was markedly more dovish than its inflation narrative

(see Chart 4.4). The Eurosystem concluded its strategy review in the summer of 2021.

See Deutsche Bundesbank (2021a) and European Central Bank (2021).

On the back of this, the Governing Council modified its communication on the future trajectory of key interest rates, known as forward guidance. It underlined its intention to maintain a persistently accommodative monetary policy stance to achieve the inflation target, even if this might imply “a transitory period in which inflation is moderately above target”.

MILA

classifies this communication as very dovish. The Governing Council remained markedly dovish in its direct communication on monetary policy decisions afterwards as well, especially in comparison with its inflation narrative. In the March 2022 press conference, the Governing Council stressed that any adjustments to the key interest rates would take place “some time” after the end of net purchases under the

APP

In June of that year, the Governing Council then decided to discontinue net asset purchases under the

APP

and announced its intention to raise interest rates again for the first time at its next meeting.

MILA

accordingly posts a surge in the Decision Hawk-O-Meter between March and June 2022.

The inflation narrative and key interest rate communication signal the monetary policy stance and telegraph future key interest rate decisions. The monetary policy stance in the euro area (see the supplementary information entitled “Measuring the monetary policy stance”) gradually became less accommodative at the end of 2021 (see Chart 4.4). At that point in time, the deposit facility rate still stood at -0.5%. However, prompted, amongst other things, by the ECB

Governing Council’s hawkish communication on the subject of inflation dynamics in its press conferences, market participants priced in a higher likelihood of an impending rise in key interest rates, pushing up longer-term market interest rates. The first key interest rate hike then came in July 2022, half a year after communication had left dovish territory, with the Governing Council raising the deposit facility rate by 50 basis points from -0.5% to 0%. Over the rest of the monetary policy tightening cycle, communication continues to act as an indicator of developments to come: the Hawk-O-Meters for inflation and key interest rates peak around nine months ahead of the final key interest rate increase in September 2023. The reduction in the degree of monetary policy restriction in 2024 through four key interest rate cuts of 25 basis points each was also preceded by less hawkish communication. This makes clear that central banks utilise their communication in press conferences to signal future monetary policy decisions, both via the economic narrative and via the key interest rate path.

Hawk-O-Meter for ECB press conferences compared with changes in key interest rates

Since 2011, the tone of the ECB

Executive Board’s monetary policy speeches has evolved in line with the

ECB

press conferences and the macroeconomic environment in the euro area.

MILA

finds that the public remarks made by the

ECB

Executive Board in the period from 2011 to 2021 were likewise predominantly dovish, with a clear pivot towards hawkish speeches at the start of 2022 (see Chart 4.5, left panel). The changes in the Hawk-O-Meter for monetary policy speeches can be attributed to macroeconomic developments in the euro area (see Table 4.2). The rates of inflation and

GDP

growth have statistically significant effects on the tone of speeches by the

ECB

Executive Board: on average, the speeches become more hawkish when inflation or growth go up. The estimated coefficients are consistent with a conventional monetary policy rule whereby key interest rates are adjusted more strongly in response to inflation. This also tallies with the

ECB

’s price stability mandate. The results indicate that monetary policy decision-makers respond to current macroeconomic developments and signal their views with regard to future monetary policy decisions.

Hawk-O-Meter for monetary policy speeches and ECB press conferences

Table 4.2: Estimated impact of macroeconomic variables on the Hawk O-Meter for monetary policy speeches by the Executive Board of the ECB
Variable	Hawk-O-Meter¹⁾
Variable	(1)	(2)	(3)	(4)	(5)
HICP inflation rate	0.0946***	0.1295***	0.0638***	0.0677***	0.0673***
HICP inflation rate	(0.005)	(0.005)	(0.010)	(0.009)	(0.009)
GDP growth rate	0.0147***	0.0024	0.0121***	0.0109***	0.0092**
GDP growth rate	(0.004)	(0.003)	(0.005)	(0.004)	(0.004)
Fixed effects	-	Speaker	Year	Speaker, year	Speaker, year, speaker-year
1 Estimated coefficients from a panel regression. The dependent variable is the Hawk-O-Meter for individual monetary policy speeches made by members of the ECB ’s Executive Board between November 2011 and August 2024. All members with at least 16 monetary policy speeches are included, yielding a total of 350 observations. Explanatory variables are the inflation rate and GDP growth in the euro area (3-month averages in each case). Specifications (2)-(5) control for year-specific and/or speaker-specific fixed effects. Standard errors are in brackets. Significance levels: * p < 0.01, p < 0.05, * p < 0.1.

5 Positive or negative? A sentiment indicator for monetary policy communication

According to MILA

sentiment in

ECB

press conferences has been predominantly positive since 2011

(see Chart 4.6, left panel). Following the lowering of the deposit facility rate to 0% in July 2012,

MILA

records negative sentiment, which persisted until around the end of 2013. This primarily reflects a pessimistic narrative about economic dynamics in the context of the European sovereign debt crisis. By contrast, sentiment was broadly optimistic between 2014 and the end of Mario Draghi’s term as

ECB

President in November 2019. The subsequent coronavirus pandemic was also reflected in

ECB

press conferences: while the lockdowns in spring 2020 and winter 2020-21 went hand in hand with negative sentiment, the increasing prevalence of vaccines and the reopening of the economy in 2021 led to optimism.

MILA

shows a renewed phase of pessimism coinciding with Russia’s war of aggression against Ukraine and the sharp rise in inflation at the beginning of 2022. The

ECB

press conference in October 2022, in which the sentiment indicator reaches a low, is particularly striking. At that time, financial markets reacted in a highly dovish way, even though the

ECB

Governing Council raised the key interest rates by 75 basis points and did not change the key interest rate outlook.

MILA

indicates that this may be attributable to the extremely pessimistic communication. Since the last interest rate hike in September 2023, sentiment has gradually become more positive.

A historical comparison of sentiment and the Hawk-O-Meter regarding the economic narrative shows that the two dimensions should not be considered equivalent (see Chart 4.6, right panel). According to MILA

the economic narrative at the

ECB

’s press conferences under President Draghi was generally dovish and balanced during the sovereign debt crisis. From the introduction of forward guidance in July 2013, however, communication was dovish and positive until the end of Draghi’s term of office. For 2014 to 2018, in particular,

MILA

’s assessment indicates satisfactory macroeconomic developments overall, which were supported by the accommodative monetary policy. Three different phases can be identified during President Lagarde’s term of office. A very dovish tone and markedly negative sentiment is recorded for the coronavirus pandemic. Sentiment during the subsequent period of monetary policy tightening from 2022 to 2023 was also negative, but communication was hawkish overall. During the phase of the reduction in the degree of monetary policy restriction that began in 2024, the economic narrative is broadly balanced and slightly dovish. According to

MILA

a combination of markedly hawkish and positive communication has only occurred for a short time over the past 14 years, namely during four

ECB

press conferences in the second half of 2018.

Sentiment of economic narratives in ECB press conferences

6 Discussion and outlook

Artificial intelligence offers promising opportunities for improving the analysis of monetary policy texts and thus considerably promoting the understanding of the characteristics of monetary policy communication. Using large language models, it is possible to evaluate different monetary policy texts from central banks in an automated and rules-based manner. AI

agents based on these models – such as

MILA

developed by the Bundesbank – also enable a detailed, transparent and multidimensional analysis, which combines the flexibility of human assessments with the consistency of formal methods.

These advances in analysis are faced with challenges. Interpreting and assessing monetary policy statements remains complex and context-dependent. In order to make an accurate assessment, communication must therefore be viewed in the context of the current economic situation, previous statements and the monetary policy stance. The risk of inaccurate assessments via automated systems continues to require careful checking and validation by human experts. In addition, it is essential to continuously adapt the employed models to new technical and linguistic developments as well as economic insights. This is the only way to ensure that AI

analysis remains useful and sufficiently precise.

From a central bank perspective, integrating AI

into the analysis of monetary policy texts holds the potential of improving central banks’ communication.

analyses can quantify the characteristics of monetary policy communication in a rules-based manner along various dimensions. Based on the quantitative indicators, it is possible to examine, for instance, the impact of these characteristics on financial market responses directly around the time of the communication. Systematic

based analysis thus contributes to a deeper understanding of the perception and impact of monetary policy statements. For example,

assisted text analysis can help central banks to improve their own communication. In particular,

could be used to check in advance whether the tone of communication adequately conveys the intended monetary policy signals. This could help to avoid undesirable market reactions such as occurred in October 2022 (see the section entitled “

Positive or negative? A sentiment indicator for monetary policy communication

”). Such an improvement could increase the effectiveness and accuracy of one of the most important monetary policy instruments. The increased efficiency and transparency of monetary policy communication resulting from this can ultimately help to strengthen the public’s trust in the central bank, anchor inflation expectations and thus fulfil the primary mandate of price stability.

However, the increasing prevalence of AI

analyses among financial market players is also associated with risks.

If market participants increasingly use

to analyse monetary policy statements and predict future decisions, their perceptions and interpretations of central bank communication could become increasingly homogeneous. This would be the case particularly if financial market players used very similar

models, relied primarily on

assessment for efficiency reasons and therefore invested less in acquiring additional information. As financial markets provide incentives to anticipate others’ assessments, market participants could increasingly align themselves with the

based assessments of other actors.

See Morris and Shin (2002) and Gao and Huang (2020).

As a result, market participants’ expectation formation process, and thus the price discovery process, would be less efficient and associated with the risk of fragile market sentiment.

See Angeletos and Pavan (2007) and Bikhchandani et al. (2024).

This could lead to increased volatility if market conditions changed unexpectedly rapidly or if the central bank made an unanticipated decision.

The availability of AI

analyses may also make monetary policy communication more challenging for central banks.

It may become more difficult for central banks to get their messages through to their target human audience effectively if the latter increasingly use

for interpretation. If central banks also use

to shape their communication, a scenario could arise in which machines end up communicating with other machines. This raises concerns about the effectiveness of monetary policy communication and the potential for feedback loops that could amplify market movements or lead to self-fulfilling expectations.

See Beretta et al. (2025).

It is vital that market participants as well as central banks are aware of these possible implications. A critical examination of

assisted analysis is essential to address these risks and ensure that central banks retain control over the impact of their monetary policy communication.

7 List of references

Angeletos, G.-M. and A. Pavan (2007), Efficient Use of Information and Social Value of Information , Econometrica, Vol. 75, pp. 1103-1142.

Altavilla, C., L. Brugnolini, R. S. Gürkaynak, R. Motto and G. Ragusa (2019), Measuring Euro Area Monetary Policy , Journal of Monetary Economics, Vol. 108, pp. 162-179.

Apel, M. and M. Blix Grimaldi (2014), How Informative Are Central Bank Minutes? , Review of Economics, Vol. 65, No 1, 2014, pp. 53-76.

Aruoba, S. B. and T. Drechsel (2024), Identifying Monetary Policy Shocks: A Natural Language Approach , NBER

Working Paper No 32417, May 2024.

Beretta, E., V. Dianova, G. Pankratz and U. Schollmeyer (2025), AI

based forecasting and market expectations: a self-fulfiling prophecy?, mimeo.

Bernanke, B. S. and A. S. Blinder (1992), The Federal Funds Rate and the Channels of Monetary Transmission , The American Economic Review, Vol. 82, No 4 (September), pp. 901-921.

Bernanke, B. S. (2022), The Fed

from the Great Inflation to

COVID

, Brookings Webinar, 23 May 2022.

Bikhchandani, S., D. Hirshleifer, O. Tamuz and I. Welch (2024), Information Cascades and Social Learning , Journal of Economic Literature, Vol. 62, No 3, pp. 1040-1093.

Blinder, A. S., M. Ehrmann, M. Fratzscher, J. De Haan and D.-J. Jansen (2008), Central Bank Communication and Monetary Policy: A Survey of Theory and Evidence , Journal of Economic Literature, Vol. 46, No 4, pp. 910-945.

Blinder, A. S., M. Ehrmann, J. De Haan and D.-J. Jansen (2024), Central Bank Communication with the General Public: Promise or False Hope, Journal of Economic Literature , Vol. 62, No 2, pp. 425-457.

Bohl, M. T., D. Kanelis and P. L. Siklos (2023), Central bank mandates: How differences can influence the content and tone of central bank communication , Journal of International Money and Finance, Vol. 130.

Brown, T. B. et al. (2020), Language models are few-shot learners , NIPS

’20: Proceedings of the 34th International Conference on Neural Information Processing Systems.

Caballero, R. J., T. E. Caravello and A. Simsek (2024), Financial Conditions Targeting , NBER

Working Paper No 33206, November 2024.

Choi, J., T. Doh, A. Foerster and Z. Martinez (2022), Monetary Policy Stance Is Tighter than Federal Funds Rate , FRBSF

Economic Letter, No 2022-30.

Christiano, L. J., M. Eichenbaum and C. L. Evans (1999), Monetary policy shocks: What have we learned and to what end? , Chapter 2, Handbook of Macroeconomics, Vol. 1, Part A, Elsevier.

Cieslak, A. and A. Schrimpf (2019), Non-monetary news in central bank communication , Journal of International Economics, Vol. 118, May 2019, pp. 293-315.

Clarida, R., J. Galí and M. Gertler (1999), The Science of Monetary Policy: a New Keynesian Perspective , Journal of Economic Literature, Vol. 37, No 4, pp. 1661-1707.

Curti, F. and S. Kazinnik (2023), Let’s face it: Quantifying the impact of nonverbal communication in FOMC

press conferences

, Journal of Monetary Economics, Vol. 139, pp. 110-126.

Deutsche Bundesbank (2025), Risk appetite in financial markets and monetary policy

925736

Article from the Monthly Report Risk appetite in financial markets and monetary policy

, Monthly Report, January 2025.

Deutsche Bundesbank (2024a), Developments in loans to enterprises in Germany since the start of the monetary policy tightening cycle

925736

Article from the Monthly Report Developments in loans to enterprises in Germany since the start of the monetary policy tightening cycle

, Monthly Report, July 2024.

Deutsche Bundesbank (2024b), Financing costs for banks in Germany in the monetary policy interest rate cycle

925736

Financing costs for banks in Germany in the monetary policy interest rate cycle

, Monthly Report, December 2024.

Deutsche Bundesbank (2023), Term structures in economic analysis , Monthly Report, January 2023.

Deutsche Bundesbank (2021a), The Eurosystem’s monetary policy strategy , Monthly Report, September 2021.

Deutsche Bundesbank (2021b), The impact of the Eurosystem’s monetary policy on Bitcoin and other crypto tokens , Monthly Report, September 2021.

Deutsche Bundesbank (2020), The impact of monetary policy on the euro’s exchange rate , Monthly Report, September 2020.

Deutsche Bundesbank (2017), Monetary policy indicators at the lower bound based on term structure models , Monthly Report, September 2017.

Deutsche Bundesbank (2013), Forward guidance – an indication of monetary policy stance in the future , Monthly Report, August 2013.

Deutsche Bundesbank (2004), Monetary policy under uncertainty , Monthly Report, June 2004.

Deutsche Bundesbank (2000), Monetary policy transparency , Monthly Report, March 2000.

Devlin, J., M. W. Chang, K. Lee, and K. Toutanova (2019), BERT

: Pre-training of Deep Bidirectional Transformers for Language Understanding

, in: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.

Doh, T. and J. Choi (2016), Measuring the Stance of Monetary Policy on and off the Zero Lower Bound , Federal Reserve Bank of Kansas City, Economic Review, Vol. 101, No 3, pp. 5-21.

Ehrmann, M. and J. Talmi (2020), Starting from a blank page? Semantic similarity in central bank communication and market volatility , Journal of Monetary Economics, Elsevier, Vol. 111(C), pp. 48-62.

Ehrmann, M. and M. Fratzscher (2007), Communication by Central Bank Committee Members: Different Strategies, Same Effectiveness? , Journal of Money, Credit and Banking, Vol. 39, No 2-3, pp. 509-541.

European Central Bank (2021), An overview of the ECB

’s monetary policy strategy

, July 2021.

European Central Bank (2019), Speeches dataset .

Gambacorta, L., B. Kwon, T. Park, P. Patelli and S. Zhu (2024), CB

LMS

: language models for central banking

, Bank for International Settlements,

BIS

Working Papers No 1215.

Gao, M. and J. Huang (2020), Informing the Market: The Effect of Modern Information Technologies on Information Production , The Review of Financial Studies, Vol. 33, No 4, April 2020, pp. 1367-1411.

Geiger, F. and F. Schupp (2018), With a little help from my friends: Survey-based derivation of euro area short rate expectations at the effective lower bound , Deutsche Bundesbank Discussion Paper No 27/2018.

Geiger, F., D. Kanelis, P. Lieberknecht and D. Sola (2025), Monetary-Intelligent Language Agent, Deutsche Bundesbank Technical Paper, forthcoming.

Gorodnichenko, Y., T. Pham and O. Talavera (2023), The Voice of Monetary Policy , American Economic Review, Vol. 113, No 2.

Grattafiori, A. et al. (2024), The Llama 3 Herd of Models , arXiv preprint 2407.21783.

Hansen, A. L. and S. Kazinnik (2024), Can ChatGPT Decipher Fedspeak ?, mimeo.

Hansen, S. and M. McMahon (2016), Shocking language: Understanding the macroeconomic effects of central bank communication , Journal of International Money and Finance, Vol. 99.

Hartmann, P. and F. Smets (2018), The European Central Bank’s Monetary Policy during Its First 20 Years , Brookings Papers on Economic Activity, Fall 2018, pp. 1-118.

Hayo, B., K. Henseler, M. S. Rapp and J. Zahner (2022), Complexity of ECB

communication and financial market trading

, Journal of International Money and Finance, Vol. 128.

Istrefi, K., F. Odendahl and G. Sestieri (2024), ECB

communication and its impact on financial markets

CEPR

Discussion Paper No. 19242.

Jansen, D.-J. and J. De Haan (2005), Talking heads: the effects of ECB

statements on the euro-dollar exchange rate

, Journal of International Money and Finance, Vol. 24, No 2, pp. 343-361.

Jarociński, M. and P. Karadi (2020), Deconstructing Monetary Policy Surprises – The Role of Information Shocks , American Economic Journal: Macroeconomics, Vol. 12, No 2.

Jude, C. and G. Levieuge (2024), The pass-through of monetary policy tightening to financing conditions in the euro area and the US

Is this time different?

SUERF

Policy Note, No 342, 8 March 2024.

Kanelis, D. and P. L. Siklos (2025), The ECB

press conference statement: Deriving a new sentiment indicator for the euro area

, International Journal of Finance and Economics, Vol. 30, No 1, pp. 652-664.

Kanelis, D. and P. L. Siklos (2022), Emotion in Euro Area Monetary Policy Communication and Bond Yields: The Draghi Era , mimeo.

Kerssenfischer, M. (2022), Information effects of euro area monetary policy , Economics Letters, Vol. 216, July 2022.

Lombardi, D., P. L. Siklos and S. St. Amand (2019), Asset Price Spillovers from Unconventional Monetary Policy: A Global Empirical Perspective , International Journal of Central Banking, Vol. 15, No 2.

Morris, S. and H. S. Shin (2002), Social Value of Public Information , American Economic Review, Vol. 92, No 5, pp. 1521-1534.

Niţoi, M., M.-M. Pochea and Ş.-C. Radu (2023), Unveiling the sentiment behind central bank narratives: A novel deep learning index , Journal of Behavioral and Experimental Finance, Vol. 38.

Odendahl, F., M. S. Pagliari, A. Penalver, B. Rossi and G. Sestieri (2024), Euro area monetary policy effects. Does the shape of the yield curve matter? , Journal of Monetary Economics, Vol. 147, Supplement, October 2024.

Picault, M. and T. Renault (2017), Words are not all created equal: A new measure of ECB

communication

, Journal of International Money and Finance, Vol. 79, pp. 136-156.

Pfeifer, M. and V. P. Marohl (2023), CentralBankRoBERTa: A fine-tuned large language model for central bank communications , The Journal of Finance and Data Science, Vol. 9.

Rostagno, M., C. Altavilla, G. Carboni, W. Lemke, R. Motto, A. Saint Guilhem and J. Yiangou (2019), A tale of two decades: The ECB

’s monetary policy at 20

, European Central Bank, Working Paper No 2346.

Sun, S., R. Yuan, Z. Cao, W. Li and P. Liu (2024), Prompt Chaining or Stepwise Prompt? Refinement in Text Summarization , arXiv 2406.00507.

Taskforce on Rate Forward Guidance and Reinvestment (2022), Rate forward guidance in an environment of large central bank balance sheets: a Eurosystem stock-taking assessment , European Central Bank, Occasional Paper No 290.

Weidmann, J. (2021), The evolving nature of central bank communication, speech at the virtual anniversary event to celebrate 50 years of Reuters Germany

Speech The evolving nature of central bank communication

, November 2021.

Weidmann, J. (2019), What the future holds – Benefits and limitations of forward guidance

660540

Speech Weidmann draws attention to the risks and side effects of a very loose monetary policy

Bundesbank President Jens Weidmann has drawn attention to the dangers of an even longer period of negative interest rates. "Holding interest rates lower for even longer would aggravate the risks and side effects of a very expensionary monetary policy", he said at this year‘s European Banking Congress in Frankfurt.

, speech at the European Banking Congress on 22 November 2019.

Weidmann, J. (2018), Central bank communication as an instrument of monetary policy , lecture at the Centre for European Economic Research on 2 May 2018.

Woodford, M. (2005), Central-bank communication and policy effectiveness , presentation at the Federal Reserve Bank of Kansas City Economic Policy Symposium, Jackson Hole.

Wu, T., E. Jiang, A. Donsbach, J. Gray, A. Molina, M. Terry and C. J. Cai (2022), PromptChainer: Chaining Large Language Model Prompts through Visual Programming , arXiv 2203.06566.

Footnotes

US
economist Michael Woodford writes: “Because the key decision-makers in an economy are forward-looking, central banks affect the economy as much through their influence on expectations as through any direct, mechanical effects of central bank trading in the market for overnight cash.” In the view of Ben Bernanke, former Chairman of the Board of Governors of the US
Federal Reserve, communication is actually more important than key interest rate decisions themselves: “Monetary policy is 98% talk and 2% action”. See Woodford (2005) and Bernanke (2022).
In addition, interviews, publication of data, economic forecasts, regular reports, or minutes of meetings could also be interpreted as communication by central banks. See, for example, Deutsche Bundesbank (2000), Blinder et al. (2008), Apel and Blix Grimaldi (2014), and Istrefi et al. (2024).
An especially strong form of direct information is “forward guidance”, whereby the central bank makes explicit statements on future key interest rates or its monetary policy reaction function. The ECB
issued forward guidance for the first time in July 2013. At that time, it communicated that it “expect[ed] the key ECB
interest rates to remain at present or lower levels for an extended period of time”. See Deutsche Bundesbank (2013), Hartmann and Smets (2018), Rostagno et al. (2019), Weidmann (2019), and Taskforce on Rate Forward Guidance and Reinvestment (2022).
This means that communication can have a similar effect to a monetary policy shock, i.e. a change in key interest rates that was not expected by market participants. See Bernanke and Blinder (1992) and Christiano et al. (1999).
See Picault and Renault (2017).
Not only the economic narratives of central banks but also their monetary policy decisions themselves can send signals about the economic situation. The literature describes this using the term “central bank information effect” (or “information shock”). See Cieslak and Schrimpf (2019), Deutsche Bundesbank (2020, 2021b), Jarociński and Karadi (2020), and Kerssenfischer (2022).
In addition to verbal communication, sentiment can also be conveyed through non-verbal communication, such as the speaker’s facial expression or tone of voice. See Curti and Kazinnik (2023), Gorodnichenko et al. (2023), and Kanelis and Siklos (2023).
In addition to the aspects mentioned here, the comprehensibility or complexity of the language used is also relevant in the central bank context. See Hayo et al. (2022).
Until June 2021, this speech was known as the “Introductory statement”. The first press conference took place on 9 June 1998, just a few days after the ECB
was established. The first press release on monetary policy decisions was published on 4 March 1999, two months after the introduction of the euro.
In rare cases, questions are also answered by the ECB
Vice-President.
The corresponding texts are available on the ECB
’s website. See European Central Bank (2019).
Transparent communication steers market participants’ expectations, promotes a clear understanding of the monetary policy stance, and strengthens public confidence in the central bank. This helps to anchor long-term inflation expectations and thereby safeguard price stability. See Clarida et al. (1999), Deutsche Bundesbank (2000, 2004), Blinder et al. (2008), Weidmann (2018, 2021), and Blinder et al. (2024).
Even minimal changes in communication compared with previous statements can be significant and trigger considerable market movements. See Ehrmann and Talmi (2020).
This approach is also referred to as “content analysis”. See Jansen and De Haan (2005) and Ehrmann and Fratzscher (2007).
See Apel and Blix Grimaldi (2014), Hansen and McMahon (2016), Picault and Renault (2017), Lombardi et al. (2019), and Aruoba and Drechsel (2024).
See Gorodnichenko et al. (2023), Nitoi et al. (2023), Pfeifer and Marohl (2023), Gambacorta et al. (2024), and Kanelis and Siklos (2025). This research is based on further developments of the BERT
language model (Devlin et al. (2019)).
See Hansen and Kazinnik (2024).
See Geiger et al. (2025).
See Grattafiori et al. (2024).
Wu et al. (2022) and Sun et al. (2024) show that breaking tasks into a number of steps improves the accuracy of results and increases transparency for the user.
See Brown et al. (2020).
Language models generate text by determining the most likely next word based on the foregoing context (see the supplementary information entitled “How large language models work”).
This article uses the terms “press conference” and “MPS
” synonymously. However, as explained in the section entitled "The role of communication in monetary policy", press conferences also include a second part, which is disregarded here.
The first generation of language models is not capable of factoring in context beyond the text that it is processing. As a result, producing a state-dependent evaluation of inflation-related communication is challenging for these models.
In principle, it would also be possible at this point in the process to have experts provide additional direct information (e.g. on the output gap) or to extract additional information.
For a more detailed description of the scoring system, see Geiger et al. (2025).
This section is always introduced with phrases such as “Let me now explain our assessment in greater detail …” or “I will now outline in more detail …”.
See Apel and Blix Grimaldi (2014), Picault and Renault (2017), Kanelis and Siklos (2025), and Geiger et al. (2025).
To accomplish this, MILA
assigns the individual sentences to various topics (topic modelling). See Deutsche Bundesbank (2020) and Bohl et al. (2023).
Speeches of a non-monetary policy nature can generate misclassifications. See Bohl et al. (2023).
Evaluating speeches is a more complex and time-consuming exercise than evaluating press conferences. This is mainly due to the less standardised language, the large number of topic areas and varying text formats. Automated analysis is thus more difficult. In principle, however, the methodology described here is also suitable for handling other monetary policy texts, such as interviews or minutes of meetings (provided due consideration is given to their specific features).
The deposit facility rate was lowered into negative territory in June 2014. In January 2015, the ECB
Governing Council launched the asset purchase programme (APP
).
Against this backdrop, the ECB
Governing Council launched the pandemic emergency purchase programme (PEPP
) in March 2020.
See Deutsche Bundesbank (2021a) and European Central Bank (2021).
See Morris and Shin (2002) and Gao and Huang (2020).
See Angeletos and Pavan (2007) and Bikhchandani et al. (2024).
See Beretta et al. (2025).

Monetary policy communication according to artificial intelligence Monthly Report – March 2025

1 The role of communication in monetary policy

Measuring the monetary policy stance

2 History of monetary policy communication analysis

How large language models work

3 Monetary-Intelligent Language Agent (MILA
)

Examples of classifications by MILA

4 Hawkish or dovish? A Hawk-O-Meter for monetary policy communication

5 Positive or negative? A sentiment indicator for monetary policy communication

6 Discussion and outlook

7 List of references

search.dialog.title

search.empty.title

Cite

Monetary policy communication according to artificial intelligence Monthly Report – March 2025

1 The role of communication in monetary policy

Measuring the monetary policy stance

2 History of monetary policy communication analysis

How large language models work

3 Monetary-Intelligent Language Agent (MILA MILA: monetary-Intelligent language agent )

Examples of classifications by MILA MILA: monetary-Intelligent language agent

4 Hawkish or dovish? A Hawk-O-Meter for monetary policy communication

5 Positive or negative? A sentiment indicator for monetary policy communication

6 Discussion and outlook

7 List of references

3 Monetary-Intelligent Language Agent (MILA
)

Examples of classifications by MILA