Accessible Version
Assessing Monetary Policy Globally: Evidence from LLMs and Semi-structural Models, Accessible Data
Figure 1. Global Monetary Policy Tone. LLM vs. Lexicon-based Measures
The figure compares two quarterly measures of the hawkish or dovish tone of central bank speeches in the Bank for International Settlements speech database over 1996–2025. The figure plots the quarterly average raw score produced by Claude Haiku 4.5 (blue solid line) and by the lexicon-based classifier (dark brown dotted line). Both methods score the same speeches on a scale from −1 to +1, where negative values indicate less hawkish language and positive values indicate more hawkish language. Because the plotted values are quarterly averages, their observed range is narrower, extending from approximately −0.33 to +0.15. A horizontal dashed line marks the neutral value of zero, and gray shaded vertical bands with thin black outlines identify the September 11 attacks, the Global Financial Crisis, and the COVID-19 pandemic.
The two tone measures display meaningful cyclical co-movement, with a Pearson correlation of 0.49, but they differ materially in magnitude. Before the Global Financial Crisis, they generally move together through alternating easing and tightening episodes and both become more dovish around the Asian Financial Crisis, the dot-com downturn, the September 11 attacks, and the Global Financial Crisis. After the crisis, the lexicon measure remains persistently below zero, whereas the LLM measure moves back toward neutral. A similar distinction appears after the COVID-19 shock: both measures shift in a hawkish direction during the subsequent inflation and tightening episode, but the LLM measure becomes clearly positive while the lexicon score mostly becomes less negative. The comparison suggests that the LLM responds not only to the presence of hawkish and dovish terms but also to their context, qualification, and policy conditionality. Because the quarterly global averages pool individual speeches, countries that communicate more frequently exert greater influence on these raw aggregates.
Figure 2. LLM-Based Monetary Policy Tone: Selected Countries
The figure displays the LLM-based quarterly average monetary policy tone for the six countries or monetary areas with the highest speech counts in the sample. The panels are arranged in a 2×3 grid, with panels headed United States, Eurozone, and Canada in the top row and Japan, United Kingdom, and Germany in the bottom row. The series span approximately 1996–2025. In every panel, the horizontal axis reports calendar time and the vertical axis reports the quarterly average raw LLM score, with negative values denoting dovish communication and positive values denoting hawkish communication. The dark blue solid line is the country-level tone measure, the horizontal dashed line marks zero, and gray shaded bands with thin black outlines identify the September 11 attacks, the Global Financial Crisis, and the COVID-19 pandemic. The common scale, which runs approximately from −0.6 to +0.5, facilitates comparison of the direction and intensity of communicated policy tone across the six panels.
The country series reveal both synchronized global shifts and substantial national differences. Several central banks communicate more dovishly around the September 11 attacks, the Global Financial Crisis, and the COVID-19 shock, followed by a broad move toward hawkish language during the post-pandemic inflation and tightening cycle. Germany's series is comparatively hawkish over much of the sample, consistent with the Bundesbank's strong emphasis on price stability. Japan exhibits an extended dovish period during the late 1990s and early 2000s and additional dovish episodes later in the sample, reflecting its prolonged experience with low inflation and unconventional policy. The euro-area series records an additional dovish phase around the European sovereign debt crisis. Canada, the United Kingdom, and the United States also show pronounced post-pandemic hawkish shifts followed by moderation. Overall, the figure indicates that the global communication measure combines a common policy cycle with meaningful country-specific timing and amplitude.
Figure 3. Monetary Policy Tone and Stance by Country
The figure compares the LLM-based communication measure with the semi-structural model (SSM) estimate for six representative economies. The panels are arranged in a 2×3 grid, with panels headed United States, Eurozone, and Canada in the top row and Japan, United Kingdom, and Germany in the bottom row, over approximately 1996–2025. The blue solid line, read against the left-hand axis, is the quarterly LLM tone score derived from central bank speeches; positive values indicate more hawkish communication and negative values more dovish communication. The dark brown dotted line, read against the right-hand axis, is the SSM interest-rate gap, defined as the policy or shadow policy rate minus the country-specific neutral nominal rate; positive values indicate a restrictive macro-implied stance and negative values an accommodative stance. The scales differ across countries and, crucially, across the two measures. Gray shaded bands with thin black outlines mark the September 11 attacks, the Global Financial Crisis, and the COVID-19 pandemic, while horizontal zero lines indicate neutral tone or a zero interest-rate gap.
Co-movement is most apparent around major global turning points, but the two measures often differ in timing and persistence. Around the Global Financial Crisis, the communication measure generally turns dovish rapidly as central banks discuss emergency support and easing, while the SSM may continue temporarily to reflect the restrictive conditions that preceded the crisis before moving into accommodative territory. A similar sequence appears around COVID-19 and the subsequent inflation episode: communicated tone first shifts toward support and then toward tightening, whereas the macro-implied stance adjusts more gradually as policy rates, inflation, activity, and long-term yields incorporate the new environment. The clearest post-pandemic sequencing appears in Canada, the euro area, the United Kingdom, and the United States, where the LLM series turns hawkish before the SSM gap reaches its restrictive peak and begins to moderate while the SSM remains elevated. The figure therefore supports interpreting the measures as complementary: the LLM captures communicated policy orientation, while the SSM provides a slower-moving benchmark anchored in macroeconomic fundamentals.
Figure 4. Global Monetary Policy Tone: Full Sample vs Selected Countries
The figure compares two standardized dynamic factors constructed from the LLM-based country series over 1996–2025. The dark brown dotted line is the communication factor estimated from the selected eight-country sample, and the blue solid line is the corresponding factor estimated from the full country sample in the BIS database. The selected sample comprises the six economies displayed in Figure 3 plus Australia; together, these seven countries account for nearly half of the speeches in the BIS sample. The horizontal axis reports calendar time, and the vertical axis reports standardized factor values, approximately in the range of −3.1 to +3.6. Positive values indicate a common shift toward more hawkish communication and negative values a common shift toward more dovish communication. Because the factors are estimated separately from country-level observations, their weights are determined by estimated factor loadings rather than directly by each country's speech count.
The full-sample and eight-country factors co-move closely throughout the sample, especially around the September 11 attacks, the Global Financial Crisis, the COVID-19 pandemic, and the post-pandemic tightening cycle. Both factors move in a dovish direction during the major downturns and subsequently shift toward tightening as economic conditions recover. Their close correspondence indicates that the eight selected countries capture much of the common variation in central bank communication across countries, although the two factors do not match exactly. During easing episodes, the dovish movements in the eight-country factor are generally less pronounced than those in the full-sample factor, whereas its hawkish movements are often stronger during tightening episodes. This asymmetry may reflect differences in how communication-intensive central banks signal tightening and easing, but it may also arise from sample composition, estimated factor loadings, standardization, and country-specific variation. It should therefore be interpreted as a descriptive pattern rather than definitive evidence of a systematic signaling bias.
Figure 5. Global Factors – LLM Policy Tone vs. SSM Policy Stance
The figure compares the LLM communication factor with the SSM stance factor over approximately 1996–2025. Both dynamic factors are estimated from the same eight-country sample—Australia, Canada, the euro area, Germany, Japan, the United Kingdom, and the United States—and each is standardized to have a mean of zero and a standard deviation of one. The blue solid line is the LLM factor, which summarizes the common monetary policy tone conveyed through central bank speeches. The dark brown dotted line is the SSM factor, which summarizes the common variation in country-level interest-rate gaps and therefore represents the macro-implied monetary policy stance. The horizontal axis reports calendar time and the vertical axis reports standardized factor values, ranging approximately from −2.5 to +3.6. Because the two factors are standardized representations of conceptually different objects, the relevant comparison concerns their direction, timing, and persistence rather than differences in their absolute magnitudes.
The two factors display broad co-movement around several major policy turning points but do not always adjust simultaneously. The LLM factor displays greater short-term variation and, in some episodes, appears to turn before the smoother and more persistent SSM factor. The post-pandemic episode provides the clearest illustration: the communication factor rises abruptly as central banks signal tightening and begins to decline before the SSM factor reaches its later restrictive peak. This sequencing is consistent with central bank communication responding to changing policy considerations before those developments are fully reflected in the observed and latent macroeconomic variables informing the SSM. The pattern is descriptive and does not by itself establish a formal lead-lag relationship, but it suggests that the LLM factor may capture elements of forward guidance and policymakers' intended direction of monetary policy.