Central bank independence is contested, most recently by one of the two candidates in this year’s US presidential election (Weissman 2024). Economists overwhelmingly support the concept of an autonomous or independent central bank (Qanas and Sawyer 2024). Recent history – the history of the last three decades – shows that polities have been voting with their feet, as more countries have buttressed the independence of their central banks (Crowe and Meade 2008, Dincer and Eichengreen 2018).
In a new paper (Dincer et al. 2024) we ask whether this recent trend extends back further in time – whether it has historical precedent. We adopt the most widely used measure of central bank independence but apply it to a longer period than in previous research, extending the analysis back to the beginning of the 19th century. Assembling legal statutes for a long period is a considerable task, given that new statutes and amendments to existing statutes are adopted frequently. The result is an unbalanced panel since some central banks were established later than others. In all, we have a total of 8,318 annual data points for 120 central banks.
Trends and precedents
We focus on legal independence, assigning numerical ratings based on the relevant passages of the central bank’s statutes. We utilise the 16-point scale developed by Cukierman et al. (1992) since this is the metric used most widely in the literature, meaning that adopting it facilitates comparisons with other studies.
The fact that stands out most clearly from our long-run analysis, summarised in Figure 1, is the rise in central bank independence since 1980. The extent of the change and the high level of legal independence in this period in fact have no precedent in earlier history.
Figure 1 Central bank independence, decadal averages, 1800-2021
Notes: The decade averages are calculated after removing missing values at the country-year observation level. The ECB’s addition in December 1999 is represented in a dotted line. LVAU is the unweighted Cukierman et al. (1992) index that takes the simple average of eight components: chief executive officer (CEO), policy formulation, objectives, advances criterion under limits on lending to the government, securitised lending criterion under limits on lending to the government, terms of lending, potential borrowers from the bank criterion, and other criteria on the limits on lending to the government. LVAW is the weighted average version of the eight components. LVES and LVESX are two narrower indices. LVES is the weighted average of components Who formulates monetary policy, who has final word in resolution of conflict and objectives. LVESX is a weighted average between the LVES and limitations on lending - miscellaneous subcomponent under the Limitations on lending to the government component.
This fact also applies across cohorts, as shown in Figure 2. In other words, the sharp rise in central bank independence post-1980 is common to central banks in existence already in the first half of the 19th century, central banks established in the second half of the 19th century, central banks created in the first half of the 20th century, and central banks founded subsequently. These trends do not reflect changes in sample composition.
Figure 2 Central bank independence (LVAU) by period of central bank establishment
Notes: The graph shows the (decade) average value of LVAU for countries grouped by their year of establishment. Each line holds fixed the composition of central banks created before 1850, established between 1850-1900, established between 1900-1950, established between 1950-2000, established after 2000, and includes a line that plots the average for all banks in our dataset. Each of the three establishment categories includes central banks created from that year onwards.
However, we document a slower rise in average levels of independence starting in the 1920s. Following WWI, League of Nations experts and so-called money doctors working under the auspices of the Bank of England, Bank of France, and Federal Reserve System spearheaded a drive for autonomous (in contemporary parlance) central banks. As we show, the increase in the extent of independence in the 1920s and 1930s was associated with the creation of a number of new, initially relatively independent central banks. We also document a modest but ongoing rise in average levels of independence spanning the four subsequent decades.
There were interruptions, to be sure: there was a sharp fall in central bank independence during WWII for obvious reasons. But this was reversed in the 1950s. For example, the Federal Reserve was required to subordinate its interest-rate-setting policy to the dictates of the US Treasury during WWII but regained interest-rate-setting independence with the Treasury-Fed Accord in 1951. Here, we show that this movement was more general.
The upward trend in levels of independence then resumed, spanning 1950 to 1980. Again, this movement is common to all birth cohorts. This suggests that the groundwork for the post-1980 revolution in central bank independence was laid in earlier years.
Human judgement versus machine learning
Assigning numerical ratings for various dimensions of legal independence based on language in central bank statutes involves human judgment in interpreting the relevant passages. We therefore compare our findings with results obtained using natural language processing and machine learning techniques. We analyse a sample of the 90 most recent central bank acts using machine-vector regression and topic modelling.
The goal of our natural language processing approach is to determine which topics have predictive power for our index of legal independence and whether they predict higher or lower values. To determine which terms that have predictive power, we first tokenised the texts using a pre-processing routine that excludes stop words, rare words (on the grounds that infrequent words add more noise than signal), punctuation, numbers, country names, month names, and extra spaces. We grouped together (‘lemmatized’) different inflected forms of the remaining terms and kept for further analysis the top 2,000 most frequent unique tokens across all statutes.
The next step consists of topic modelling the entire corpus of central bank statutes to classify terms contributing to central bank independence. This is performed using the latent dirichlet allocation (LDA) algorithm. We choose an optimal number of topics by comparing the prediction capacity of each model and selecting that with the best performance achieved with the minimum number of topics. Here, the optimal number of topics is 11. Table 1 shows the list of words that constitute topics when applying the LDA method. We then assign a label to each group of words or topic based on the IMF’s Central Bank Legislation Database categories, in which articles and extracts of the statutes are classified. Readers will note that one of these categories (‘Objectives of Microprudential Supervision’) appears twice since two separate groups of tokens have multiple references (such as supervision, financial, institutions, establishment, branches, capital) that are plausibly grouped under this heading.
Table 1 Topics using terms from the corpus of central bank statutes
(tokens with highest topic weightage, starting with highest weightage)
Notes: This table displays the list of topics generated by a topic modelling exercise employing the Latent Dirichlet Allocation algorithm on a corpus of words comprised of central bank statute datasets from the repository of IMF Central Bank Legislation. Column (1) displays labels assigned to the topics, which were selected from the categories of the IMF's Central Bank Legislation Database, in which articles and excerpts of the statutes are classified. Column (2) lists the subjects and words from the corpus of central bank statutes that correspond to each subject. We list the words that contribute the most to each topic (this is not an extensive list). As described in the paper, terms comprising each topic were estimated using the Latent Dirichlet Allocation algorithm.
The topic with the largest positive contribution to explaining variations in central bank independence is related to disclosure, transparency, and reporting obligations. This is consistent with the presumption that independence and accountability go together and that transparency is a mechanism by which central banks provide information on their operations so that they can be held accountable by politicians and the public. The topic with the largest negative contribution has to do with regulatory powers over, inter alia, securities markets. These powers complicate the central bank’s mandate, make accountability more difficult, and thereby render independence more problematic.
Conclusion
We have taken two approaches to adding to the already large literature gauging central bank independence: a historical analysis of central bank statutes examined using judgmental techniques and text analysis using natural language and machine-learning methods. The first approach enables us to document the sharp rise in central bank independence starting in the early 1980s and to identify it as historically unique. In addition, however, we document a slower rise in independence starting in the 1920s. As we show, the increase in the extent of independence in the 1920s and 1930s was associated with the creation of a number of new, relatively independent central banks. We also document an ongoing increase in average levels of independence spanning the four subsequent decades.
The second approach identifies terms and topics that are positively and negatively associated with central bank independence, as captured by the first approach. The topic with the largest positive contribution to explaining the variation in central bank independence focuses on disclosure, transparency, and reporting obligations. This is consistent with the presumption that independence and accountability go together and that transparency is a mechanism by which central banks provide information on their operations such that these institutions can be held suitably accountable by politicians and the public. The topic with the largest negative contribution concerns regulatory powers over, inter alia, securities markets, powers that complicate the central bank’s mandate and render accountability more difficult, thereby making independence more problematic.
This survey has sought to demonstrate that the empirical analysis of central bank independence remains a fruitful area for study. Future work might seek to supplement measures of de jure independence, as analysed here, with measures of de facto independence, as in Binder (2021), but extend the latter backward in time. Furthermore, the application of machine-learning methods, similar to those outlined in this study, could be extended to analyse various types of central bank materials, including reports, press releases, conference transcripts, speeches, minutes, and transcripts. Despite these advances, there is still much ground to cover in this area, with ample opportunities for further exploration and analysis.
References
Binder, C (2021), “Political Pressure on Central Banks”, Journal of Money, Credit and Banking 53: 715-744.
Crowe, C and E Meade (2008), “Central Bank Independence and Transparency: Evolution and Effectiveness”, IMF Working Paper WP/08/119.
Cukierman, A, S Webb and B Neyapti (1992), “Measuring the Independence of Central Banks and its Effects on Policy Outcomes”, World Bank Economic Review 6: 353-398.
Dincer, N and B Eichengreen (2018), “Central Bank Transparency and Independence: Updates and New Measures”, International Journal of Central Banking 10: 189-253.
Dincer, N, B Eichengreen and J Martinez (2024), “Central Bank Independence: Views from History and Machine Learning”, Annual Review of Economics 16: 393-428.
Qanas, J and M Sawyer (2022), “’Independence’ of Central Banks and the Political Economy of Monetary Policy”, Review of Political Economy 36: 565-580.
Weissman, J (2024), “Could Donald Trump Break the Fed?”, The Atlantic, 21 August.
