A FORWARD GUIDANCE INDICATOR FOR THE SOUTH AFRICAN RESERVE BANK: IMPLEMENTING A TEXT ANALYSIS ALGORITHM

† We would like to extend our gratitude to Lars Christensen and Laurids Rising for sharing their generic sentiment library and for their valuable input in the early stages of this project. We would like to also thank all the reviewers (official and unofficial) for their valuable comments throughout the research and writing process of this project.

Abstract

The expansion of central bank communications and the increased use thereof as a policy tool to manage expectations have led to an area of research, semantic modelling, that analyses the words and phrases used by central banks. We use text-mining and text-analysis techniques on South African Reserve Bank monetary policy committee statements to construct an index measuring the stance of monetary policy: a forward guidance indicator (FGI). We show that, after controlling for market expectations, FGIs provide significant explanatory power for future changes in the repurchase interest rate (the primary monetary policy instrument). Their out-of-sample predictive power is, however, weak. Furthermore, we show that FGIs are primarily driven by inflation expectations, which highlights the strong link between the SARB’s communication strategy and its inflation targeting mandate. In fact, we observe a systematic anti-inflation bias in the communicated stance of monetary policy—both absolutely and asymmetrically. Overall, Monetary Policy Committee (MPC) statements reflect relevant information on the inflationary stance and policy decisions of the South African Reserve Bank (SARB), but, since forecasts are conditional on current information, they provide unreliable forward guidance. Given this finding, MPC statements should emphasize the conditional nature of the SARB’s stance, and what that implies for the future path of the policy rate.

Notes

† We would like to extend our gratitude to Lars Christensen and Laurids Rising for sharing their generic sentiment library and for their valuable input in the early stages of this project. We would like to also thank all the reviewers (official and unofficial) for their valuable comments throughout the research and writing process of this project.

1 In fact, over the last few decades, all central banks have changed how they communicate to financial markets and the broader public. Historically, central banks chose to keep public announcements to a minimum because of the perceived benefits to keeping markets guessing (Mishkin, 2004: 1). However, since the 1990’s there has been a movement towards greater transparency and openness (Stein, 2014: 1). A primary argument for greater transparency is the notion that independent central banks should be more accountable to the public (Blinder, Ehrmann, Fratzscher, de Haan & Jansen, 2008: 912). In doing so, central banks place substantial emphasis on improving their communication to enhance the public’s confidence in the bank’s ability to adhere to its mandate (Weidmann, 2018: 2). This objective has largely been facilitated by way of more timely release of meeting minutes, more speeches by central bankers, embracing new communication channels on social media, increasing the scope and frequency of economic projections, introduction of post-meeting news conferences, and more news conferences (Stein, 2014: 2; Shin, 2017: 1).

2 Forward guidance is not confined to information about future interest rate trajectories, but it embodies all information about future monetary policy decisions (Weidmann, 2018: 5).

3 The form of forward guidance differs amongst central banks and can materialize in a qualitative or quantitative manner. Quantitative forward guidance includes the explicit publishing of expected future policy rates, whereas qualitative forward guidance provides suggestive (non-numerical) forecasts of policy rates communicated through the content of monetary policy statements and other material generated by central banks.

4 Rosa and Verga (2007) formulated a glossary of words and phrases, which served as a guide to establish their index. Following Rosa and Verga (2007) and Ehrmann and Fratzscher (2007), Reid and Du Plessis (2010) subjectively construct a discrete index of monetary policy “inclination” (i.e., the likelihood of a policy change), based on the information content of the SARB’s monetary policy statements that accompany each monetary policy committee (MPC) meeting. The primary objective of Reid and Du Plessis (2010) is to assess how successful the SARB’s monetary policy committee has been in communicating to the public its policy since adopting an inflation targeting framework. As such, the index serves as an analytical tool to analyse the consistency of the SARB’s communication. Other related papers include: Jansen and de Haan (2005); Musard-Gies (2006); Gerlach (2007); and Berger, de Haan and Sturm (2011). A detailed overview of the literature on central bank communication, forward guidance, and quantitative measures of forward guidance can be found in the thesis version of this article at: http://hdl.handle.net/10019.1/107163.

5 Some of the most pre-eminent techniques include boolean and dictionary, latent semantic analysis, latent Dirichlet allocation, and descending hierarchical classification. Bholat, Hans, Santos and Schonhardt-Bailey (2015) provide a succinct discussion of these techniques – specifically in the context of central bank research. There has also been growing interest in supervised machine learning methods such as support vector machines (e.g., Tobback, Nardelli & Martens (2017)) and Naïve Bayes (e.g., Moniz & de Jong (2014)) classifiers to construct sentiment indices such as FGIs.

6 Text-mining is a blanket term for a series of computational tools and statistical techniques that quantify text (Bholat et al., 2015: 1). Text mining is also commonly referred to as computational linguistics or natural language processing.

7 We use a weekly measurement since it circumvents the need to work with months that vary in duration, which introduce inconsistencies. The various meeting dates and the associated inter-meeting intervals are presented in Table B.1 in the Technical Appendix.

8 The MPC documents prior to 18/07/2013 were only available in a non-readable PDF format. This problem was addressed by using Adobe’s “Acrobat Pro DC” software to generate readable renditions of the documents.

9 This example was taken from Bholat et al. (2015).

10 Stemming, which entails “cutting” off affixes and counting stems, is also widely used in practice. For example, the word “banking” contains the stem “bank” and the affix “-ing”, hence the two words would (after stemming) be considered two instances of the same token. We do not apply this technique since: (1) it can result in errors (overstemming and understemming); and, (2) it can prove difficult to infer sentiment from a particular stem that has been derived from a set of distinct words which exhibit contrasting sentiment.

11 Although case folding sometimes obscure the meaning of proper nouns, this does not pose a problem for our analysis. Misleading occurrences of case folding can occur, but the adoption of multi-word tokens largely prevents this from occurring.

12 See Tables C.1 and C.2 in the Technical Appendix for the list of words contained in the Henry and Christensen dictionaries. Owing to the scope of the Loughran dictionary we do not provide a detailed list thereof, however, it is available at: https://sraf.nd.edu/textual-analysis/resources/#LM%20Sentiment%20Word%20Lists.

13 An advantage of using a continuous FGI emanates from its ability to accommodate both marginal and acute changes in sentiment, whereas a discrete FGI can merely accommodate acute changes.

14 The variance of repo rate changes has become smaller in recent times. Therefore, we provide results for the full sample as well as a reduced sample that corresponds to that used by Reid and Du Plessis (2010) in order to make results directly comparable.

15 Supplementary exploratory data analyses are provided in the Technical Appendix.

16 Figure D.1 in the Technical Appendix shows that FGI (Henry) fell persistently below zero only from 2008 to 2010: the period of the Great Recession.

17 The regression results for FGI (Loughran) are available upon request.

18 In the case of the Christensen FGI, the consumer confidence coefficient is significant, but also very small. ¹⁹ For the Christensen FGI, the coefficient is significant and equal to 0.19.

19 A 2-year interest rate swap is used since it is slightly less volatile compared to the 1-year interest rate swap.

20 For FGI (Christensen), a unit increase in the FGI index predicts a 0.2 pp increase in the repo rate within the next two months and a 0.4 pp increase within the next 6 months.