The power of words: A content analytical approach examining whether central bank speeches become financial news

Abstract

Few studies have examined the impact that central bank indirect intervention has on exchange rates. Efficient market theory predicts that new information within central bank communication will become a component of information used by currency traders. This study applies a novel methodology to examine whether information contained within Bank of Canada and the Reserve Bank of Australia communications does in fact get embedded within the information reported on the financial newswires. The primary data are speeches that are made public by the two central banks and from news as reported by Reuters from 1995 to 2009. Applying content analysis and an innovative use of information science theoretic measures, we demonstrate the flow-through of information contained within central bank communications to the information set used by traders.

Keywords

Central bank speeches content analysis foreign exchange information theoretic measures text mining

1. Introduction

The exchange rate is a fundamental macroeconomic variable within any economy influencing, and being influenced by, key macroeconomic variables, including inflation and interest rates [1]. Excessive volatility of the exchange rate has the potential to create serious price distortions that can perpetuate a chain-reaction of structural inefficiencies that can spread throughout the economy. To mitigate against this, central banks periodically intervene in foreign currency markets with a view to stabilizing the exchange rate. They do this in conjunction with their other monetary policies which impact on other macro-economic variables, like interest rates and money supply, so that the overall financial system is nudged towards equilibrium. Traditionally, central banks have used ‘direct intervention’ (DI) to influence the exchange rate. They do this by buying or selling foreign currencies in the open market. However, since the mid-1990s many central banks appear to have moved towards ‘indirect intervention’, mostly via oral communication through speeches delivered by key central bank officials. The extent to which indirect intervention is effective in stabilizing the exchange rate depends on two underlying factors – whether the foreign exchange market receives the conveyed information and whether that received information actually exerts an influence over the level and volatility of the exchange rate. It is important to alert the reader at this point that the purpose of this paper is to examine only the first factor, that is, whether or not key information embedded within central bank communications actually flows through to the information set that traders use in their buy/sell decisions. We are not examining the second factor – whether or not the information that flows through then is actually used by traders, thereby exerting an influence on the level and volatility of the exchange rate. This will be the subject matter of another follow-up paper. Also the goal is not ‘testing’ the significance of the impact of underlying macroeconomic variables on exchange rates but rather examining the efficacy of indirect intervention as a central bank policy in managing exchange rate stability under a certain economic condition (i.e. for given ‘levels’ of the other key variables).

At this stage it may be of some benefit to the reader if we quickly revisit the principal roles of central banks in a nation’s economy. The central bank is typically seen as the ‘lender of last resort’ and the highest monetary agent of the government which is responsible for executing government monetary policies. It is the primary prerogative of the central bank to maintain full employment in the economy and stimulate economic growth. It is also the responsibility of the central bank to stabilize the nation’s currency as excessive volatility increases risk. The central bank is also the gatekeeper of a nation’s money supply as it is the only institution empowered to issue bank notes and maintain the nation’s gold and foreign currency reserves. In this regard the roles of the Reserve Bank of Australia (RBA) and the Bank of Canada (BOC), being the central banking institutions in their respective nations (i.e. Australia and Canada), are not dissimilar.

In our paper we have considered the speeches of the BOC and the RBA and have examined how their information content flows through to the relevant financial newswires. The primary reason we have focused on Australia and Canada is because they are both small, open economies that are socio-economically similar but are geographically far apart and also engaged in different markets. Both these countries have a long tradition of free-floating exchange rates and have similar economic systems. However, the weighted importance of their trading partners is significantly different, and furthermore, their respective currencies are not used as international reserve currencies.

Our work draws inspiration from prior research, notably that carried out by Beine et al. [2] and Fratzscher [3 –5]. However, we take an entirely different approach and use an information science methodology rather novel to the field of financial research, which makes our present work stand out from the previous ones in terms of its unique contribution to the relevant literature. Not only do we demonstrate the flow-through of information embedded in the central bank communications to the information set available to traders in an efficient foreign exchange market, but in doing so we also prove the effectiveness of information science methods and tools as a financial research methodology.

Indeed, a key contribution of our paper is the successful application of information science tools to answer an important question that arises within the realms of mainstream finance. We use an innovative methodology of tracking speeches from the ‘source’ (the central bank) to the ‘sink’ (the information set that is common knowledge to currency traders in an efficient foreign exchange (FX) market). Our approach involves the use of Leximancer, a state-of-the-art text mining software, to detect hidden patterns in the central bank speeches and newswires published in days surrounding the speeches, and the computation and application of key information theoretic measures to determine the similarity or lack thereof between the information contents of central bank speeches and what is being reported by the newswires.

2. Background material

In this section we review the literature on communication policy and financial markets in general, and its effects on the exchange rate in particular. It is necessary to present a macro view of the impact of communication on financial markets as it paves the way for a micro analysis aimed at assessing the impact of communication on the FX market. The case for using communication as a policy tool for managing exchange rates is based on three factors: (1) the failure of DI to influence the level and volatility of the exchange rate; (2) the costly nature of DI; and (3) the growing awareness of the value of oral intervention, as demonstrated by a number of empirical studies [2 –7].

2.1. Communication policy and financial markets

Communication is a powerful tool that can influence markets [8]. Bernanke [9] argues that effective central bank communication encourages financial market efficiency and has positive effects on macroeconomic performance. Its usefulness in financial markets is based on the extent it can move asset prices. Research examining the effect of oral intervention to this point is rather limited [2, 3]. Research on the Federal Reserve Bank of America (FRB), the Bank of England (BOE) and the European Central Bank (ECB) confirm that communication does influence markets [10]. This research indicates that communication by central bank authorities in the USA and the UK does have an impact on financial markets; however, not all components of communication have the same effect. In both the USA and the UK, routine speeches have little impact on markets. In the USA it is the Federal Open Market Committee statements and the chairman of the FRB’s congressional testimonies [10] that have the most pronounced effect. In the UK, while the speeches and testimonies to parliamentary committees have a small impact, the BOE Minutes and the Inflation Report are much more influential [11]. Studying the relatively small Swiss economy, Ranaldo and Rossi [12] found that monetary policy announcements, speeches and interviews by the Swiss National Bank were the factors that most significantly impacted on asset values.

This research examines whether communication from the central banks flows into the information set of traders. This is a fundamental research question which earlier research has either assumed or ignored. If central banks cannot get their messages through to the market in the first place, then the question of whether or not these messages can influence the direction of exchange rate movements becomes irrelevant! The next section focuses on the extant research examining communication policy as it specifically relates to the FX markets.

2.2. Communication policy and the exchange rate

Major central banks such as the FRB and the ECB believe that oral intervention can influence exchange rates [3, 5]. Signalling from the central bank in terms of future monetary policy intentions or planned interventions in the FX market is what potentially impacts the exchange rates [2]. When the central bank releases information, it intends for sufficient signals to be generated in order to induce market participants to act in a certain way.

Studies like Fratzscher [3 –5, 13], Park and Song [7], Beine et al. [2] and Bernal and Gnabo [6] have explored oral intervention and its impact on the exchange rate. Findings suggest that oral intervention is effective in influencing the level of the exchange rate and reducing its volatility. The dampening of volatility is in sharp contrast to the effect of DI, which has been found to exacerbate volatility [14].

Fratzscher [3, 5, 13] examined the US$/Euro and Yen/US$exchange rates and found that communication by central banks is a useful tool to manage short-term exchange rate volatility and that it also moved the exchange rates in a desired direction [5, 13]. Communication can be used, therefore, as a stand-alone policy tool available for use in the short term. Fratzscher [4] further highlighted that G3 economies are now using communication as the primary tool in managing exchange rates. An independent study by Park and Song [7] on the Yen/US$exchange rate using data from 2000 to 2003, also found that oral intervention was effective when ‘leaning against the wind’, that is, in countering the exchange rate volatility when it is moving away from fundamental equilibrium.

The extant research therefore indicates that oral intervention has a place within the policy toolkit of central banks.

3. Data

This study uses two data sources – primary data in the form of actual speeches delivered by central bank officials, and secondary data derived via content analysis of those speeches using the Leximancer software. The primary data was extracted from two sources – speeches from the websites of the BOC and the RBA and the Reuters newswires queried from the Factiva database maintained by the National Library of Australia. Secondary data was derived from the co-occurrence matrix, which is a Leximancer output. Total number of speeches analysed is different between the two central banks. This is because we extracted speeches for a pre-defined period, from 1995 to 2009, but central bank officials actually make the speeches as and when they see the need. Commencing with 263 BOC speeches and 273 RBA speeches, a data cleaning process resulted in 38 BOC speeches and 15 RBA speeches. We removed the speeches in which there were three or fewer common concepts between the speeches and their corresponding Reuters newswires, as it would otherwise be difficult to calculate meaningful information theoretic measures in such situations. The retained speeches are shown in Tables 1 and 2.

Table 1.

Summary of 38 BOC speeches used in information analytics.

Year	Speech	Date	Year	Speech	Date
1998	Speech 4	23 September 1998	2005	Speech 2	17 February 2005
1998	Speech 5	15 October 1998	2005	Speech 3	21 March 2005
1998	Speech 6	7 December 1998	2005	Speech 4	30 March 2005
1999	Speech 1	20 January 1999	2007	Speech 31	5 December 2007
1999	Speech 5	4 May 1999	2008	Speech 2	24 January 2008
2000	Speech 1	27 January 2000	2008	Speech 4	18 February 2008
2000	Speech 2	9 March 2000	2008	Speech 17	17 July 2008
2000	Speech 7	14 September 2000	2008	Speech 20	25 September 2008
2000	Speech 9	4 December 2000	2008	Speech 21	23 October 2008
2001	Speech 6	20 April 2001	2008	Speech 26	17 December 2008
2001	Speech 7	1 May 2001	2009	Speech 2	22 January 2009
2001	Speech 8	26 June 2001	2009	Speech 4	10 February 2009
2002	Speech 4	28 January 2002	2009	Speech 8	30 March 2009
2003	Speech 2	29 January 2003	2009	Speech 15	18 June 2009
2003	Speech 13	12 June 2003	2009	Speech 20	25 August 2009
2004	Speech 2	11 February 2004	2009	Speech 22	23 September 2009
2004	Speech 13	16 June 2004	2009	Speech 23	28 September 2009
2004	Speech 15	22 July 2004	2009	Speech 30	19 November 2009
2004	Speech 16	20 September 2004	2009	Speech 31	16 December 2009

Table 2.

Summary of 15 RBA speeches used in information analytics.

Year	Speech	Date	Year	Speech	Date
1995	Speech 9	19 October 1995	2004	Speech 11	16 November 2004
1997	Speech 5	15 May 1997	2005	Speech 6	8 April 2005
1997	Speech 7	12 August 1997	2005	Speech 14	11 October 2005
1997	Speech 14	4 December 1997	2005	Speech 16	19 October 2005
1997	Speech 15	4 December 1997	2008	Speech 26	31 October 2008
1998	Speech 3	12 March 1998	2008	Speech 27	19 November 2008
1998	Speech 10	16 July 1998	2008	Speech 28	19 November 2008

Note that the numbers were assigned to each speech by the authors of this paper.

We chose the start of the sample period such that it was in line with a discernible shift towards oral intervention by major central banks. This study focuses on Reuters newswire as it is the one most commonly used [15] by market participants. Oberlechner and Hocking [16] in their study found that foreign exchange traders relied heavily on financial newswires as their most important source of information. We did not pre-select specific newswires to include, but rather included all those captured using a specified criterion. We did not choose to actively ‘filter out’ non-financial news simply because Reuters newswires concentrate on financial news anyway, so the presence of non-financial news was assumed negligible.

For each speech we generated a separate folder containing the newswires, covering a four-day window from one day before the speech date to two days after the speech. We did this in order to track whether there were any newswires that emanated from the speech. One day before is included to account for the possibility of information leakage. We adopted this conservative approach because there have been incidences of information leakage reported in the past. For example, the RBA prematurely released market-sensitive monetary policy information on 2 February 2000, which resulted in subsequent reviewing of its actions [17]. While applying prudence, we realize this approach will result in a downward bias on our results since one day out of the four day window actually represents the day before the actual speech. However, we did this in recognition of the fact that information leakage does occur.

To demonstrate our empirical approach, the 1998 Speech 4 of BOC is used as an illustrative example and Section 5 shows the information analytics output from that speech. Results from other speeches analysed are summarized in Section 7.

4. Methodology

This research applies a two-pronged methodology. The first is the use of text mining, applying content analysis to identify the themes and concepts in central bank speeches and their corresponding newswires. We follow this up with determination of the modified Kullback Leibler Divergence (KLD) [18, 19] and the Cumulative Shannon Differential (ShD) [18], two of the information theoretic measures of entropy. These are measures of distance between two distributions: P, the conditional distribution, and Q,– the approximate distribution. The closer or more similar P is to Q, the lower the values of the KLD and the ShD. Having determined cut-off values for KLD and ShD, it is possible to make a determination of the similarity or lack thereof between the contents of the central bank speech and the newswire reports.

4.1. Text mining

Ghazinoory et al. [20] define text mining as a process of extracting useful information from data, involving the identification and detection of unpredictable and significant patterns. Pons-Porrata et al. [21] state that text mining involves extracting non-trivial information and knowledge from unstructured text. The goal is therefore to detect previously unknown information [20], implying that the documents containing the data must be unstructured. According to Fayyad [22], text mining is one of the processes of knowledge discovery to identify valid, novel and potentially useful patterns in data. We use clustering methodology (in particular, the latent semantic approach), deemed suitable as it allows an unsupervised discovery of concepts that are hidden in the text [23]. It is an appropriate methodology to use as it does not require the data to be pre-classified and makes no assumptions regarding its structure. The latent sematic approach uses a nonlinear machine learning algorithm based on Bayesian analysis of word co-occurrences to perform the content analysis.

4.1.1. Content analysis

Content analysis is a research methodology used in the summarization and categorization of documents and text. Stemler [24] views it as a systematic and replicable approach of reducing text in documents to a few categories. The categorization of documents can be interpreted as a form of pattern recognition [25]. The Leximancer software identifies ‘concepts’ in the text, which it defines as collection of words that travel together in the text [26]. It objectively assigns theme names to concepts that travel together within the document and shows them as theme circles on concept maps [26]. The concept map schematically charts the themes and their defining concepts, placing them in proximity to each other according to their connectedness [26] and showing to what extent they are semantically linked.

The themes are significant in that they represent the message that is contained in the central bank speeches or the central issues in the newswires. If the message in the central bank speeches is flowing to the news stream, the same or similar themes would be present in the newswires. Thus we aimed to first identify the themes and related concepts in the source data, which are the speeches, and in the newswires. Having done that, we then analysed whether the themes and related concepts were similar or different.

In order to perform the above we utilized information summarized on the concept maps, and the co-occurrence matrices. The matrix is made up of the number of times (frequency) a particular concept occurs when a given concept occurs (co-occurrence). For instance, when the concept ‘economy’ occurs, how many times does the concept ‘dollar’ occur with it? A comparison of the co-occurrence matrices from the speech and the corresponding newswires is necessary to identify how close the concepts are within the two sets of documents.

4.1.2. Justification for the use of Leximancer

The choice of Leximancer in this study stems from its relative stability and repeatability [27]. This means that, with the same document and the same settings or parameters in the processing stage, it will generate consistent results. Rooney [28] attributes the stability to a high level of coding stability given that Leximancer will perform the same deterministic machine learning phase despite the number of times it is run if the same settings are maintained. Kabanoff and Brown [29] describe Leximancer as operating at almost the full end of the automated scale of computer-aided text analysis programs.

4.2. Information theory measures

Information theoretic measures have been widely used in research, for example, in the approximation of probability distributions, signal processing and pattern recognition [30]. Their central focus is the determination of how similar or divergent distributions or variables are. Thus these measures have wide applicability in the statistical and mathematical fields in areas where signal analysis [30] is critical, such as radar, transport, and in the digital field where image analysis is important [31]. These measures are usually termed distance measures as they give an indication of how far variables could be from each other. In this paper the calculated measures are referred either as divergence or as similarity measures interchangeably.

In Hoffmann et al. [32] information theoretic measures are used to identify similarity between frequency of speech signals. Gokcay and Principe [33] use them in clustering data as a means of pattern recognition. Since information theoretic statistics determine information content in data, according to Chandola et al. [34], they could be used for anomaly detection, which is the identification of nonconforming patterns in data.

Information theoretic measures are attractive as they are easy to understand, general in nature and simple to use [35]. They also provide an objective alternative to Bayesian statistics approaches in performing inferential analysis, enabling the determination of differences between parameters without using techniques requiring hypothesis testing [35]. The data extracted from the common concepts of the speeches and the newswires do not lend themselves to the use of Bayesian statistics, given that the distributions vary from speech to speech and from newswire to newswire. While the extent of similarity or otherwise between the datasets could have been gauged to an extent by using one of Leximancer’s in-built functionalities, the information theoretic tools would yield a more robust, quantitative measure of similarity which can be better used as a ‘benchmark’ in subsequent follow-up studies. This provides a further justification for using measures developed within the information sciences. Specifically, we chose to use the KLD and ShD to measure the divergence between the concepts in the speech and in the newswires [18]. The two measures are related but have different formulations. The KLD has been used in many studies, such as ecological studies [35], in the detection of copying in a multiple choice examination [36] and in studies involving information retrieval systems [37], among other uses.

The formula for KLD [18, 19] is:

KLD = 0.5 \sum P_{k} \log_{2} P_{k} / Q_{k} + 0.5 \sum Q_{k} \log_{2} Q_{k} / P_{k}

(1)

The formula for ShD [18] is:

ShD = \sum | {\sum P}_{k} \log_{2} P_{k} - {\sum Q}_{k} \log_{2} Q_{k} |

(2)

where P is the probability of each of the values of the distribution P, and Q is the probability of occurring of each of the values of the distribution Q. The inclusion of log base 2 is because the KLD and ShD are in units of bits. This transformation is in line with the convention of information theory [38]. The two measures have been calculated to two decimal places as in Alguliev et al. [39], where the calculated Jaccard, Overlap and Normalized Google Distance measures were calculated correct to two decimal places.

In order to test for similarity between the central bank speech and its corresponding newswires, we compared document representations as demonstrated by Egghe and Rousseau [40]. The document representations are the common concepts between the speech and its newswire equivalent. These provide data for distributions P and Q. The smaller the values of KLD and ShD, the closer or the more similar distribution P is to Q. A value of 0 indicates that they are identical. The KLD is non-negative and takes the value zero when two distributions are identical [41]. Since the two measures have no upper bounds, it is necessary to determine an upper cut-off value. With such a figure, a calculated similarity figure that is above the cut-off would suggest dissimilarity. Determination of the cut-off values is discussed in Section 6.

We noticed that in some cases, even though the calculated KLD and ShD are based on the same distributions, their values can result in conflicting conclusions. This occurs when the KLD value may indicate similarity whereas the ShD indicates divergence. In such cases we calculated a third measure, the chi-square (ChiSQ). Its purpose is to provide an alternative measure when KLD and ShD are inconclusive. For instance, if the magnitude of the ChiSQ is relatively low and this is similar to the KLD value, then the overall conclusion would be the two distributions are close. It must be noted that, while the ChiSQ is usually used as a parametric measure for data extracted from an underlying distribution, in this study it is being used simply as a ‘tie-breaker’, and is not used in any parametric test for significance of the calculated values.

5. Empirical findings

In this section we discuss empirical findings from the quantitative and qualitative output of Leximancer. While many speeches were analysed, the empirical illustration is based on 1998 Speech 4 BOC and its corresponding newswires.

5.1. Illustration using 1998 Speech 4 BOC and its newswires

In 1998 Speech 4 BOC, Leximancer identified seven key themes: Canadian, dollar, economy, exchange, currency, international and crisis. In the corresponding newswires, the main themes are Canadian, currency, economy, market, developments, international and crisis. The themes for the speech are schematically represented on the concept map as in Figure 1, in which the ones that are directly connected to the theme ‘Canadian’ are ‘economy’, ‘exchange’ and ‘dollar’.For the corresponding newswires of 1998 Speech 4 BOC, the themes identified by Leximancer are shown in Figure 2.

Figure 1.

Concept map for 1998 Speech 4 BOC (adapted from concept map drawn by Leximancer).

Figure 2.

Concept map for 1998 Reuters newswires for Speech 4 BOC (adapted from concept map drawn by Leximancer).

The fact that the bulk of the themes in the 1998 Speech 4 BOC and its newswires are common is the visual indicator of some similarity between the two. Structured analysis is required because themes can be assigned a common name by Leximancer in the two documents but represent different concepts. For instance, the word ‘economy’ could mean frugality on an individual basis in one document, while in another it could be referring to the economy of a country. This is why the calculation of similarity or divergence values is necessary. These measure the closeness, or lack thereof, of the content within the speech and the newswires. The data used for the calculation of measures of similarity is extracted from the co-occurrences matrices of concepts that make up the themes – we focus on common concepts between the speech and the corresponding newswires. Tables 3 and 4 show co-occurrence matrices of the common concepts between 1998 speech 4 BOC and its newswires. The common concepts are Canada, economy, currency, dollar, rates, international and crisis.

Table 3.

Co-occurrence matrix of common concepts in 1998 Speech 4 BOC.

Concepts	Canada	Economy	Currency	Dollar	Rates	International	Crisis
Canada	30	12	6	15	8	4	3
Economy	12	25	6	2	4	6	2
Currency	6	6	19	7	4	1	2
Dollar	15	2	7	20	3	1	2
Rates	8	4	4	3	10	0	1
International	4	6	1	1	0	12	3
Crisis	3	2	2	2	1	3	9

Table 4.

Co-occurrence matrix of common concepts in Reuters newswires for 1998 Speech 4 BOC.

Concepts	Canada	Economy	Currency	Dollar	Rates	International	Crisis
Canada	35	12	13	19	8	3	3
Economy	12	21	6	4	2	3	1
Currency	13	6	21	8	6	1	2
Dollar	19	4	8	22	4	1	2
Rates	8	2	6	4	12	1	1
International	3	3	1	1	1	13	2
Crisis	3	1	2	2	1	2	9

6. Analysis of data

This section analyses co-occurrence data. The main focus is on relational analysis which is based on the co-occurrence matrices of common concepts between the central bank speech and its newswires. If the BOC speech and its corresponding newswires contain the exact replica of the speech, then the central bank speech has been captured in its entirety. In such a case there would be 100% similarity or 0% difference or divergence between the two sets of documents.

6.1. The similarity measures

The KLD and ShD are calculated based on distributions extracted from the matrices of common concepts in Tables 3 and 4. Instead of finding the correlation between the two matrices as a whole, each concept is taken in turn. For instance, if the concept ‘economy’ is the subject of analysis, its co-occurrence with the other concepts, that is, ‘Canada’, ‘currency’, ‘dollar’, ‘rates’, ‘international’ and ‘crisis’, is taken. This data from the BOC forms distribution P. The equivalent distribution from the newswires is Q. Thus, P = [12;6;2;4;6;2] and Q = [12;6;4;2;3;1] for 1998 BOC Speech 4 and its corresponding newswires.

The objective is to determine the extent of similarity or divergence of the two distributions. In essence we are determining similarity between the pairs of common concepts [42], but we are doing it collectively, pitting one distribution against another. The calculation of similarity measures is one of the approaches recommended by Fu et al. [43] in a content-based similarity measure article. The measure is based on a number of common terms in two information sets – in this particular case, on the number of common terms between query 1 and query 2. In our study similarity is determined by taking each concept in turn and calculating the KLD and ShD values. In the cases where the KLD and ShD provide mixed results, the chi-square is used as a tie breaker.

To be able to decide on whether or not distribution P is close to distribution Q, cut-off values of KLD and ShD were determined. KLD and ShD values from both the BOC and RBA calculations were aggregated in order to form a distribution from which the upper bounds could be estimated. The KLD value was capped at 0.25, that of ShD at 0.22 and the ChiSQ at 0.42. Low values of KLD indicate closeness of two distributions whereas large values indicate a wider separation [41]. Any values above these cut-off points would lead to a conclusion of divergence between the distributions. Following are the procedures undertaken in the calculation of the cut-off values.

6.2. Determination of the cut-off values of KLD and ShD

The cut-off values are based on the mean from each of the distributions plus one standard deviation. This is appropriate if the distribution is normal. Following suggestions by Osborne [44], data transformations were applied to the statistics. The transformation process commenced with anchoring the distributions at 1 by first removing zeros and then dividing each value by the minimum value. The resulting distribution was termed ‘Adj XXXX’. For instance, for the KLD, the ‘Adj KLD’ distribution was created. On the KLD, the eighth root of ‘AdjKLD’ distribution was determined while the fourth root transformation was applied to the ‘AdjShD’ values and the natural logarithm transformation to the ‘AdjChiSQ. The transformed distributions were then tested for normality using the Kolmogorov–Smirnov and the Shapiro–Wilk tests [45]. While the log transformation is deemed special and has some advantages over other transformations [46], in this study it produced normality in the ChiSQ distribution only. Table 5 shows the results of the normality tests.

Table 5.

Tests of normality.

Similarity measure	Transformation performed	Kolmogorov–Smirnov			Shapiro–Wilk
Similarity measure	Transformation performed	Statistic	Degrees of freedom	Significance	Statistic	Degrees of freedom	Significance
KLD	Eighth root	0.045	252	0.200	0.993	252	0.307
ShD	Fourth root	0.034	251	0.200	0.991	251	0.111
ChiSQ	Natural log	0.044	252	0.200	0.990	252	0.068

Since the p-values (i.e. the significance values) are greater than 0.05, this indicates normality [45]. The results did not provide enough evidence to reject the null hypothesis, which assumes that there is no difference between the distribution in question and the standard normal distribution. However, the normality test results combined with assessing the QQ plots by eye gave confidence for normality. Following on from these results we proceeded to calculate the cut-off values for KLD, ShD and ChiSQ (Tables 6 –8). It must be noted that the figures used in the determination of cut-off values were to three decimal places following Bigi [47] and Dragalin et al. [48].

Table 6.

Determination of cut-off values of KLD.

Minimum value of KLD distribution	0.001
Mean of eighth root of Adjusted KLD	1.713
Standard deviation of eighth root of adjusted KLD	0.277
Upper bound (mean + 1 standard deviation)	1.990

Calculation of the original KLD value for the cut-off point is [1.990⁸ = 245.937] x 0.001 = 0.25 (multiplying by the minimum value of KLD reverts to the original distribution as figures are from the transformed distribution).

Table 7.

Determination of cut-off values of ShD.

Minimum value of ShD distribution	0.001
Mean of fourth root of adjusted ShD	3.023
Standard deviation of fourth root of adjusted ShD	0.810
Upper bound (mean + 1 standard deviation)	3.833

Calculation of the original ShD value for the cut-off point is [3.833⁴ = 215.852] x 0.001 which is 0.22. The calculated cut-off values for the chi-square were determined in case we needed to use ChiSQ as a tie breaker between KLD and ShD.

Table 8.

Determination of cut-off values of ChiSQ.

Minimum value of ChiSQ distribution	0.001
Mean of natural log of adjusted ChiSQ	4.624
Standard deviation of natural log of adjusted ChiSQ	1.422
Upper bound (mean + 1 standard deviation)	6.046

Calculation of the original chi-square value for the cut-off point was done by taking the following steps in a sequential manner: taking antilogarithm of 6.046 we obtained e^6.046 = (2.718^6.046) = 422.155. Now multiplying by the minimum value of the chi-square of 0.001 gave us a cut-off value of 0.42.

Table 9 contains similarity values calculated for the 1998 Speech 4 BOC that are used as the illustrative example and the application of cut-off values is shown. Table 9 shows that in six out of seven concepts, that is, Canada, economy, currency, dollar, rates and international, the calculated values of KLD and ShD are not above the respective cut-off values. With regards to the concept ‘crisis’, the ShD is slightly above the cut-off value of 0.22. However, its ChiSQ value is below, thus making us conclude that there is similarity. This leads to the conclusion that 1998 Speech 4 BOC is similar to its corresponding newswires. Therefore the message contained in the speech is being transmitted to the news stream of traders.

Table 9.

Similarity statistics for 1998 Speech 4 BOC with 1998 Factiva newswires.

		KLD	ShD	Chi-SQ	Decision
Canada	economy, currency, dollar, rates, international, crisis	0.06	0.06	0.08	Similar
Economy	Canada, currency, dollar, rates, international, crisis	0.12	0.12	0.18	Similar
Currency	Canada, economy, dollar, rates, international, crisis	0.07	0.12	0.10	Similar
Dollar	Canada, economy, currency, rates, international, crisis	0.02	0.01	0.02	Similar
Rates	Canada, economy, currency, dollar, crisis	0.09	0.05	0.15	Similar
International	Canada, economy, currency, dollar, crisis	0.04	0.15	0.06	Similar
Crisis	Canada, economy, currency, dollar, rates, international	0.08	0.23	0.10	Similar

A similar approach was taken on 38 BOC speeches and 15 RBA speeches. The results are summarized in Section 7.

7. Discussion of results

The 38 BOC and 15 RBA speeches cover a 15 year period (1995–2009). Table 10 is a snapshot of important news events that occurred during that period.

Table 10.

Major news events 1995–2009.

Year	Event
1995	US rescues Mexican Economy with $20 billion aid programme. Nerve gas attack in Tokyo subway.
1996	Beginning of US budget deficit crisis.
1997	Hong Kong returns to China. Beginning of Asian financial crisis.
1998	Europeans agree on a common currency, the Euro. Russia enters financial crisis. Japan officially declares a recession.
1999	Asian financial crisis spreads to Latin America.
2000	Collapse of Dot-com bubble.
2001	September 11 terrorist attacks. Enron and Arthur Anderson file for bankruptcy.
2002	WorldCom files for bankruptcy. IMF signs emergency loan to Brazil.
2003	US commences military operations in Iraq.
2004	Chechen terrorists take schoolchildren hostage in Russia.
2005	Angela Merkel becomes Germany’s first female chancellor.
2006	Subprime market collapse occurs in USA.
2007	US Federal Reserve injects billions into US money markets to increase liquidity.
2008	Global stock markets fall dramatically. Lehman Brothers files for bankruptcy.
2009	Newly elected President Obama signs US$787 billion economic stimulus package.

Over this period 38 BOC and 15 RBA speeches were identified by Leximancer has having common concepts between the speeches and their corresponding newswires. For ease of presentation and explanation, the results have been aggregated into four categories: (1) a summary of the speeches in which 100% of the concepts show similarity between the main speech and its corresponding newswires; (2) from 70 to 99% of concepts show similarity; (3) from 50 to <70% of concepts show similarity; and (4) <50% of concepts show similarity – this category was deemed to reflect dissimilarity.

For the BOC, Table 11 shows that, in 34 out of a total of 38 speeches (89.5%), there is similarity between the speech and the corresponding newswires. Of the RBA speeches, 14 out of a total of 15 (93.3%) show similarity with the corresponding newswires.

Table 11.

Summary of the results.

Category of results	BOC	RBA
Category A: all concepts showing similarity	17	11
Category B: between 70 and 99% of concepts showing similarity	9	0
Category C: from 50 to 70% of concepts showing similarity	8	3
Total showing similarity	34	14
Category D: dissimilarity (<50% of concepts showing similarity)	4	1
Total number of speeches	38	15

These findings empirically support the position that communication from the central banks indeed flows into the information set available to traders. This therefore clears the path for the next stage of our research, which is to establish statistically the extent to which this communication is effective, that is, the extent to which currency traders actually use this information in their transactions. The significance of our contribution is quite considerable given that it would potentially help to inform the relevant policy-making bodies to appropriately tailor their indirect intervention measures to influence the currency markets. At this stage we have shown that the information does trickle through to the currency traders’ available information set – so central banks can, following the publication of our work, at least be assured that their words of wisdom are not getting lost in transmission! Furthermore, this would also alert foreign exchange traders on tapping better into the latent signals generated by the central bank via speeches about the directions it would like to see the currency market go in the near to medium term.

8. Conclusion

In this paper we examined whether speeches made by central banks flow into the information set that is available to traders in a financial (specifically currency) market with informational efficiency. While earlier studies have largely glossed over the mechanics via which an indirect intervention process is purported to work, we have posited and demonstrated that the information embedded in speeches of central banks does flow into the news arena and become part of market information set that is available to all traders. We tracked the central bank speeches in our sample not by simply reading the newswires published after the speech, but inductively through the use of Leximancer text-mining software, which was used to perform content and relational analysis of both the speech as well as the resulting newswires. This study provides empirical evidence that information contained in central bank speeches does in fact flow into the information set available to currency traders. Methodologically too, our paper breaks new grounds. First, we matched the contents of the central bank speeches and the newswires by comparing the common concepts in the relevant co-occurrence matrices generated via Leximancer. Subsequently, we quantitatively established the extent of the match by computing mathematical measures of information divergence – and in doing so, arguably for the first time, successfully applied a pure information science technique to what is essentially a financial research problem.

Our present paper is the first of a two-part study. In the next part of this research project we will endeavour to determine if, and to what extent, that information actually influences the exchange rate.

Footnotes

Funding

This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

References

Singer

. Migrant remittances and exchange rate regimes in the developing world. American Political Science Review 2010; 104(2): 307–323.

Beine

Janssen

Lecourt

. Should central bankers talk to the foreign exchange markets?Journal of International Money and Finance 2009; 28(5): 776–803.

Fratzscher

. Communication and exchange rate policy. ECB Working Paper no. 363 2004.

Fratzscher

. Strategies of exchange rate policy in G3 economies. Economics Letters 2005; 89(1): 68–74.

Fratzscher

. Communication and exchange rate policy. Journal of Macroeconomics 2008; 30(4): 1651–1672.

Bernal

Gnabo

J-Y

. Announcements, financial operations or both? Generalizing central banks’ FX reaction functions. Journal of the Japanese and International Economies 2009; 23(4): 367–394.

Park

Song

. Japanese vocal intervention and the yen/dollar exchange rate. Japan and the World Economy 2008; 20(1): 61–81.

Ehrmann

Fratzscher

. Communication by Central Bank Committee members: Different strategies, same effectiveness?Journal of Money, Credit and Banking 2007; 39(2-3): 509–541.

Bernanke

. Fedspeak. Remarks at the Meetings of the American Economic Association, San Diego, CA, 3 January 2004.

10.

Kohn

Sack

. Central bank talk: Does it matter and why? Board of Governors of the Federal Reserve System (US). Finance and Economics Discussion Series 2003; 55.

11.

Reeves

Sawicki

. Do financial markets react to Bank of England communication? European Journal of Political Economy 2007; 23(1): 207–227.

12.

Ranaldo

Rossi

. The reaction of asset markets to Swiss National Bank communication. Journal of International Money and Finance 2010; 29(3): 486–503.

13.

Fratzscher

. On the long-term effectiveness of exchange rate communication and interventions. Journal of International Money and Finance 2006; 25(1): 146–167.

14.

Cheung

Y-W

Chinn

. Currency traders and exchange rate dynamics: A survey of the US market. Journal of International Money and Finance 2001; 20(4): 439–471.

15.

Ehrmann

Fratzscher

. The timing of central bank communication. European Journal of Political Economy 2007; 23(1): 124–145.

16.

Oberlechner

Hocking

. Information sources, news, and rumors in financial markets: Insights into the foreign exchange market. Journal of Economic Psychology 2004; 25(3): 407–424.

17.

Macfarlane

. Review of the Reserve Bank’s arrangements for releasing market-sensitive information. Reserve Bank of Australia 2000.

18.

Bhattacharya

Kumar

. An ANN-based auditor decision support system using Benford’s law. Decision Support Systems 2011; 50(3): 576–584.

19.

Kim

Cho

(eds). Diverse evolutionary neural networks based on information theory. Neural information processing. Berlin: Springer, 2008.

20.

Ghazinoory

Ameri

Farnoodi

. An application of the text mining approach to select technology centers of excellence. Technological Forecasting and Social Change 2013; 80(5): 918–931.

21.

Pons-Porrata

Berlanga-Llavori

Ruiz-Shulcloper

. Topic discovery based on text mining techniques. Information Processing and Management 2007; 43(3): 752–768.

22.

Fayyad

. Data mining and knowledge discovery: Making sense out of data. IEEE Expert 1996; 11(5): 20–25.

23.

Chen

Liu

. The contribution of data mining to information science. Journal of Information Science 2004; 30(6): 550–558.

24.

Stemler

. An overview of content analysis. Practical Assessment, Research and Evaluation 2001; 7(17): 137–146.

25.

Ding

Xia

Jin

Shi

. Novel fuzzy information proximity measures. Journal of Information Science 2007; 33(6): 678–685.

26.

Leximancer. Leximancer from words to meaning to insight. Manual, version 3.5, 2010.

27.

Smith

Humphreys

. Evaluation of unsupervised semantic mapping of natural language with Leximancer concept mapping. Behavior Research Methods 2006; 38(2): 262–279.

28.

Rooney

. Knowledge, economy, technology and society: The politics of discourse. Telematics and Informatics 2005; 22(4): 405–422.

29.

Kabanoff

Brown

. Knowledge structures of prospectors, analyzers, and defenders: Content, structure, stability, and performance. Strategic Management Journal 2008; 29(2): 149–171.

30.

Lin

. Divergence measures based on the Shannon entropy. IEEE Transactions on Information Theory 1991; 37(1): 145–151.

31.

Chang

C-I

. An information-theoretic approach to spectral variability, similarity, and discrimination for hyperspectral image analysis. IEEE Transactions on Information Theory 2000; 46(5): 1927–1932.

32.

Hoffmann

Kolossa

Köhler

B-U

Orglmeister

. Using information theoretic distance measures for solving the permutation problem of blind source separation of speech signals. EURASIP Journal on Audio, Speech, and Music Processing 2012; 2012(1): 1–14.

33.

Gokcay

Principe

. Information theoretic clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence 2002; 24(2): 158–171.

34.

Chandola

Banerjee

Kumar

. Anomaly detection: A survey. ACM Computing Surveys 2009; 41(3): 15.

35.

Burnham

Anderson

. Kullback–Leibler information as a basis for strong inference in ecological studies. Wildlife Research 2001; 28(2): 111–119.

36.

Belov

Armstrong

. Automatic detection of answer copying via Kullback–Leibler divergence and K-index. Applied Psychological Measurement 2010; 34(6): 379–392.

37.

Carpineto

De Mori

Romano

Bigi

. An information-theoretic approach to automatic query expansion. ACM Transactions on Information Systems 2001; 19(1): 1–27.

38.

Johnson

Gruner

Baggerly

Seshagiri

. Information-theoretic analysis of neural coding. Journal of Computational Neuroscience 2001; 10(1): 47–69.

39.

Alguliev

Aliguliyev

Ganjaliyev

. Investigation of the role of similarity measure and ranking algorithm in mining social networks. Journal of Information Science 2011; 37(3): 229–234.

40.

Egghe

Rousseau

. An approach to similarity measurement of absence–presence data: The case that common zeros matter. Journal of Information Science 2004; 30(6): 509–519.

41.

Lee

. Selecting diagnostic tests for ruling out or ruling in disease: The use of the Kullback–Leibler distance. International Journal of Epidemiology 1999; 28(3): 521–525.

42.

Crestani

. Combination of similarity measures for effective spoken document retrieval. Journal of Information Science 2003; 29(2): 87–96.

43.

Goh

DHL

Foo

SSB

. The effect of similarity measures on the quality of query clusters. Journal of information Science 2004; 30(5): 396–407.

44.

Osborne

. Improving your data transformations: Applying the Box–Cox transformation. Practical Assessment, Research and Evaluation 2010; 15(12): 1–9.

45.

Pallant

. SPSS survival manual: A step by step guide to data analysis using SPSS. Milton Keynes: Open University Press, 2010.

46.

Keene

. The log transformation is special. Statistics in Medicine 2007; 14(8): 811–819.

47.

Bigi

. Using Kullback–Leibler distance for text categorization. Advances in Information Retrieval. Berlin: Springer, 2003, pp. 305–319.

48.

Dragalin

Fedorov

Patterson

Jones

. Kullback–Leibler divergence for evaluating bioequivalence. Statistics in Medicine 2003; 22(6): 913–930.