Big Data and central banks

Abstract

This commentary recaps a Centre for Central Banking Studies event held at the Bank of England on 2–3 July 2014. The article covers three main points. First, it situates the Centre for Central Banking Studies event within the context of the Bank’s Strategic Plan and initiatives. Second, it summarises and reflects on major themes from the event. Third, the article links central banks’ emerging interest in Big Data approaches with their broader uptake by other economic agents.

Introduction

On 2 and 3 July 2014, the Bank of England hosted an event titled “Big Data and Central Banks.” The purpose of this commentary is to articulate the motivations for the event, key themes emerging from it and explain why Big Data is likely to become of increasing importance to central banks in the years ahead. The article proceeds as follows. The first section explains how the event fits with other activities at the Bank aimed at expanding its data sources and enhancing its data analysis capabilities. Relevant related activities include a predecessor Centre for Central Banking Studies (CCBS) event, the creation of the Bank’s Data Lab and Data Community, and, most significantly, the release of the Bank’s Strategic Plan. The second section then summarises presentations made at the event. Among these themes are the benefits for central banks in having standardised granular data, the importance of legal considerations in enabling and constraining the scope of granular data collections, and the development of inductive analytical approaches to complement deductive approaches that traditionally have held sway in central banks. The article then concludes by speculating that Big Data might not only change how central banks operate, it also might be transforming how financial firms and other economic agents do business. To the extent that this transformation is occurring, it is important for central banks to understand how Big Data is changing the structure of the economy in ways that might impact monetary and financial stability, as well as economic growth and employment.

Background

One standard definition of Big Data is that it is data displaying one or more of the following characteristics:

These data are of high volume, often because data are reported on a granular basis, that is, item-by-item, for example, loan-by-loan or security-by-security ¹ ;

these data are of high velocity, because these data are frequently updated and, at the limit, collected and analysed in real-time;

these data are qualitatively various, meaning they are either non-numeric, such as text or video, or they are extracted from novel sources, such as social media, Internet search records or biometric sensors.

Judged by this definition, the Bank of England traditionally has not dealt much with Big Data. Data volume historically has been low because its primary sources have been summary financial statements compiled by the Bank’s Statistics and Regulatory Data Division and aggregate statistics compiled by the Office for National Statistics. Data velocity also has been slow because these statistics and financial statements are reported in the main at most on a quarterly frequency and revised with lags. And although the Bank does have a history of gathering qualitative data through surveys and interviews with external contacts, overall the variety of data has been minimal because most formal analysis undertaken in the Bank uses structured data sets, that is, numeric data stored in relational databases, i.e. row and column format.

More recently, however, the Bank has been one of the leading central banks when it comes to research using higher volume, higher velocity and qualitatively more various data sets. For example, McLaren and Shanbhogue (2011) used Google data as an indicator of UK labour and housing market conditions. Benos and Sagade (2012) used equity transaction data to understand the consequences of high frequency trading on stock markets, while Benos et al. (2013) used transactional trade repository data to investigate the structure and dynamics of the UK credit default swap market. And both Davey and Gray (2014) and Merrouche and Schanz (2009) have used high-value, high-velocity payment systems data to analyse banks’ intraday liquidity management.

The steady increase in research done by the Bank using Big Data in part reflects its greater availability. This is because the financial crisis of 2007–2008 prompted a number of statutory and supervisory initiatives that require greater disclosure by financial firms of their data to central banks and regulators. These include the reporting of firms’ large exposures on a counterparty-by-counterparty basis on Common Reporting (COREP) templates; security-by-security reporting of insurers’ assets mandated by Solvency II scheduled to come into force in 2016; and the reporting of transactional derivatives data as required by European Market Infrastructure Regulation (EMIR). Such granular data became more readily available to the Bank of England after it assumed supervisory and regulatory responsibilities last year with the establishment of the Prudential Regulation Authority (PRA).

Consequently, the Bank hosted an event in 2013 titled “The Future of Regulatory Data and Analytics.” Its focus was on how to best integrate the Bank’s new supervisory and regulatory data collections with existing statistical compilations. ² Much of the discussion at the event centred on developing a new post-crisis data strategy for central banks, prefiguring the Bank’s three-year Strategic Plan released in March 2014 (Bank of England, 2014). Although the Strategic Plan has many facets, one of its major points of emphasis is data. This is particularly evident in the three strategic initiatives that fall under the “Analytical Excellence” pillar. The “One Bank research agenda” initiative commits the Bank to opening up to the public previously proprietary data sets in order to crowd-source solutions to challenging policy questions. The “New Approach to Data and Analysis” initiative created an Advanced Analytics Division with the objective of establishing a centre of excellence for the analysis of Big Data. And the “One Bank Data Architecture” initiative is to be overseen by the first Chief Data Officer in the Bank’s history, with the goal of integrating data across the Bank, partly through the enforcement of metadata standards to enable easier information sharing across the organisation.

Since announcing these strategic initiatives, the Bank has made strides toward their achievement. Three milestones are worth highlighting. The first is the establishment of a Data Lab. The Lab is a room in the Bank with computers that are uploaded with state-of-the-art IT tools. Bank employees who visit the Lab are supported by a small team of IT experts who help them store, manipulate, visualise and analyse granular and unstructured data. A second and related development is the formation of a new Bank-wide Data Community. The Data Community organises a set of activities and events designed to raise awareness among staff about Big Data issues, including monthly seminars, a new Bank intranet site with information on novel ways staff can use data, and a Data Art Gallery event exhibiting examples of innovative approaches for visualising data. Finally, a third key milestone was the convening of the “Big Data and Central Banks” event. More information about that event is contained in the unabridged version of this commentary found on the Bank’s website. ³

CCBS event

Granular data

Many salient issues were raised during the two-day event. One recurring theme was the benefits to central banks in having access to granular data from financial firms. As noted earlier, central banks typically have collected aggregate data from firms using reporting returns structured like standard financial statements. These returns tend to translate end user requirements literally. For example, an end user might request data on firms’ liquid assets, defining that term as deposits and government securities with original maturities less than one year. A regulatory return might then be sent to firms containing a field for reporting a single “liquid assets” figure, as defined by the end user. However, liquidity is a notoriously fluid concept. Assets that ordinarily might be sold with minimal loss to the seller may no longer be so under changed conditions; for example, if the credit profile of the issuer of the securities held by financial firms has changed. Furthermore, the definition of liquid assets can vary across analysts and central bank areas. For example, some end users might conceive of shares held in money market funds as liquid assets, while others might not. To the extent that each conception of liquid assets gives rise to discrete data collections, multiple returns might have to be filed by financial firms, even though the data reported are highly repetitive except at the margin. Yet the duplication of reported data can still leave data gaps since the aggregated nature of the figures precludes end users from drilling down and asking more detailed questions as circumstances require. For example, if an end user needs to assess whether government securities held by a firm are issued by the central government or municipal authorities. Circumstances might then require the costly ad hoc collection of these data from firms at short notice.

An alternative approach is to collect granular data once. A number of speakers at the CCBS event advocated this approach. A representative view was that different returns often specify varying consolidation bases, valuation methods, accounting conventions, definitions and report at different frequencies, making it difficult to stitch data together. When one speaker’s country suffered a financial crisis, it was discovered that the aggregate data available to the central bank were incomplete and incompatible for pinpointing financial fragilities. This prompted the central bank to introduce exposure-by-exposure reporting. According to the speaker, such granular data now make possible the coherent mapping of the banking system as a whole, enabling the central bank to better spot systemic risk and manage it with macro-prudential policy.

At first glance, the idea that amassing micro-data will make it easier for central banks to discover macro patterns might appear paradoxical. On the contrary, it might make intuitive sense to think that having more data would make it harder to identify the wood from the trees. To paraphrase information economist Herbert Simon, a wealth of information might create a poverty of attention, leading to worse decision-making (quoted in Haldane, 2013). However, a number of speakers noted that granular data become analytically tractable if overlaid with visual analytic tools. Instead of eyeballing millions of rows and columns of granular data, end users are able to quickly picture the data at different units of analysis, drilling down to identify granular fault lines which might be otherwise concealed at an aggregate level. ⁴

The benefits in gaining access to granular data might not only accrue to central banks. They also may accrue to the financial firms central banks regulate. For instance, one central banker noted that by having access to the granular detail on loans pre-positioned by firms with central banks as collateral, this may result in reduced haircuts and thus cheaper emergency liquidity for firms because the central bank can better judge the quality of the underlying loans against which they are lending.

However, greater data granularity in itself is not a panacea. If granular data collections are introduced unevenly across a central bank and managed in end user silos, then the organisation runs the risk of reproducing the inconsistencies and inefficiencies of the current approach of collecting aggregate data using multiple returns. According to one speaker, one way to prevent this from occurring is to harmonise and enforce common definitions of granular data attributes across the organisation. In loan-by-loan databases common attributes include the original value of the loan, the currency in which it is denominated, its purpose, outstanding balance, original and residual maturity, the repayment schedule, the interest rate, the reference rate (if applicable) and information on financial firms’ counterparties, such as the income of the obligor, their domiciled country and credit rating.

Legal considerations

As the foregoing list of attributes indicates, the existing granular data sets collected by some central banks tend to have good coverage of information on counterparties and contracted cash flows. However, one area where the scope of these collections could be extended is in capturing more detail on legal terms and conditions. Even apparently minor legal provisions can have major systematic consequences. Consider the recent Greek sovereign debt crisis. The absence of collective action clauses that would have permitted bondholders to write-down Greece’s outstanding debt with binding effect on credit minorities, coupled with the fact that most of the bond issues named Greek courts as arbitrators in case of dispute, are key factors explaining the European Central Bank’s (ECB’s) decision to purchase Greek sovereign debt (Grant et al., 2014). Among the key clauses in financial agreements which central banks may want to capture are provisions limiting personal liability, creditors’ ranking in bankruptcy proceedings, breach of contract and default triggers and protective covenants. ⁵

However, several speakers at the CCBS event observed that embracing Big Data did not necessarily require that central banks enlarge their data collections. For these speakers, the key task facing central banks is not getting more data. Rather they argued that it is doing more with the data central banks have already. One speaker cited payment systems’ data as a good example. Historically, these real-time data have been used to monitor operational risks. For example, by looking at the data, central banks might observe that a payment system is too reliant on a small number of settlement banks. The central bank might then address this excess concentration by inviting indirect participants accessing the payment system through these settlement banks to become direct participants (Finan et al., 2013). But these same data also can be useful for other purposes. For example, these data might be used to monitor the settlement behaviour of individual firms. Such information might provide a timelier indicator of firms’ liquidity position than liquid assets figures submitted on regulatory returns. Payment systems data also might be linked with other data to achieve new insights. For example, payment data might be blended with loan-by-loan mortgage origination data to identify shortfalls in mortgage repayments much sooner than those shortfalls are reported as arrears by firms in their quarterly financial disclosures.

Indeed blending data gives central banks a ‘third way’ between the false dichotomy of either buying all their data from external parties or building all data capture systems in-house. For example, one speaker presented an analysis of a country’s housing market that blended different data sets – some proprietary, others open source, others purchased from commercial vendors. The combined data set contained mortgage origination information blended with up-to-date information on property prices and obligors’ credit scores from credit rating agencies. The speaker noted that the value of this combined data set was greater than the sum of its individual parts. At the same time, because each part had been collected by an organisation with a comparative advantage in its provision, the speaker claimed that the cost–benefit calculus had been optimised.

However, there are technical and legal obstacles in blending different data sets. The technical obstacle is that data are often stored in different formats so it can be laborious to transform them into a common technical type. The legal obstacle is that the use of data collected by central banks is sometimes restricted only to those purposes explicitly expressed in legislation. That might mean, for example, data collected for supervisory purposes may have restrictions on its use by monetary economists situated in a different part of the central bank, even though the data might be useful for secondary purposes. One speaker argued that these types of legal strictures should be relaxed if central banks are to achieve economies of scope when collecting data and so also reduce the regulatory reporting burden borne by financial firms. In a related vein, another speaker felt more multilateral agreements were needed to allow greater cross-border sharing of data between regulators. The speaker noted that although their central bank has access to highly granular data on the domestic counterparty exposures of their banks, it does not have access to similarly detailed data on the exposures of these domestic banks to foreign counterparties.

New approach to data and analysis

If these types of technical and legal obstacles to sharing and blending data can be overcome, the resulting quantitative increase in data might add up to a qualitatively new approach for analysing the economic and financial system (Varian, 2014). In recent decades, the dominant approach inside central banks to analysing these systems has been deductive. A deductive approach starts from a general theory and then uses particular data to test it. Suppose an analyst starts by positing an accounting identity that the product of the quantity of money (M) and its velocity (V) is equal to the product of the price level (P) and expenditures on goods and services in the economy (Q). ⁶ If the analyst further assumes that the velocity of money is stable, then an increase in money might be hypothesised to result in inflation. ⁷ The analyst might then seek to test the validity of the theory using money and price data over a particular period of time.

An alternative point of departure for research to deduction is induction. An inductive approach starts from data and then seeks to generate theoretical explanation of it. Induction may mitigate confirmation bias, that is, the tendency to seek data which confirms ex-ante assumptions. For instance, one speaker at the event noted that many commentators simply assumed that the recent wave of defaults in subprime mortgages in the United States was caused when adjustable rate mortgages reset to a higher interest rate. According to the speaker, however, careful analysis of granular mortgage data actually revealed that subprime mortgage defaults spiked before their initial, so-called ‘teaser’ rate expired.

Of course, deduction and induction are ideal types. In reality, explanatory approaches are always mixed. ⁸ Nevertheless, the reason why a more self-consciously inductive approach was advocated by some event participants is that the recent crisis punctured many deductive models dominant before the crisis which purported to explain how economies and financial systems work universally. So in the aftermath of the crisis, a space has emerged for a form of induction called abduction. In other words, inferring the best explanation for a particular puzzle given patterns in the data, without pretence to making generalised theoretical claims. As Linnet Taylor et al. (2014) have noted in this journal, one can read the shift to Big Data as “an epistemological change” for economics from “a science based on the notion of the mean and the standard deviation from the ‘normal’ to one based on individual particularity.” ⁹

Conclusion

This article has summarised the Bank’s recent “Big Data and Central Banks” event and placed it within the context of the organisation’s new strategic approach to data analysis. In brief, and to paraphrase one event participant, the new approach involves a shift in tack from analysing structured, aggregated sample data collected with a specific set of questions in mind, to analysing data that are more heterogeneous, granular and complete such that these data are fit for multiple purposes. Throughout the article, emphasis has been placed on the ways bigger and better data might enhance the Bank’s analytical toolkit and improve its operational efficiency, with the end goal being to promote the good of the people of the United Kingdom by maintaining monetary and financial stability.

Viewed in isolation, central banks’ increasing interest in Big Data might be viewed as a conjunctural phenomenon, that is, as a response to the recent financial crisis. However, viewed more broadly, it appears instead to reflect a more fundamental structural shift toward the exploitation of Big Data by other economic agents (Bholat, 2013). This broader embrace of Big Data has both supply and demand sources. On the supply side, increases in the volume, velocity and variety of data have been driven by technological advances that have increased storage capacity and processing power while lowering costs. And on the demand side, there is increasing interest from economic agents in understanding how analysis of their data might enhance productivity and profits (Bakhshi et al., 2014; Brown et al., 2014; Einav and Levin, 2013).

To date, Big Data tools and techniques have had less of an impact on financial services than they have had on other sectors of the economy such as the information and communications industry. However, the situation appears to be changing rapidly. Some of the largest and most established banks are now taking a fresh look at their customers’ transactional data to tailor their customer offer, and to enhance early detection of fraud (Davenport, 2014). At the same time, some of the new ‘challenger’ financial services providers are using innovative risk models that exploit novel sources of data like social media (King, 2014). Taken in sum, these developments may have a positive impact on financial stability in the long term to the extent that they improve the financial decisions made by firms and their counterparties. But there are also nearer term risks if early adopters of Big Data significantly disrupt the business models and profitability of incumbent firms. The impact of Big Data on the wider economy may be similarly double-edged. While it might boost productivity and lower costs, it may also alter the productive structure of the real economy and wealth distribution in ways that are difficult to forecast and measure (Rifkin, 2014). Given the pace and depth of these possible changes, central banks will likely need to make further advances to “nowcast” (Bell et al., 2014) the economy, building on existing initiatives to exploit timelier data on prices, employment and output (Galbraith and Tkacz, 2013; Koop and Onorante, 2013).

In sum, Big Data is likely to become a topic of increasing interest to central banks in the years ahead. This is because Big Data is likely to change both the internal operations of central banks, and transform the external economic and financial systems they oversee.