2. Monitoring Financial Stability in a Complex World

The DFA is officially the Dodd-Frank Wall Street Reform and Consumer Protection Act; see U.S. Congress (2010). The OFR provisions of the DFA were based on an earlier bill introduced by Sen. Jack Reed; see U.S. Senate (2010). Among many other things, the DFA created the Financial Stability Oversight Council (FSOC) and Office of Financial Research (OFR) to monitor threats to financial stability in the U.S. The Federal Reserve Board established a new Office of Financial Stability Policy and Research. The Federal Deposit Insurance Corporation (FDIC) established a new Office of Complex Financial Institutions. Similar significant initiatives exist at other central banks, regulatory agencies, and multilateral institutions worldwide.

Workhorse regulatory data collections in the U.S. include the SEC’s 10-K (annual) and 10-Q (quarterly) reports, bank Call Reports (quarterly), and Thrift Financial Reports (quarterly). While it is difficult to generalize, reporting abroad tends to be less frequent than in the U.S.

Blundell-Wignall and Atkinson (2010) describe many of the issues that arise in this approach. Adrian and Shin (2010) discuss ways in which myopic VaR implementations can exacerbate boom-bust leverage cycles in financial markets.

The literature on systemic risk measurement is large and growing. Bisias et al. (2011), IMF (2009), and ECB (2010b) provide overviews.

The final rule on fair value measurement was adopted by the Financial Accounting Standards Board (FASB) and the International Accounting Standards Board (IASB) in 2011; see IASB (2011). This harmonizes the FASB and IASB approaches, and replaces earlier, very similar guidance under FASB Topic 820, formerly the Statement of Financial Accounting Standards (SFAS) 157.

Pozsar et al. describe the shadow banking sector in greater detail. Lo (2011b) asserts that outsiders know almost nothing about the nature and magnitude of the risk exposures of the hedge fund industry, and are forced to resort instead to plausible “conjectures.” Based on data from Institutional Investor, he emphasizes that the size of the now defunct LTCM is an order of magnitude smaller (in 1998 dollars) than a number of current hedge funds.

The general literature on technological innovation is largely beyond our scope. See Antonelli (2009) for a general overview. See Tufano (2003) for an economist’s overview of the literature on financial innovation.

The terms hard and soft are close to their usage in the knowledge management (KM) literature, and related to the notions of tacit and explicit knowledge (see Hildreth and Kimble 2002). Very roughly, knowledge is soft (hard) if it cannot (can) be fully and accurately written down. Examples of soft knowledge are technical skills or the capacity for rapid cognition; examples of hard knowledge include loan amounts or monthly income. Knowledge is tacit (explicit) if it is not (is) written down. In practice, knowledge typically exists in a continuum or mixture of hard and soft, tacit and explicit. See Choo (2006) for a survey of the field of KM in organizations.

See, for example, Berger et al. (2005), Agarwal and Hauswald (2010), Liberti and Mian (2012), and Petersen and Rajan (2002). In the mortgage industry, the two most common automated underwriting systems are Freddie Mac’s Loan Prospector and Fannie Mae’s Desktop Underwriter.

For example, Cordell et al. (2011, p. 25, emphasis in the original), citing IOSCO (2008, p. 2), note that, “clearly data quality was a problem, fueled as it was by declining underwriting standards. One very valid point on the data is that the quality of the data being provided deteriorated significantly in the buildup to the crisis because of declining underwriting standards, by the IOSCO’s reckoning, ‘beginning in late 2004 and extending into early 2007.’”

See Flood (2009) on the costs of data reconciliation across multiple schemas and systems. The discussion here implicitly assumes that the information collected at origination comprises a stable and well-defined set of attributes. Because relational databases are costly to change in production systems, back-office practices typically require static data models describing instruments that traded repeatedly. Front-office dealers, on the other hand, frequently prefer customized deal structures built from one-off or semi-standardized components with idiosyncratic data representations. This can overwhelm back-office data modelling.

Pozsar et al. (2010) provide estimates of the size of the shadow banking markets. On the other hand, other segments, such as the hedge fund industry are much murkier; see Lo (2011b).

The numbers provided here are intended to be suggestive of the situation in financial markets, rather than conclusive. The growth in processing power represented by Moore’s Law is particularly relevant as a benchmark for the growth in storage requirements in finance, since advances in processor power help enable the development of new market segments. Valuation of structured securitizations, for example, makes frequent use of CPU-intensive Monte Carlo analyses; see, for example, Berthold et al. (2011). Similarly, while high-frequency trading is typically latency-dependent, it nonetheless benefits from high-performance processing power; see, for example, Intel (2010).

For at least two reasons, the S&P 500 trading volume depicted here represents a lower bound on the growth in data generated by the financial system. First, it does not encompass the vast increase in derivative markets that has occurred since 1980. Comprehensive data on outstanding balances (not trading volumes) for OTC derivatives are available only since 1998; see BIS (2010). These have shown a roughly order-of-magnitude increase over the past decade, with approximately $600 trillion notional outstanding in June 2010 (ca. $25 trillion in market value), dominated by interest-rate swaps. The growth in trading is also reflected in and compounded by the growing “financialization” of the economy: the share of GDP represented by the U.S. financial sector (including insurance) has tripled since World War II, and nearly doubled since 1980 (see Philippon 2008, Fig. 1, p. 36). Second, each transaction generates a number of internal and external versions of the trade information for financial reporting, regulatory compliance, risk management, etc. These ancillary data sets should all be kept consistent, but the number of reconciliations required does not typically scale linearly with the number of positions or transactions (see Flood 2009). Note that time scales in financial markets have also been shrinking, evidenced by the growth of algorithmic trading; see Castura et al. (2010) or Hendershott et al. (2011). Because more must happen faster, the consequences of process failure are correspondingly larger.

Extrapolating from their 23 % approximate annual growth rate over the 1986–2007 period–and assuming it applies at least equally to the financial services sector—we see that data storage requirements are on the order of 10,000 times greater in 2005 compared to 1980. For comparison, they estimate annual growth rate for worldwide computing capacity at 58 %/year, and telecommunications traffic at 28 %/year. At the same time, advances in processing power are also creating engineering challenges as applications impose heavier demands on legacy database technologies; see, for example, Stonebraker et al. (2007) and Pandis et al. (2010).

TMPG (2011) offers main mechanisms for settlement fails: miscommunication, operational problems (e.g., the September 11, 2001 disruption), “daisy chain” fails in which failure to receive collateral on one deal leads to failure to deliver on another (this is an example of “tight coupling” as described by Perrow 1999 and Bookstaber 2007), and “strategic” fails in which the “short” counterparty intentionally reneges, typically because the cost of borrowing securities to fulfill his commitment approaches or exceeds the time-value opportunity cost of postponing delivery. Strategic fails are thus exacerbated by episodes of low interest rates.

Note that this is the delinquency rate for mortgages overall, including both prime and subprime loans. The delinquency rate for subprime loans in isolation was much worse, peaking at over 15.5 % in the final quarter of 2009. Prime mortgage borrowers are easy for mortgage servicers to handle: until the crisis, defaults and foreclosures were rare, and loans typically had very standard structures. As a result, the mortgage servicing business became concentrated in a handful of specialized banks that invested in the relevant information technology infrastructure. In contrast, subprime mortgages employed a variety of innovative terms ostensibly intended to constrain the monthly mortgage payment to a level expected to be sustainable for the borrower. In addition to a more complex servicing process, subprime loans exhibit very different default rates. In hindsight, it is apparent that both underwriting standards and credit pricing were too lax for an enormous number of subprime mortgages, especially those originated after 2005. Dungey (2007a) provides a good introduction to the mechanics of the mortgage servicing process. Dungey (2007b) is a similar overview of the foreclosure process on the eve of the crisis.

Robo-signing is the practice of attaching signatures to affidavits and other foreclosure documents so quickly that it is inconceivable that a reasonable review occurred. This is a data-validation issue on two levels: first, the signature is an attestation, based on (supposedly) diligent human review, of the validity of the information in the affidavit. Second, because it seems in many cases that the task was delegated to unauthorized and unqualified shills as an expedient, the signatures themselves become data requiring subsequent validation.

Noll (2009) chronicles a much earlier instance of robo-signing, perhaps the first in history. As the U.S. Treasury systematically shrank the denominations of the new greenback currency it was issuing to finance the Union efforts in the Civil War, the number of signatures required quickly outstripped the capacity of Treasury staff, even after signature authority was broadly delegated. By 1863, the Treasury building was home to industrial printing operations, including engraved signatures.

Fleming and Garbade (2005) provide a contemporary analysis of settlement fails in the Treasuries market. The Counterparty Risk Management Policy Group Report (CRMPG 2005), a statement by participants of industry best practices, was a catalyst for change at the time. When the operational costs are small and/or not internalized, unilateral remediation is difficult to justify.

See Markham (2002, pp. 362–367) and SEC (1972, pp. 3–6).

On the need for straight-through processing, see CPSS-IOSCO (2001) and CPSS (2008). For an example of an implementation perspective, see Ciulla et al. (2010). CPSS (2011) identifies five main categories of financial market infrastructure, each of which encompasses a multitude of processes, and each of which might benefit from STP:

Ironically, technological advances may also encourage novel practices—such as transacting via text messages from wireless devices—that place further demands on data management and validation.

Madnick and Zhu (2006) offer some concrete examples of the role of semantic context in defining the quality of a data set, as well as suggestions for effective management of that semantic context to improve data quality. Fuerber et al. (2011) indicate a similar path forward, defining a data quality constraints language targeted at the Semantic Web.

See Lo (2011a, at 13: 18). The study he refers to is Khandani and Lo (2011).

Flood et al. (2010) also discuss some of the implications of financial complexity for information management.

Taub (2008), IMF (2001), and Copeland et al. (2010) describe the mechanics of the repo markets in greater detail. The repo markets are very large, and there are naturally numerous variations.

A prime broker is a specialized firm that provides a range of related services to hedge funds and other investment managers. Typical services include custody, securities settlement, tax accounting, and account-level reporting. Lehman Brothers acted as prime broker for a number of large hedge funds at the time of its demise. In the example here, the hedge fund is the “collateral pledger” and the prime broker is the “collateral pledgee.”

Deryugina (2009) describes the structure of rehypothecation transactions and related legal considerations in detail. She emphasizes the importance of the relatively lenient U.K. rules on rehypothecation in attracting prime brokerage business to London.

Pozsar and Singh (2011) further explore the complexities introduced by rehypothecation of collateral.

Shleifer and Vishny (2011) survey the issues surrounding fire sales and contagion.

A “graph” is an abstract mathematical formalism of a set of elements, called “nodes” (or vertices or points), and a corresponding set of “edges” (or lines) that connect the nodes. Graph theory has developed a large body of proved propositions describing the nature of graphs. See, for example, Diestel (2006) for further details.

The literature on network models of systemic risk is large and growing. For recent overviews, see Haldane (2009), ECB (2010a), or Moussa (2011). Engle and Weidman (2010) specifically consider the technical capabilities needed for supervising systemic financial risk.

See, for example, Gooch and Klein (1997), especially pp. 63–64.

The experience of the MERS system (see Hunt et al. 2011) is emblematic of the difficulties and unintended consequences endemic to the automation of long-standing processes. See Gilbert and Lynch (2002) on eventually consistent architectures and the so-called “CAP theorem.” See Srivastava (2006) on other recent advances in data architectures.

The DFA, at §154(b) (2) (A) (i), also requires the OFR to build a “financial company reference data base.” This will not be trivial because many individual obligors exist in parent-subsidiary hierarchies with de facto cross-guaranties. In some cases, these are de jure cross-guaranties: the DFA (at §616) reiterates and extends the “source of strength” doctrine that requires bank and thrift holding companies provide financial support to their subsidiary depository institutions.

Situational awareness is a concept that originated in a military context to describe the outcome of a tactical process of perception, comprehension, and projection onto a near-term decision space; see, for example, Leedom (2001). The issues of organizational capacity for systemic surveillance are better developed and understood in certain other disciplines. See, for example, Wagner et al. (2006).

There are important exceptions, of course. Unstructured data, for example, articles from newspapers and the trade press or interviews with regulators or industry participants, will be an important source of information. The information on settlement fails–which by definition do not result in contracts–presented by Bradley et al. (2011) might provide the basis for a systemic key risk indicator. Bisias et al. (2011) identify a class of early warning models that are based solely on macroeconomic aggregates.

Provenance is a technical term for the metadata to support repeatable collection or derivation of the data. In many cases where issues regarding chain of custody or data lineage apply, establishing accurate data provenance can be crucial. Data source tagging–i.e., citation of the source–is a basic technique. There are standard markup languages, such as the Data Documentation Initiative (see DDI 2009) for capturing provenance metadata in a structured format.

Similarly, efforts to build a “semantic repository” for finance–a comprehensive set of standard, structured, and interrelated definitions to augment the data model and help specify the attributes of contractual relationships; for example, see Enterprise Data Management Council (EDMC) (2011) or Madnick and Zhu (2006)—are extremely useful, but not sufficient. A semantics repository is also only one input into the process of understanding, and not a full solution or a methodology. Other important techniques include object definition, unique entity symbology, information standardization, and business process flow; these are beyond the scope of the present paper.

For example, Vogels (2009), in a discussion of the “eventual consistency” model of distributed and replicated data, cites Brewer’s (2000) “CAP (consistency, availability, partition-tolerance)” proposition that, “of three properties of shared-data systems–data consistency, system availability, and tolerance to network partition—only two can be achieved at any given time.” A formal proof is given by Gilbert and Lynch (2002).

Regarding data visualization, see Hansen et al. (2009), Johnson et al. (2007) and Lemieux et al. (2012).