Since the Lehman Brothers collapse and financial crisis in 2008, regulators, investors, lawmakers and the public in general have repeatedly demanded that banks prove their financial health, particularly in a stressed economic environment. US and European regulators have established a more formal procedure, which periodically requires banks to stress test their capital base given a certain scenario. This scenario is expressed in macroeconomic factors and financial indicators, and regulators provide these factors on a quarterly basis for a period of two or three years ahead. For example, regulators might produce a scenario that includes the S&P 500 falling 30% and US unemployment reaching 12% (and many other factors) in a certain quarter.
Based on this information, banks would then assess the impact of this economic scenario in terms of market and credit losses in their portfolios and how their capital base would behave in this situation. The need for these stress tests derives from governments’ recent experience of bailing out banks that did not have enough capital to cope with extremely unlikely negative scenarios. The novelty is that banks are also required to analyse the impact of this scenario on operational risk. The relationship between these macroeconomic factors and indicators to market and credit risks is straightforward, but what about operational risk?
In operational risk terms, finding a consistent statistical relationship between these factors and incorporating them in a sound way into the framework is a challenge. The most obvious reason is data. Though progress has been made across the industry to improve the quality of operational loss data, some issues remain:
- Completeness: The completeness of internal and external data, while an objective for the industry, is elusive. Even when using external data to assess correlations, it is open to question whether loss databases such as the Operational Riskdata Exchange (ORX) are comprehensive and whether members are reporting all losses they suffer.
- Varying collection thresholds: Several banks started with high collection thresholds and have been reducing them – this makes it difficult to find a long time series of standard events.
- Natural scarcity: Operational losses are sparser. For some risk types, losses would not happen with daily or even weekly frequencies, while economic indicators are available daily. The solution in this case is to aggregate losses on a monthly or quarterly basis. However, as the aggregation increases, quite a few spurious correlations could appear that would have no logical support (for example, WTI crude oil prices, aggregated quarterly using ORX data, show a 32% correlation with business disruption and system failure losses).
- Dates: Operational losses would have many dates associated, for example, ‘occurrence date’ (when losses occur), ‘impact date’ (when losses are realised), and ‘account date’ (when losses are booked to the general ledger). Changing the type of date used would affect correlations.
For several important operational risk types, the lag that exists between a macroeconomic event and the losses can be many years, well beyond the scope of the exercise proposed by the regulators. One example is litigation losses (mostly under the ‘clients, products and business practices’ risk type). For example, banks did not start to set reserves for litigation arising from the mortgage crisis in the US in 2007–8 until 2011. The cycle for litigation can take from three to six years, or longer. Considering the regulatory stress tests only span a couple of years ahead, it is difficult to find a meaningful correlation between a certain macroeconomic scenario and litigation losses within this time frame.
Given these constraints, modellers need to take cautionary measures. The first is to break down operational risk events into their Basel risk types. Operational risk is an amalgamation of different risk types, and the impact of macroeconomic factors can vary among them. For execution losses (the ‘execution, delivery and process management’ type), a steep decline or high volatility in the financial markets usually increases the trading volume, which can increase execution losses.Using ORX data, this relationship is not apparent using daily, weekly or even monthly data, but starts to show up on a quarterly aggregation. Considering that execution risk would represent about 20% to 40% of the total operational risk, a volatile macroeconomic scenario can have some significance.
However, there is no robust and conclusive correlation using this data. An example of a strong correlation is when it maintains at any aggregation level. For example, if we analyse the relationship between the Dow Jones Industrial Average and the S&P 500, we will find a strong relationship on a daily basis and if we extend this to weekly, monthly and quarterly the association will hold. This does not happen when data from ORX is compared against any of these macroeconomic factors.
For some risk types, such as employment risk (the ‘employment practices and workplace safety’ category), a stress scenario can lower the risk. Analysing unemployment data against employment-related losses in the US, it can be seen that higher unemployment levels reduce the risk, as most employees are more worried about securing their jobs and avoid litigation with employers.
Bearing in mind these difficulties, many banks prefer to use subjective modelling of these correlations and relationships in the preparation of these scenarios. The danger of this method is that we can establish relationships that seem logical but lack support in the data – and therefore lack a connection with reality.