RDP 2015-06: Credit Losses at Australian Banks: 1980–2013 1. Introduction

Credit risk – the risk that borrowers will not repay their loans – is one of the main risks that financial intermediaries (such as banks) face. Credit risk has been the underlying driver of most systemic banking crises in advanced economies over recent decades (von Westernhagen et al 2004; Bernanke 2010). As credit risk materialises and borrowers fail to make repayments, banks are forced to recognise the reduction in current and future cash inflows this represents. These ‘credit losses’ reduce a bank's profitability and can affect capital. In extreme cases, credit losses can be large enough to reduce a bank's capital ratio below regulatory requirements or minimum levels at which other private sector entities are willing to deal with a bank, so can cause banks to fail.

This paper explores the historical credit loss experience of the Australian banking system. It does so using a newly compiled dataset covering the bank-level credit losses of larger Australian banks over 1980 to 2013. Portfolio-level credit loss data – data that break losses down by type of lending (e.g. business, housing and personal lending) – are available for a broad range of banks only from 2008 onwards, so this paper mainly uses total loan portfolio data.

This paper provides the first narrative account of banking system credit losses in Australia that includes both the early 1990s and global financial crisis episodes. Credit losses rise sharply during economic downturns, and are the main influence on banking system profitability during such periods. The Australian credit loss experience over the past three decades is dominated by two episodes: the very large losses around the early 1990s recession and the losses during and after the global financial crisis. During both episodes, banks' credit losses appear to have had a close relationship with changes in business sector conditions (such as commercial property prices and the business sector's interest burden). Losses during the earlier period totalled around 8½ per cent of lending; losses during and after the global financial crisis were around 2½ per cent of lending. The earlier episode was a more severe downturn – business sector conditions declined to a greater extent – but anecdotal evidence indicates that differences in lending standards also played a role in the different levels of credit losses across these two episodes.

As well as macroeconomic conditions and lending standards, portfolio composition turns out to be important for credit losses. The very limited portfolio-level data available for the early 1990s indicate losses during this episode were incurred mainly on business lending. The better data available for the global financial crisis episode make it clear that the elevated losses during this episode were almost entirely incurred on business lending. Credit losses on housing loans during the global financial crisis episode were minimal.

Other authors have applied econometric models to the ex post credit risk experience of Australian banks. Gizycki (2001) modelled bank-level measures related to credit losses – impaired asset and return-on-asset ratios – over periods that end in 1999. She found the interest burdens of the household and business sectors, real credit growth, the real interest rate, the share of construction in GDP, as well as commercial and residential property prices, to be the macro-level conditions that influenced credit risk measures. This is informative, but the dependent variables that Gizycki used do not have straightforward relationships with credit losses, so these conclusions are not directly transferable to credit losses.^[1] Hess, Grimes and Holmes (2009) did model credit losses, but did not consider some of the macro-level variables that Gizycki found to play key roles, particularly financial variables. Esho and Liaw (2002) is the only paper on credit losses in Australia that considers banks' portfolio composition. These authors use measures of portfolio composition from capital data as stand-alone explanatory variables in a model for credit losses over 1991–2001. They found residential mortgage lending to be indistinguishably risky from bank lending to governments, and much less risky than lending to businesses and (non-housing) personal lending.

The econometric models of banks' credit losses in this paper add to past Australian work in several ways. As the new dataset covers 1980–2013, they include both the early 1990s episode and the global financial crisis. They also consider a wide range of macro-level variables as potential explanators of credit losses. Most importantly, the main econometric model presented in this paper allows the effect of macro-level variables on bank-level credit losses to vary depending upon each bank's portfolio composition. An example of the underlying intuition is that a fall in the profitability of the business sector should lead to more credit losses (as a share of each bank's lending) for banks with a higher share of their portfolio devoted to business lending. This variability is achieved using interactions between bank-level portfolio composition variables and macro-level variables. This modelling strategy exploits the panel nature of the newly compiled credit loss dataset, as well as that of a regulatory dataset – the bank-level data underlying the aggregate measures of business, housing and personal credit. Interaction variables are clearly suggested by the available data on portfolio-level loss rates – which indicate losses on different portfolios respond differently to macro-level conditions – but a systematic approach of this type is novel in the literature. Pain (2003), Gerlach, Peng and Shu (2005) and Glogowski (2008) allow interactions between the share of one portfolio and a limited number of macro-level variables; I interact all macro-level variables with portfolio shares.

This model with portfolio interactions explains bank-level credit losses over recent decades reasonably well. The macro-level conditions that are statistically and economically significant are business sector conditions. As these variables are interacted with the shares of each bank's portfolio made up by business lending, this indicates business lending has been the main source of credit losses over recent decades. Analogous interactions between household sector conditions and the shares of banks' portfolios made up by housing or personal lending are not significant in the model. This result is consistent with the narrative account of credit losses in Australia over this period.

The econometric models in this paper do not explain all of the variation in credit losses. For example, they cannot explain why credit losses were so large at several state government-owned banks during the early 1990s. This accords with the omission of most of the variation in lending standards – roughly, the average riskiness of a bank's borrowers – from the models (quantitative measures that comprehensively summarise bank lending standards are not available). It also accords with anecdotal evidence that state government-owned banks had below-average lending standards. An alternative measurement strategy, based on quantile regressions, indicates that credit losses at banks with similar portfolios can respond very differently to macro-level downturns, providing further support for the importance of lending standards. While this evidence is not definitive, it suggests that poor lending standards may have been the cause of the very worst credit loss outcomes seen in Australia over recent decades.

As well as underlining the importance of lending standards, these findings have practical implications for the conduct of financial stability monitoring and stress testing. However, past performance does not necessarily predict future performance. A point of caution in projecting forward past patterns of credit losses is that the residential mortgage market has developed considerably since the early 1990s and now represents a much larger proportion of banks' lending activity.

The next part of this paper, Section 2, sets out the way I measure credit losses. Section 3 provides the narrative account of credit losses in Australia since 1980. Section 4 contains the econometric analysis of credit losses. Section 5 summarises my conclusions and discusses the implications for stress-testing practice and broader financial stability policy.

Footnote

As an example, impaired assets are not a sufficient statistic for credit losses. See Section 2.1 below. [1]