RDP 2015-06: Credit Losses at Australian Banks: 1980–2013 5. Summary and Policy Implications

Credit losses in Australian banking in the post-deregulation period have been concentrated in two episodes: the very large losses around the early 1990s recession and the smaller losses during and after the global financial crisis. They have a close temporal relationship with the economic cycle, peaking close to troughs in GDP during downturns. A narrative account attributes the key roles in driving credit losses to business sector conditions such as business indebtedness and commercial property prices. The available data on portfolio-level losses indicate that elevated losses during these downturns stemmed from banks' lending to businesses, rather than their lending to households. Data available from 2008 onwards indicate losses on housing loans barely rose (from very low levels) during the global financial crisis, even though housing prices and employment fell noticeably in some geographical areas.

One of the main contributions of this paper is an econometric panel-data model that properly controls for bank-level portfolio composition. This model indicates business sector conditions, rather than household sector conditions, have been the driver of domestic credit losses over the period studied. The relevant business sector conditions – interest burden, profitability and commercial property prices – are indicators of the ability of this sector to service its debts and of the value of the collateral behind these debts. As a corollary, the model indicates that most losses over the past three decades were incurred on banks' business lending, and household losses were largely unresponsive to economic conditions in that period. Unlike past work, these results are consistent with the narrative account of credit losses in Australian banking.

Descriptive accounts attribute the scale of losses during the early 1990s to poor lending standards, and the data support this. One piece of evidence, based on quantile regressions, indicates that changes in macro-level conditions have had very different impacts upon banks with similar portfolios (in terms of the shares of business, housing and personal lending). Most compellingly, standard models cannot explain the extremely high credit losses experienced at some state government-owned banks in the early 1990s. Given the anecdotal evidence that these banks had below-average lending standards, this is consistent with the conclusion that poor lending standards have caused the very worst credit loss outcomes over recent decades.

These conclusions have practical implications for stress testing. The credit loss models in this paper that use least squares estimation, and include bank-level variables, are unable to explain, and so unlikely to predict, the very worst credit loss outcomes. Many stress-testing exercises use similar (and in some cases simpler) econometric models (see, for example, IMF (2012)). As the worst credit loss outcomes are the most relevant when stress testing, this suggests that alternative models are needed. Covas, Rump and Zakrajsek (2013) show that a type of quantile regression (quite different to that in this paper) can provide out-of-sample forecasts that encompass the credit losses experienced by the US banking system during the global financial crisis. In an Australian context, Durrani, Peat and Arnold (2014) show that allowing variation in credit risk outcomes across banks, rather than applying the same average risk parameters to all banks, can lead to significantly larger loss estimates.

Stress-testing models could also be improved by incorporating better data on lending standards. The Federal Reserve collects and makes use of loan-level data on borrower characteristics in its annual stress tests of the largest US banks (Board of Governors 2014). This captures some aspects of the risk profile of borrowers; more work is probably needed to make it possible to systemise and accurately record banks' lending standards.

The historical experience of credit losses at Australian banks this paper describes should help to guide overall understanding of the credit risk they currently face. It supports a continued focus on the analysis of the financial health of the business sector (one output of this work is a chapter of the Reserve Bank's semiannual Financial Stability Review). As another example, credit loss measures appear to peak before asset performance measures, potentially providing an early signal of future improvement in financial system stability.

The lack of a historical relationship between household sector conditions and credit losses should be used cautiously in contemporary debates on the riskiness of housing lending. It indicates that the macroeconomic shocks experienced by the household sector during the past three decades have been small relative to the lending standards in place for housing lending over this period. Future macroeconomic shocks may, however, have a larger impact on households. There have been, for example, no large nationwide falls in house prices during recent decades. In addition, a rise in unemployment on par with that in the early 1990s could be expected to have a more severe influence on household credit losses, given the large rise in household indebtedness over the intervening period. A corollary of this rise in household indebtedness is the greater share of banks' lending now made up by housing and personal lending. These considerations suggest that any weakening in lending standards in these areas could have a larger systemic impact than in the past.