Since 2001, Norges Bank has used an empirical model, the SEBRA model (2), to estimate bankruptcy probabilities for Norwegian limited companies. The model is also used to estimate banks' expected losses on loans to enterprises in different industries. This article presents two new versions of the model: an extended version of the original model, and a basic version which makes less use of variables which correlate with the size of the enterprise. We show that the basic version is better suited to predicting and projecting banks' overall loan losses. However, the accuracy rate for bankruptcies is slightly lower at enterprise level. The extended version is better suited to analyses where the emphasis is more on bankruptcies than on aggregate loan losses.
Norges Bank's SEBRA model estimates bankruptcy probabilities using key figures calculated on the basis of enterprises' annual accounts, and information on their age, size and industry classification. Multiplying these bankruptcy probabilities by each enterprise's bank debt and then adding up the figures for all enterprises gives us an estimate of banks' expected loan losses due to bankruptcy, assuming that the entire loan amount is lost. Analyses based on such estimates are published regularly in Norges Bank's report Financial Stability and are included in its continuous assessment of the outlook for banks' financial strength. In analyses of enterprises' credit risk, we look at the situation both in different industries and in different regions. The SEBRA model is also used for projecting and stress testing banks' loan losses in various macro scenarios, for analyses of banks' pricing of loans to enterprises, and for assessing the potential effects of changes in the capital adequacy rules) Kredittilsynet (the Financial Supervisory Authority of Norway) uses bankruptcy probabilities from the model in its on-site supervision of banks and in its analyses of the state of financial markets.
This broad use of the SEBRA model has over time provided useful experience and ideas for further development over the years. In addition, access to data has improved since the model was developed. The original SEBRA model's accuracy rate for bankruptcy at enterprise level has been high and stable over time. The model also captures the surge in banks' recorded loan losses during the banking crisis of the early 1990s. However, the next increase in banks' loan losses, which came in 2002 and 2003, is not captured to the same extent.
In this article, we look more closely at various needs for the further development of the SEBRA model. We present two new versions of the model: an extended version of the original model, and a basic version which uses a smaller number of explanatory variables. After evaluating the accuracy and predictive power of these models, we describe briefly how banks' recorded loan losses can be projected. The article concludes with a summary.
The original SEBRA model in brief
In the original SEBRA model, the probability of bankruptcy is modelled mainly using key figures for an enterprise's earnings, financial strength and liquidity, see Eklund et al. (2001). Thus, the model's predictions are driven by quantities that reflect key business economic conditions at the individual enterprise. These will always be crucial for an enterprise's capacity to service its debt. Besides key financial figures, the model includes measures of an enterprise's size and age, and industry variables based on aggregates of the key financial figures. It is useful to differentiate between variables which reflect financial conditions and variables which are more indirectly related to these conditions but still contribute to the model's overall explanatory power. Examples of the latter are the level of tax payable, trade accounts payable and dividend provisions.
The model does not include additional information such as negative credit history, absence of auditor approval, or late or non-filing of annual accounts. This ensures that the model attaches more importance to the financial factors behind movements in risk, which is important given that the model's main purpose is to contribute to an understanding of movements in credit risk. Furthermore, it would be very difficult to project such variables. The model is also more stable, as experience shows that the registration quality of this additional information varies from year to year. The model does not take explicit account of historical variations in bankruptcy frequency between industries. These differences are instead represented through variables for industry averages and variances of basic key variables based on a detailed industry classification. In this way, changes in risk levels in different industries over time can be captured, and the model becomes less retrospective.
The need to further develop the SEBRA model
Long experience of the use of the SEBRA model has meant that we have discovered various weaknesses in it. In this section, we discuss the most important needs for improvement. There are also other reasons to reassess the model. For example, the way in which the explanatory variables are measured in enterprises' annual accounts may have evolved over time, due in part to new accounting rules. There may also have been changes in the registration of bankruptcies over time. Access to new and more data is another factor which makes the further development of the model desirable.
Better prediction of the risk of losses on loans to large enterprises
The risk of losses is not the same as the risk of bankruptcy. The original SEBRA model's accuracy rate for bankruptcy at enterprise level has generally been high and stable over time. In the original SEBRA model, size (measured as the logarithm of total assets) is included as an explanatory variable. It appears that small enterprises go bankrupt more often than large enterprises for given values of the explanatory variables. If this size effect applies less to the probability of a loan loss, it will be problematic using bankruptcy as a substitution variable for losses in a model that uses size as an explanatory factor. Such a model will overestimate the effect of size on defaults and losses. Small enterprises often have little bank debt in...