The essence of loss given default - a factor that should not be overlooked

 

How do lending institutions control losses while expanding into such a retail market? Though measuring credit losses is indispensable, it has received considerably less attention from lending institutions over the years. One of the key reasons is the analytical complexity associated with maintaining loss data and mastery of methods that permit robust forecasting. 

Banks and lending institutions in Asia-Pacific rest at different levels of the Readi-ness Matrix in relation to their credit-loss data and ability to undertake cutting-edge forecasting and modelling. Things have evolved for the industry since 2000.

As we know, the three workhorse credit-risk parameters that comprise provisioning against loan losses are the probability of default (PD), loss given default (LGD) and exposure at default (EAD). 

Of these, PD almost always commands centre stage. Yet to add to its power, the soon-to-come IFRS 9 accounting standards have incorporated the concept of lifetime expected loss – once there is a significant increase in the credit risk of obligors, the default estimate (expected loss projection) will need to be revised to the lifetime expectation. 

On the other hand, LGD, the proportion of actual loan loss in the event of a default, always plays a lesser role and has been treated as non-essential in playing the provisioning role. To the authors though, LGD modelling is considered as crucial going forward, as it adds considerable merit to the consistency of the loss model estimate. 

Among its benefits, it helps project recovery more efficiently and quantifies the effectiveness of a bank’s recovery management system and strategy. 

The process of data collation and computation of LGD at the facility level involves the consistent tracking of recoveries made on each delinquent facility and discounting of both the recovered amounts and the cost amounts to arrive at the net discounted recovery and LGD on the facility. 

Computation of the LGD metric is heavily contingent on the quality of recovery data maintained at the level of each credit facility. It needs to be noted here that recovery experience varies significantly from lender to lender and is both a function of the time elapsed since the facility entered delinquency and the effectiveness of the recovery strategy of the lending institution. 

Though the LGD metric would almost always lie in the unit interval of (0, 1), its shape when analysed across all defaulted facilities within a lender’s portfolio would vary from institution to institution and by region of operation – depending largely on the timing, type of recovery intervention and the nature of the credit facilities – besides certain other characteristics.

 

Loss-Based Segmentation – |historical experience

Banks and lending institutions have often tried to analyse their historical loan loss databases in an attempt to find pockets of their loan portfolios that contribute to higher LGDs. Lenders have examined a set of characteristics as proximate “determinants” of LGD. These include the obligor’s industry, demographics, facility type and collateral details. 

However, success in creating robust forward-looking profiles of facilities differentiated by their expected LGD seems to be less than modest. The primary difficulties in this domain are the depth of technical expertise needed to address the widely varying recovery experiences, data quality and therefore the modelling approach tailored to each lending entity.

 

LGD – shape and challenges 

Data analysis of any risk metric is driven by the likely set of values it is expected to take – characterised by its probability density function. Conventional forecasting techniques largely require a degree of symmetry in the shape of this function, indicating that while all defaulted loan facilities tend to centre around a certain value/range of LGD, there are a few outliers, with a roughly equal proportion of such extreme cases around LGD=0 and LGD=1.

However, many lending institutions have LGD data that do not conform to that expected shape. In fact, we may observe a large proportion of defaulted accounts having an LGD=0, with very few having values above 0. This is true of lenders that have over time optimised their recovery strategy such that the loss is minimal from accounts that enter a default bucket. 

 

LGD – profiling and modelling 

Developing statistical models for LGD consists of taking cognisance of the peculiarities of the nature and shape of LGD. The associated experience and expertise are at the heart of designing and successfully validating such profiles, bringing about measurable benefits to lenders, which can pre-emptively design their credit risk management strategy in accordance with these solutions.

Also, LGD-based profiling of facilities and statistical modelling facilitate the generation of readily applicable solutions under more complex scenarios. These include portfolios wherein the observed LGD density function comprises multiple peaks – usually resembling a bi-modal shape.

A reason for this is the presence of two very large clusters of facilities – one having an LGD=0 and the other having an LGD=1, with the remaining facilities lying in between. 

Such models and predictive techniques add onto themselves layers of complexity, but as a result promise sharper predictions of LGD at the level of each credit facility, enabling more accurate loss management strategies for banks to be implemented with ease time and time again. 

 

The authors are Tana Tanarugsachock, director of enterprise risk services, and Sayan Sen, manager of enterprise risk services, at Deloitte Touche Tohmatsu Jaiyos Advisory Co.