Leveraging Analytics to Improve Loss Forecasting
Losses on loans are a big cost of doing business for retail banks. These losses are reserved for by earnings that the banks have and are typically one of the key drivers of profitability. Banks need to account for future losses from current earnings in part due to regulatory requirements. Predicting losses is a key task that needs to be just right. Too high and you will over-reserve, leading to finances being underutilized and affecting earnings. Too low and you will not be able to cover losses which will impact profitability. Over the last 10 years, delinquencies have been ranging from 1.5% to 7.5% for all loans and leases. While individual banks have to weather localized challenges based on loan portfolios and macroeconomic conditions, the losses have significant volatility and impacted the bottom line of these banks. There are repercussions to future funding when loan losses are not accurately predicted as the banks credibility is lost. This can increase the cost of borrowing in the short term for banks and affecting overall returns.
Retail banks have had to improve their loss forecasts from an accuracy perspective, after having had to borrow from the FED during the economic crisis. The FED has required multiple stress tests and these have required banks to show the validity of their models as well as ensuring that their forecasts would be covered by their loss reserves. However, there have been many challenges to achieving that accuracy. There have been significant changes in the overall economy. Apart from that, regulatory changes in the wake of the economic crisis have changed loss reserve requirements. There are requirements to show adaptability to rapidly changing macroeconomic scenarios and the need for more granular loss forecasts. With all these requirements, there has been an increased need to look at more sophisticated approaches to forecasting losses.
Typical approaches to loss forecasting include roll rate models and vintage based models. These two approaches have been widely used and have had their fair share of successes. Roll rate models depend on historic movement of customers into various delinquency stages. These models use historic data of outstanding balances in these delinquency stages and predict the one-month forward movement using some historical average (like 3, 6 or even 12 months). The losses are based on the flow into the final delinquency (charge-off) bucket and can be predicted accurately with a 12 month range but are not very good going forward. Vintage based models track losses by grouping similarly aged sections of the portfolio. The losses of older vintages are calculated by month. More recent vintages’ losses are calculated for as many months as data is available and then extrapolated based on the loss curves from the older vintages. These forecasts are usually accurate over a long term in stable conditions. When conditions change rapidly, these forecasts break down significantly.
The above mentioned approaches do not account for economic factors. This is an important requirement as it is driven by regulatory concerns on the loss models not being very amenable to stress testing. To stress test models, it is important to show how losses react to changing macroeconomic conditions.
A more detailed approach is to look at loan-level loss models. These models are very granular and can account for macroeconomic conditions (locally as well as globally). Specifically, classification models are built for assessing the probability of the transition to the next delinquency stage given that they are already at a given stage. These models (typically Markov models) can incorporate product and customer level characteristics while also having local macroeconomic conditions in them. They can also incorporate the advantages of vintage based models by combining the similarly dated loans’ historical loss ratios to set curves. Given enough history (and more importantly, a rapidly changing macroeconomic environment), they are quite adept at accounting for these characteristics within the model. These models can then be used to aggregate transition of dollars through the roll-rate matrix, enabling good forecasts for the 18-24 month horizon. While the accuracy of these models is, at the minimum, comparable to other methods described above, their biggest advantage lies in being able to adjust to changing conditions rapidly. These allow product managers to also define better selection criterion in different macroeconomic conditions.
A key modeling improvement that can be bundled with this methodology is to incorporate account level forecasts for new business as well as recoveries. This can be significant again when changing economic conditions forces recoveries to fail first or even new business is not able to help sufficiently with improving loss ratios. A bank’s acquisition stage model can set the foundation for a new business component and the recoveries models can incorporate various elements of portfolio characteristics (like collateral, customer’s willingness to pay etc.) to improve the overall loss forecasting process.
Two key challenges need to be accounted for. The loan level models integration is a significantly complex exercise. In cases where portfolios are similar, or the requirements are during stable times, roll rate models are sufficiently useful for the required accuracy levels. With multiple models, there needs to be significant oversight into the modeling process to ensure that the right data and techniques as well as modeling granularity are used. This will ensure that the models themselves are robust overall in terms of loss forecasting. Another important issue is operationalizing complexity. The multitude of models can also create enough implementation challenges that would require dedicated IT bandwidth and investment.
While regulatory requirements might have forced the innovations in loss forecasting and increased the challenges in operationalizing them, it is of utmost importance to leverage these models for what they provide:
- Improved accuracy – With granular data, accuracy can be improved. Loss reserves can be set accurately and banks have the ability to adapt faster (on a quarterly basis) to changing macroeconomic and portfolio conditions.
- More visibility – The loan level models give better insights into key drivers of changes in losses. This will enable banks to understand whether challenges are internal or external, portfolio level or product level, customer level or geography level. This can allow banks to change strategies on the fly in terms of managing these losses well.
- Enhanced flexibility – Given that banks have to comply with various stress tests required by the FED, they now have a tool that will permit to forecasting for other scenarios. They can create new scenarios for not just loss forecasting, but in general managing their portfolios to decide where they should grow and where they should button down and manage their portfolio well.
Harsha Rao is currently an Associate Director at Mu Sigma
Tweet This