Forecasting Portfolio Performance in an Uncertain Economy

Transcription

1 Economic Environment Forecasting Portfolio Performance in an Uncertain Economy In an uncertain environment, having a forecasting process in place to assess risk makes good business sense. by Jeffrey S. Morrison We a r e in the midst of a severe recession. Some are even venturing to call it the Great Recession. If only we had known! Where were the warning signs? Can the future be predicted? Would financial institutions have benefited from an in-house forecasting function highlighting potential weaknesses? If history teaches us anything, it is that the economy is susceptible to ups and downs. In fact, they are inevitable. Since the 1970s, we have experienced seven recessions, including the current one. The average length has been a little under one year 1 (Table 1). The current recession already has exceeded the average. The dramatic rise in mortgage delinquency rates since the end of 2007, the official beginning of the recession, highlights a worthwhile comparison with the last downturn. The 2001 recession, which began in March and ended in November of that year, resulted from a collapse of the dotcom bubble combined with the terrorist attacks of September 11. During that time, the mortgage delinquency rate went up almost 28%, considered a large increase at the time. Today it might be viewed as modest. During the current recession, the national average mortgage delinquency rate had increased more than 100% as of March Although the characteristics of each recession vary, common among them is a substantial increase in unemployment that may not recede until after the downturn ends. The current economic environment best mirrors the recession, when unemployment hit 10%. Many economists forecast the current unemployment rate to reach that level or greater by the end of The Value of Forecasts Without a doubt, the ability to predict the future is priceless, although such a task is never easy. Regardless, in its compliance requirements, the new Basel Capital Accord highlights the need for a forecasting process, advocating benchmarking and stress tests for computing loss reserves. Even if Basel compliance is not relevant to your organization, being able to assess risk in an uncertain environment simply makes good business sense. For example, a bank looking to move into new markets would be interested in geographical areas that reflect a lower risk environment relative to the traditional footprint. Would the macroeconomic environment of a particular metropolitan area better insulate some consumers from continued shocks in the credit markets? Do trends in the average credit scores for a select group of counties show significant differences over the past 10 years? What about average balances, credit utilization, or delinquencies? In our current recession, what areas of the country are most affected by delinquency or bankruptcy? If those areas had been identified ahead of time and trend scenarios made, would the bank have been able to implement policies to offset some of its losses? These are just a few of the strategic questions that could be M. Dykstra/Shutterstock 52 July August 2009 The RMA Journal

2 Table 1 Recession Lengths Since 1969 Recession Months Dec Nov Nov March Jan July July 1981 Nov July 1990 March March 2001 Nov Dec. 2007?? asked to support potential corporate marketing and risk mitigation efforts. The purpose of this article is to give risk managers a taste of the tools and techniques used for portfolio forecasting, while pointing out similarities to traditional credit-scoring models. Some techniques are relatively simple. Some are more complicated and require substantial data collection. All are practical and can be done with new user-friendly statistical software packages such as SAS, EViews, and Forecast Pro. Different Data Let s begin with the fundamentals for forecasting: the data. Much has been written about predicting delinquencies in the credit industry. Most of the literature focuses on the financial institution s ability to predict whether an individual account will become 90-or-more days delinquent in the next 18 to 24 months. Although the exact timing of the delinquency can be addressed at the individual account level by the tools and techniques of survival analysis, quantifying aggregate portfolio credit behavior often is left to econometricians who deal with time-series data. Risk managers can use the same econometric tools, however, to provide realistic forecasts of loan performance at the portfolio level. Because risk managers may not be comfortable using time-series data, let s provide an orientation. Time-series data is information aggregated over time to a grouping above the individual account level. Examples include the national bankruptcy rate over the past 10 years or the average bank card balance between 1999 and 2007 for California. Although traditional credit-scoring techniques rightly make use of the advantages of individual credit information, they focus on tactical and operational objectives. If strategic objectives are in play, aggregating credit information over time offers advantages not available otherwise. Keep in mind that historical data used in time-series forecasting may not be fixed in stone. Many economic time series are really estimated actuals, meaning the economic service provider (government or private) has to make educated guesses to fill gaps in data availability and reporting. For example, revisions are common in series such as unemployment and gross domestic product. Forecasts built using the same predictors over the same time periods may come out differently, depending on the extent of data revisions. Often, these changes are substantial and reach back one or two years. Different Modeling Techniques The technique of choice in building credit-scoring models is regression. Regression is an efficient way to compute weights for the prediction algorithm where correlations are consolidated into one succinct formula. For credit scoring, logistic regression often is selected because it works with individual-account data from two populations good and The RMA Journal July August

3 Table 2 State Date Delinquency Rate GDP Cross-sectional Time-Series Data Disposable Income Unemployment Rate Credit Limit Utilization Average Score Alabama 2005Q $10,914 $117, $78, Alabama 2005Q $11,002 $119, $79, Alabama 2005Q $11,115 $120, $78, Alabama 2005Q $11,164 $123, $78, Alabama 2006Q $11,316 $124, $79, Alabama 2006Q $11,388 $125, $81, Alabama 2006Q $11,444 $127, $81, Alabama 2006Q $11,507 $129, $82, California 2005Q $10,914 $1,141, $146, California 2005Q $11,002 $1,152, $151, California 2005Q $11,115 $1,169, $157, California 2005Q $11,164 $1,188, $160, California 2006Q $11,316 $1,211, $166, California 2006Q $11,388 $1,213, $169, California 2006Q $11,444 $1,231, $175, California 2006Q $11,507 $1,243, $175, Colorado 2005Q $10,914 $153, $137, Colorado 2005Q $11,002 $154, $140, Colorado 2005Q $11,115 $157, $141, Colorado 2005Q $11,164 $159, $142, bad accounts. However, other regression techniques are available to handle time-series data in cases where our goal might be to predict aggregate trends for example, portfolio delinquency rates. Regardless, the analyst still goes through the same modeling steps in doing exploratory data analysis, determining the best variables to use as predictors, computing the weights, producing the forecast, and evaluating predictive accuracy. Unlike in credit-scoring models, the regression techniques associated with predicting aggregate time-series behavior require forecasts for each predictor in the model. This is especially true if one wants to provide forecasts for one or two years into the future. Some predictors might reflect leading behavior that is, changes in these variables occur before they find their way to delinquencies. For example, changes in the unemployment rate may lead credit behavior by one or more quarters. Other variables may serve as coincidental Capturing turning points is a primary strength of a regression framework, as well as its capacity to construct whatif analysis, scenario planning, and stress tests. indicators of delinquency, with effects occurring in the same period of time as delinquencies. In constructing a portfolio forecasting process, the risk manager must weigh the pros and cons of obtaining forecasts for these economic or demographic variables, while considering strategic objectives and the resulting incremental benefits. Many benefits of a regression-based time-series model are found in its medium- to long-term predictive power often two or more periods out over the forecast horizon. This is because it is a causal model where the determinants of the series have been identified structurally. For better short-term accuracy (one or two periods out), other techniques such as ARIMA or exponential smoothing often are used. They are simple to build (often automatic) and do not require external predictors, but they are incapable of capturing turning points, such as an abrupt upward shift in delinquencies in a new recessionary environment. Capturing turning points is a primary strength of a regression framework, as well as its capacity to construct what-if analysis, scenario planning, and stress tests. Sample design for credit-scoring models requires credit attributes to be pulled at what is called the observation period. Payment performance is calculated over the performance window usually 18 months to two years. Time-series 54 July August 2009 The RMA Journal

4 forecasting models need a longer stream of continuous data both predictor information and performance. Many econometric textbooks recommend having at least10 years of quarterly data. However, most financial organizations do not keep information that long. The good news is that these data requirements can be minimized if the analyst uses a particular type of time-series model called a pooled cross-section time-series regression. Pooled models capture the best of both worlds. From a cross-sectional perspective, they allow simultaneous estimation of models for a collection of geographies (or loan vintages), while looking at how the forecasted series varies over time. The result is a more robust model where weights better capture, for example, how California differs from Georgia while incorporating correlations over time. To quote a famous econometrics textbook: The combination of time series with cross sections can enhance the quality and quantity of data in ways that would be impossible using only one of these two dimensions. 3 To understand this type of information, a picture is worth a thousand words. Although the process used to aggregate the data is not trivial, the end result can fit onto an Excel spreadsheet, as illustrated by the fictitious example shown in Table 2. Finally, the pooling framework offers the risk manager additional tools to satisfy a variety of modeling assumptions and to impose consistency across forecasts. For example, why should the employment rate and disposable income be used as part of the forecast equation for California but not be included for Florida and Montana? Imposing these kinds of consistencies often provides more realistic forecasts and makes it easier to explain the models to senior management. Going from a regional-level forecast to national projections can be done through a simple weighting based on the number of consumers or accounts in each state. Variable Selection As mentioned earlier, a pooled cross-section time-series framework allows us to simultaneously forecast, for example, the average mortgage delinquency ratio for each state in the U.S. However, a key to success is finding the proper set of predictor variables. This is typically done in a manner very similar The combination of time series with cross sections can enhance the quality and quantity of data in ways that would be impossible using only one of these two dimensions. to credit scoring, with a few caveats. In a credit-scoring model, the credit attributes are not generally transformed. They are entered into the regression in their normal form. In a portfolio forecasting model, the analyst has the option to model changes in both the predictor variables and delinquency, or enter them in log form (which has some practical advantages), or use a combination of both. Furthermore, the analyst must determine whether the predictor would be better if it were specified to lead credit performance. If a leading specification is selected, the analyst can then decide how many leading periods will be most effective. Luckily, some of these decisions can be determined by simple graphics. As seen in Figure 1, house prices in Florida begin to change before the rise in delinquency, making it a Figure 1 Florida: Mortgage Delinquency vs. Home Prices Actual Delinquency Ratio (Source: TransUnion) House Price Index (Source: Economy.com) The RMA Journal July August

5 Figure Florida Ex-Post Forecast (2008:Q3 to 2008:Q4) Actual Mortgage Delinquency Ratio Predicted Mortgage Delinquency Ratio good candidate for a leading indicator. As in credit-scoring models, however, only variables that have well-formulated theories behind them should be included in the model. Fortunately, as in credit-scoring models, statistical tests are available to determine whether these results were merely a coincidence or if systematic behavior was being observed. For illustrative purposes, let s build a model to predict the ratio of mortgages 60-or-more days past due based on house prices, vacancy rates, percentage of adjustable rate mortgages, unemployment rates, consumer confidence, and disposable income per capita. An advantage of building our forecasting model, other than to provide a forecast, is to better understand the drivers of delinquency. For example, what is the current role of the unemployment rate and has it changed over time? This can be answered by changing the time periods for which the regression model is estimated and looking at how the coefficient (or weight) of each variable changes. Interestingly, the model as described above showed a significant drop in the size of the coefficient for unemployment as we add years of information say, , , and The implications are meaningful. The unemployment rate, although still very important, is not as strong a driver of mortgage delinquency today as it was before the current recession. The reason: There is a relatively new player in town. The abrupt declines in regional house prices occurring in the summer of 2007 have had a growing impact Many consumers who once showed positive equity on their homes now owe more than what their houses are worth. on consumers ability to repay their debt obligations. Many consumers who once showed positive equity on their homes now owe more than what their houses are worth. This is one fundamental reason why today s recession is different from that of Now, increases in unemployment are feeding the delinquency problem, adding insult to injury. Before leaving this topic, it is important to note another advantage of building econometric models other than to provide a forecast. Credit managers need rules of thumb to better articulate the cause-and-effect relationships between the economy and their policy decisions and the implications for risk. The good news is that the forecasting model can easily produce these rules by using a log transformation on all variables when the model is estimated. When structured this way, the coefficients of each variable can be interpreted as elasticities. For example, if the coefficient for disposable income in such a logged model is 0.65, this implies that a 1% drop in income would tend to increase the delinquency ratio by 0.65%, all other factors remaining equal. Validation Now that we have a model, how do we test its accuracy? Best practice in the design of credit-scoring models is to hold out a sample of observations from model development where you have known performance and see how your model works in distinguishing good from bad accounts. Typically, a KS (Kolmogorov-Smirnov) statistic for measuring accuracy is computed ranging from zero to 100, in which 100 reflects a model that is perfect. Measuring forecast accuracy for time-series models is similar, but differences exist. For time-series models, the holdout period is usually at the end 56 July August 2009 The RMA Journal

6 An interview with TD s CRO Mark Chauvin_p.12 Interdependent risks must be managed on an enterprise-wide basis_p.22 May 2009 of the historical series (one to four periods). The forecasts for these holdout periods are then compared to actual history and the mean absolute percent error (MAPE) is calculated. The smaller this value, the more accurate the forecast. So how small should the error be? That depends on the general volatility of the series. If the historical fluctuations are not excessive, you can get your MAPE down to 1% or less. If the history shows substantial volatility, making the series more difficult to explain, a MAPE of 5% to 10% (or more) may be reasonable. In credit-scoring models, a benchmark comparison often is helpful. Most credit scores are compared to a score made up of random values, implying a comparison to no information. The true credit score should provide substantial lift over a random score if the model was developed properly. For time-series models, such a comparison might be done using a regression model, with only a simple time trend built as the challenger. Over a long period of time, one would hope that the MAPE from the econometric model would be smaller than that derived from the simple time-trend model. Figure 2 shows a comparison between actual and forecasted mortgage delinquency for our Florida example. The gray area represents the holdout period, called the ex-post forecast period in the statistical literature. This says that, given known values for all variables in the model, the forecasts would be very accurate. However, forecast error is sometimes on the optimistic side some accuracy is lost due to errors associated with predicting driver variables such as unemployment and income. Conclusion As in credit scoring, portfolio forecasting is as much art as science. Although we would like a crystal ball to help us develop the perfect prognostication, such a tool is neither realistic nor even necessary. In the run-up to the recent recession, a financial institution with strategic processes in place to continually evaluate the determinants of mortgage delinquency, from both a theoretical and empirical perspective, might have picked up on the likelihood that some consumers would not be able to pay their ever-increasing monthly mortgage bill, given the relatively slow growth in real incomes. Although the exact timing of the tipping point, as well as the full impact associated with the complex world of mortgage-backed securities, might have been elusive, financial institutions could have identified areas in the country that were potentially at greatest risk if a meltdown occurred. Scenarios could have been developed not only from a mortgage perspective, but also across other portfolios such as bank card, auto, and installment loans. Supported by a set of strategic forecasting models and an adequate corporate planning function, simulations, stress tests, and contingency plans could have been readily available for implementation to mitigate losses when the housing bubble began to burst in the summer of v Jeffrey S. Morrison is manager of research and economics at TransUnion. Contact him by at jmorris@transunion.com. Notes 1. TransUnion s Trend Data. 2. National Bureau of Economic Research. Although we would like a crystal ball to help us develop the perfect prognostication, such a tool is neither realistic nor even necessary. 3. Gujarati, D. Basic Econometrics, 4th ed., pp New York: McGraw-Hill, Letters to the Editor The RMA Journal welcomes letters from our readers. Letters can be ed to editor@rmahq. org, or mailed to Kathleen M. Beans, Editor, The RMA Journal, 1801 Market Street, Suite 300, Philadelphia, PA We look forward to hearing from you. E N T E R A April 2009 T Y O U R O W N R I S K FINANCING MEDICAL OFFICE BUILDINGS REQUIRES LENDERS TO ACCOUNT FOR SPECIAL REQUIREMENTS_p Staffing Community Banks Community bankers foresee challenges in attracting and keeping qualified staff_p. 18 The Five Cs of Credit Character, one of the five Cs, 8 9 deserves closer scrutiny_p.34 Why Are Canadian Banks the Envy of the World? Key Requirements for ERM 1 Construction Lending in a Recession Banks should take a second look at their construction loan portfolio and practices. Regulators Take a Hard Look at Risk Bank regulators are calling for comprehensive, forward-looking ERM programs. June 2009 The RMA Journal July August