Presentation by AIS\\’ CEO, Mr. Ramon Trias, at GARP\\’s chapter "Stress testing: new approaches for a new paradigm" (New York, May 13, 2010)
8. RDF Method- Risk Dynamics into the Future Scenario Asset structure Macroeconomic model Micro models (portfolios) Conditional distribution Losses given risk drivers Global Risk measures “ Cul d’olla” Integral Partial instances: GDP 4th Q 2010=xxx && Unemployment 3rd Q 2010=yyy alternative structure Information & Decision system Intrinsic Risk correction Conditioned Economic Capital CVar,Cshortfall,CUL, Risk Contribution VAR-models Segmented Micro-models Marginal distribution Consolidated Distribution
Good afternoon, Ladies and Gentlemen I would like take advantage of this moment to thank to GARP for the opportunity to giving this conference and the audience for the interest in this subject. I am sure it is going to be useful to you. In the next fifteen minutes, I am going to present some experiences and analysis around the use of stress testing to support good decisions on management and strategic plan, helping analysts and managers to improve their position when faced with an eventual crises and extending their effect to economic capital calculation, pricing assessment and reporting. I will overview the new method that supports it. At the end, we will have fifteen minutes more for questions and answers, please, keep in mind we have a short time for this complex matter, I invite you to discuss it in a more informal way over cocktail.
We will start with some reflections about models and crisis, Trying afterwards to identify some drawbacks on currents methods of calculation –Vasiceck model, Lee model, Basel II … It will be followed by a short survey of a tool designed to avoid drawbacks and mistakes Concluding with a summary approach to the future of methods like RDF-AIS
The crisis has been very vast and deep. Lehman Brothers, Stock markets crash, distrustful models, Mr. Li’s formula, government rescues, Greece… As you know, we have as players bankers, financial analysts, rocket scientist, fat cats, supervisors and neo-classic economists and everybody is blaming each other. The rating agencies are not very happy either, but in general you have to be carefully when speaking about them. Without aiming to analyze this mess, as that could be very dangerous, I believe that we should make an interior, humble and honest look to shed some light on what has failed in the battle front that we defend and, if some simplification has served other players to justify their actions, to evaluate how we can improve the calculation methods and the control of its application as an aid to take the correct decisions. A complementary note, in the other end: at the beginning of the XIX century there was in Nottingham, UK, an important social movement called “luddites” whose followers destroyed the machines to consider them guilty of the loss of jobs for many skilled textile workers. Watch out!, there are desperate neo-luddites among us.
Some biases are implied in currently used methods such as Vasicek, Basel II, Ho & Lee Formula, Merton formula: Non-measurable uncertainty components A principle relay behind the models: all events that can happen in the future, have a common origin with the past, usually is assumed a distribution of the Exponential Family (Normal, Bernoulli, Negative Binomial, Gamma, …). In the current crises, chaos were foreseen by experts and badly predicted by models. Non-symmetrical interactions among risks Risk of the same source can produce changes into variables with diverse time-lags. Moreover, correlation as a measure throughout all cycle hide the simultaneity of default in bad times. A spurious diversification can hide a non-contemporaneous dependence Non observable correlations Measuring correlation of bond’s returns can be at least feasible, but information on retail loans cannot be obtained seemingly, it means the concept of risk is equal but the way to obtain their signals are very different. Poor estimation of correlations between investments implies poor portfolio design. Missing significant idiosyncratic risk. Since the fifties, the theory of investment, the Portfolio Selection, the CAPM and the like, has been teaching us that the idiosyncratic risk vanishes with diversification, so, it can be negligible under some hypothesis. The lack of granularity (Herfindal coefficient high) mainly when produced by a mix of granular elements together with a very short number of very big investments produces fat tails loaded by the not-vanished idiosyncratic risk and, as a consequence, underestimation of capital needs and risk contribution and a false (and dangerous) measurement of solvency Lack of scalability and integration It is not easy to take into account the consequences of changing the structure of our portfolio in terms of losses distribution function using formulas that deals with the hypothesis of inelasticity of idiosyncratic risk. It can hardly be used as a tool for strategic planning if it is inelastic respect the change in portfolio volume and structure. Poor estimation of the diversification effects Mainly in low or strongly irregular granularity International Institutions can compensate the loss of credit quality operating outside the natural markets by the positive effects of diversification (if it is well evaluated) Slow answers, poor dialog user-algorithm When Montecarlo is applied, the answers need more and more processing time. Simulating different portfolio structures when faced with different scenarios, needs a fast response. Moreover, the foreseeable increase in capital regulatory ratios, will change the question “How much capital need” by “which is the best combination of investments subject to capital limitations”. A mathematical optimization of this kind, need a very fast response in each single evaluation.
In general, the risk models look for an equilibrium between expected value and variability, but in the real world these decisions are tied to the bad case. Gathering capital is another example. Solvency capital makes sense if able to cover a stressed situation, no matter related with the average thru the cycle.
A clever design on new capital and stress testing concepts can help improving the former frame: Not-measurable uncertainty components Expert or extra-model Scenarios carries the uncertainty No model can predict chaos, no expert can draw a detailed scenario and the path to reach it. Merging models and extra-models scenarios gives risk and uncertainty integrated. Not-symmetrical interactions among risks Multi-period and multi-equation models, macro and micro covers identifiable relations Effect of Macroeconomic variables on losses at each portfolio and extra-model scenarios, both defined with the most representative variables at some data, left few chances to be complete and coincident, so, a model embedded with the temporal structure of them will be necessary. VAR models fits well, integration with scenarios can be treated as a marginal distribution. Micro economic models –PD, LGD, EAD …- can be produced with current functions – Logistic, Logarithmic, Merton, … Not observable correlations in extreme events Correlations can be obtained as an output: Conditioned to a scene Correlations conditioned to events of stressed scenarios can offer more information (even being loosely defined) that correlations along the complete cycle. Moreover, instead of correlations on single risk drivers, total losses correlations is more useful. This Matrix Correlation should be obtained as an output, not as an input to the calculations. Missing significant idiosyncratic risk Poor estimation of the diversification effects Control of the idiosyncratic risk as well of the shared one gives a realistic estimation of fat tails Effect of the residual risk can be calculated Conditioned to a scene Lack of scalability and integration Taking into account non-lineal effects of volume on probability losses distribution Slow answers, poor dialog user-algorithm One to five seconds per simulation on a medium sized retail bank
To define an scenario means to fix some variables in some periods to the guessed figures given by the expert and optionally the variance around them. The modified VAR model, estimates the rest of variables with the most probable values and their correlation matrix, both of them conditioned to the economic scenario. As a consequence, after this step we have the PDF of all the variables by all the periods in the horizon of analysis. RDF has been designed to integrate this information with the micro models, the output is the losses distribution of probability, from which it can be obtained VaR, Shortfall, Expected Losses, Unexpected Losses and required Capital. All this risk measures are calculated conditioned to the scenario defined in the first step.
We observe that losses and Point In Time VaR have a cyclical variation with a fixed capital structure. In the stress periods Losses can be more greater than the Long Run VaR. The structure of whole portfolio is decomposed in 7 portfolios. An optimization of structure between portfolios (OPTIM) is made to reduce the VaR in streed periods. Currently, RDF is used by Bancaja to Re-structure of their Credit-Portfolio Price management of Real State assets Commercial planning Strategic Investments Plan Other case: Financiera Rural , a big Mexican Development Bank Credit Insurance Capital Calculation for Development targets
Credit risk analysis Concentration / diversification Limits by debtor and portfolio Marginal risk for debtor and portfolio Risk premium and RAROC by debtor and portfolio Risk management Solvency and stress Provisions forecasting Portfolio structure planning Securitization valuation Admission & follow-up Scoring limits based on objective PD for the next 3 years Back testing Strategic Planning “ Bottom-up” estimation objectives by branch, subject to macroeconomic situation