Stress testing & scenarios analysis
- Thread starter ajsa
- Start date
Hi ajsa,
Good question. I am not aware of a consensus distinction; e.g., Basel II refers to "stress scenarios" and some authors seem to use them synonymously. Although I read L Allen as suggesting that one difference is, scenario analysis can be used overcome the "correlation breakdown" (better than stress testing).
As you suggest, Jorion (VaR assigned Ch 14, Stress Testing) consistently has scenario analysis as a subclass of (type of) stress testing: in Jorion, stress testing includes: (1) scenario analysis, (2) stressing models, and (3) policy responses. As in: stress testing includes the generation of several scenarios. So, my take on Jorion is that scenario anaylsis is one means to an end (stress testing ).
I checked BIS' recently issued Principles for Sound Stress Testing (see here for my notes on this) and they have the following (interesting, I think) paragraph (emphasis mine): "A stress test is commonly described as the evaluation of a bank’s financial position under a severe but plausible scenario to assist in decision making within the bank. The term “stress testing” is also used to refer not only to the mechanics of applying specific individual tests, but also to the wider environment within which the tests are developed, evaluated and used within the decision-making process. In this paper, we use the term “stress testing” in this wider sense."
In regards to my own personal connotation, it seems to me stress test connotes, and includes, mechanical shocks of risk factors (sensitivity) and sets of risk factors; but scenario analysis (which has sub-categories) connotes "narratives that are quantified" (e.g., historical, hypothetical). So, we may collect a set of shock factors together (e.g. + yield, + volatility) and that does not imply we have a logic for the simultaneous shock under a stress test. But (it just seems to me that) scenario analysis implies that we least has a narrative/logic/story as to why we are shocking a set of risk factors (e.g., we shock this set of risk factors b/c we imagine it possible under a hypothetical scenario).
Is the correlation breakdown a limitation or a strength of scenario analysis? what are the “implications” as the AIM indicates?
The correlation breakdown refers to the general problem is using historical data to calibrate going-forward loss estimates (includes historically informed volatility into a parametric VaR, or historical simulation, or monte carlo informed by historical data). So, I think Linda Allen's point is: stress testing/scenario analysis needs to somehow address/hand correlation breakdown (namely, correlation spiking in crisis). She writes, "A simulation using SMC based on a covariance matrix that was estimated during normal times cannot generate scenarios that are economically relevant for times of crisis. A common, but potentially erroneous, remedy is to increase the number of simulations and go further out in the tail to examine more “severe stress.”
In short, the "implication" is "be sure to account for correlation breakdowns even as you don't see them in historical data." I think she implies that this is a difference btwn stress tesing and scenario analysis; i.e., scenario analysis "stresses" the risk factors but also incorporates the "narrative" or correlation breakdown/spike: "Generating a larger number of simulations cannot remedy the problem. As we will see in the remainder of this section, scenario analysis may offer an alternative that explicitly considers future events."
sorry for length, reflects my lack of decisiveness (but still interesting!)
David
Good question. I am not aware of a consensus distinction; e.g., Basel II refers to "stress scenarios" and some authors seem to use them synonymously. Although I read L Allen as suggesting that one difference is, scenario analysis can be used overcome the "correlation breakdown" (better than stress testing).
As you suggest, Jorion (VaR assigned Ch 14, Stress Testing) consistently has scenario analysis as a subclass of (type of) stress testing: in Jorion, stress testing includes: (1) scenario analysis, (2) stressing models, and (3) policy responses. As in: stress testing includes the generation of several scenarios. So, my take on Jorion is that scenario anaylsis is one means to an end (stress testing ).
I checked BIS' recently issued Principles for Sound Stress Testing (see here for my notes on this) and they have the following (interesting, I think) paragraph (emphasis mine): "A stress test is commonly described as the evaluation of a bank’s financial position under a severe but plausible scenario to assist in decision making within the bank. The term “stress testing” is also used to refer not only to the mechanics of applying specific individual tests, but also to the wider environment within which the tests are developed, evaluated and used within the decision-making process. In this paper, we use the term “stress testing” in this wider sense."
In regards to my own personal connotation, it seems to me stress test connotes, and includes, mechanical shocks of risk factors (sensitivity) and sets of risk factors; but scenario analysis (which has sub-categories) connotes "narratives that are quantified" (e.g., historical, hypothetical). So, we may collect a set of shock factors together (e.g. + yield, + volatility) and that does not imply we have a logic for the simultaneous shock under a stress test. But (it just seems to me that) scenario analysis implies that we least has a narrative/logic/story as to why we are shocking a set of risk factors (e.g., we shock this set of risk factors b/c we imagine it possible under a hypothetical scenario).
Is the correlation breakdown a limitation or a strength of scenario analysis? what are the “implications” as the AIM indicates?
The correlation breakdown refers to the general problem is using historical data to calibrate going-forward loss estimates (includes historically informed volatility into a parametric VaR, or historical simulation, or monte carlo informed by historical data). So, I think Linda Allen's point is: stress testing/scenario analysis needs to somehow address/hand correlation breakdown (namely, correlation spiking in crisis). She writes, "A simulation using SMC based on a covariance matrix that was estimated during normal times cannot generate scenarios that are economically relevant for times of crisis. A common, but potentially erroneous, remedy is to increase the number of simulations and go further out in the tail to examine more “severe stress.”
In short, the "implication" is "be sure to account for correlation breakdowns even as you don't see them in historical data." I think she implies that this is a difference btwn stress tesing and scenario analysis; i.e., scenario analysis "stresses" the risk factors but also incorporates the "narrative" or correlation breakdown/spike: "Generating a larger number of simulations cannot remedy the problem. As we will see in the remainder of this section, scenario analysis may offer an alternative that explicitly considers future events."
sorry for length, reflects my lack of decisiveness (but still interesting!)
David
Hi David,
I have a doubt about whether “Stress Testing” is a Qualitative or Quantitative measure?
On reading “Implementing Robust RAF” it says that VaR has drawbacks such that it does not account for nonquantifiable risks, that is why stress testing is needed to provide a “qualitative view”. But later, on chapter “Range of practices and issues in economic capital frameworks”, stress testing is included in the “six quantitative validation processes to consider”.
Is it considered a Qualitative or Quantitative measure?
Thanks!!
I have a doubt about whether “Stress Testing” is a Qualitative or Quantitative measure?
On reading “Implementing Robust RAF” it says that VaR has drawbacks such that it does not account for nonquantifiable risks, that is why stress testing is needed to provide a “qualitative view”. But later, on chapter “Range of practices and issues in economic capital frameworks”, stress testing is included in the “six quantitative validation processes to consider”.
Is it considered a Qualitative or Quantitative measure?
Thanks!!
Similar threads
