Stochastic Methods The Power of Numbers Presented by Roger M. Hayne, PhD, FCAS, MAAA CAS Spring Meeting 16-18 June 2008 Quebec City, Quebec Why Bother with Stochastic Methods? We all know stochastic methods are: – Complicated – Black boxes – Leave no room for actuarial judgment – Are impossible to describe – Take way too long to implement – Take far more data than we could ever imagine obtaining – Don’t answer the interesting questions – Bottom line, inconvenient and too complex for their own good Well, some of you might know that, others (yours truly included) need to be convinced 2 July 26, 2016 What is A Stochastic Actuarial Model? Many definitions – lets use: A simplified statement about of one or more aspects of a loss process that explicitly includes the potential for random effects Two main features – Simplification – Explicit incorporation of random effects Both important and both informative In effect it is a statement about all possible outcomes along with their relative likelihood of occurring, that is a statement of the distribution of outcomes and not just a single “selection” 3 July 26, 2016 Why is This Important? Consider the following very simple loss development triangle: AY 12 24 36 48 2004 1.20 1.10 1.02 ?? 2005 1.25 1.08 ?? ?? 2006 1.15 ?? ?? ?? 2007 ?? ?? ?? ?? Simple chain ladder method: – First pick a “typical” number for each column – Square the triangle with those numbers Not a stochastic model, though a simplified statement of loss process 4 July 26, 2016 Traditional “Deterministic” Approaches Chain ladder – pick factors thought to be representative of column What happens “next year” when new information available? – Often entire exercise is repeated “afresh” – Sometimes we ask “what did we pick last year?” If “actual” varies “too much” from “expected” then we might reevaluate the “expected” How much is “too much” is often dictated by experience, with line of business or particular book being reviewed That indefinable quality – “actuarial judgment” 5 July 26, 2016 Let’s Parse The Traditional Start out with the chain ladder recipe, i.e. a “model” We pick “selections” that are somehow representative of a particular age Experience and “actuarial judgment” often inform us as to what we expect to see (e.g. auto physical damage = stable, umbrella = volatile) Wait a minute – we have a simplified statement about the loss process and an implicit statement about random fluctuation The traditional is almost stochastic already! Why not write down the recipe and expectation of randomness explicitly 6 July 26, 2016 More Info in a Stochastic Context Stochastic approaches gain significant advantages over deterministic ones from many sources – Practitioner is forced to explicitly state his/her assumptions – Not only will a good model give projections, but also estimates of certain data points along the way – we can measure next year’s actual vs. expected Parametric models have some advantages too – They allow for extrapolation beyond the observed data under the assumptions of the model – Good methods for estimating the model parameters also provide estimates of how volatile those parameters themselves are • Maximum likelihood • Bayesian 7 July 26, 2016 We May Never Pass This Way Again Two schools of statistical thought – Frequentist – Bayesian Two distinct approaches in dealing with uncertainty – Frequentist makes the most sense with repeatable experiments – Bayesian attempts to incorporate prior experience in a rational, rigorous fashion Actuarial problems usually do not relate to repeatable experiments, unless you use the dice example… Actuarial judgment is essentially a Bayesian “prior distribution” Bayesian prior is also a way to handle model uncertainty 8 July 26, 2016 All Models are Wrong … The banking sector has “sophisticated” risk models setting capital to be adequate at very high (well above 99) percentiles All is fine … until something like the “subprime crisis” comes along But the models were well founded and based on “considerable” data Think about it – using 10 years of data to estimate a 1-in-1,000 year, or even a 1-in-100 year event really does not make a whole lot of sense The only way to extrapolate from such data is to assume an underlying parametric model and assume that you can extrapolate with it 9 July 26, 2016 Model Uncertainty Mentioned before a good parameter estimation method also gives an estimate of uncertainty in the parameter estimates within that model The subprime issue was not one of parameter estimation but one of model mis-estimation Traditional methods long recognized this problem and solved it by using several forecast techniques At end of day an actuary “selected” his/her “estimate” based on the projections of the various models – stochastically he/she calculated an expected value “forecast” using weights (probabilities) that were determined by “actuarial judgment” Thus there was a Bayesian prior dealing with model uncertainty 10 July 26, 2016 More is Better Stochastic methods can be thought of as extensions of traditional approaches can – Be based on same recipes as traditional methods – Give rigor in “making selections” avoiding the ever-present – – – – temptation to “throw out that point – it is an obvious outlier” Provide more information as to the distribution of outcomes within the scope of the particular model Provide more information as to how well model fits with reality Be evolutionary and evolve as data indicate Be adapted to recognize “actuarial judgment” as well as a multiplicity of potential models All in all stochastic reserving models can give you everything that traditional methods do and much, much more 11 July 26, 2016