Keywords: Nested simulation, loss distribution, value-at-risk, expected shortfall, jackknife estimator, dynamic allocation
Risk measurement for derivative portfolios almost invariably calls for nested simulation. In the outer step one draws realizations of all risk factors up to the horizon, and in the inner step one re-prices each instrument in the portfolio at the horizon conditional on the drawn risk factors. Practitioners may perceive the computational burden of such nested schemes to be unacceptable, and adopt a variety of second-best pricing techniques to avoid the inner simulation. In this paper, we question whether such short cuts are necessary. We show that a relatively small number of trials in the inner step can yield accurate estimates, and analyze how a fixed computational budget may be allocated to the inner and the outer step to minimize the mean square error of the resultant estimator. Finally, we introduce a jackknife procedure for bias reduction and a dynamic allocation scheme for improved efficiency.
JEL Codes: G32, C15
For a wide variety of derivative instruments, computational costs may pose a binding constraint on the choice of pricing model. The more realistic and flexible the model, the less likely that there will exist an analytical pricing formula, and so the more likely that simulation-based pricing algorithms will be required. For plain-vanilla options trading in fast-moving markets, simulation is prohibitively slow. Simple models with analytical solutions are typically employed with ad-hoc adjustments (such as local volatility surfaces) to obtain better fit to the cross-section of market prices. As such models capture underlying processes in crude fashion, they tend to require frequent recalibration and perform poorly in time-series forecasting. For path-dependent options (e.g, lookback options) and complex basket derivatives (e.g., CDO of ABS), simulation is almost unavoidable, though even here computational shortcuts may be adopted at the expense of bias.1
Risk-management applications introduce additional challenges. Time constraints are less pressing than in trading applications, but the computational task may be more formidable. When loss is measured on a mark-to-market basis, estimation via simulation of large loss probabilities or of risk-measures such as Value-at-Risk (VaR) calls for a nested procedure: In the outer step one draws realizations of all risk factors up to the horizon, and in the inner step one re-prices each position in the portfolio at the horizon conditional on the drawn risk factors. It has been widely assumed that simulation-based pricing algorithms would be infeasible in the inner step, because the inner step must be executed once for each trial in the outer step.
In this paper, we question whether inner step simulations must necessarily impose a large computational burden. We show that a relatively small number of trials in the inner step can yield accurate estimates for large loss probabilities and portfolio risk-measures such as Value-at-Risk and Expected Shortfall, particularly when the portfolio contains a large number of positions. Since an expectation is replaced by a noisy sample mean, the estimator is biased, and we are able to characterize this bias asymptotically. We analyze how a fixed and large computational budget may be allocated to the inner and the outer step to minimize the mean square error of the resultant estimator. We show how the jackknifing technique may be applied to reduce the bias in our estimator, and how this alters the optimal budget allocation. In addition, we introduce a dynamic allocation scheme for choosing the number of inner step trials as a function of the generated output. This technique can significantly reduce the computational effort to achieve a given level of accuracy.
The most studied application of nested simulation in the finance literature is the pricing of American options. An influential paper by Longstaff and Schwartz (2001) proposes a least-squares methodology in which a small number of inner step samples are used to estimate a parametric relationship between the state vector at the horizon (in this case, the stock price) and the continuation value of the option. This "LSM" estimator is applicable to a broad range of nested problems, so long as the dimension of the state vector is not too large and the relationship between state vector and continuation value is not too nonlinear. However, some care must be taken in the choice of basis functions, and in general it may be difficult to assess the associated bias (Glasserman, 2004, §8.6). Our methodology, by contrast, is well-suited to portfolios of high-dimensional and highly nonlinear instruments, can be applied to a variety of derivative types without customization, and has bias of known form.
Our optimization results for large loss probabilities and Value-at-Risk are similar to those of Lee (1998).2 Lee's analysis relies on a different and somewhat more intricate set of assumptions than ours, which are in the spirit of the sensitivity analysis of VaR by Gouriéroux et al. (2000) and the subsequent literature on "granularity adjustment" of credit VaR (Gordy, 2004; Martin and Wilde, 2002). The resulting asymptotic formulae, however, are the same.3 Our extension of this methodology to Expected Shortfall is new, as is our analysis of large portfolio asymptotics. Furthermore, so far as we are aware, we are the first to examine the performance of jackknife estimators and dynamic allocation schemes in a nested simulation setting.
In Section 1 we set out a very general modeling framework for a portfolio of financial instruments. We introduce the nested simulation methodology in Section 2. We characterize the bias in and variance of the simulation estimator, and analyze the optimal allocation of computational resources between the two stages that minimizes the mean square error of the resultant estimator. Numerical illustrations of our main results are provided in Section 3. In the last two sections, we propose some refinements to further improve computational performance of nested simulation. Simple jackknife methods for bias reduction are developed in Section 4. Our dynamic allocation scheme is introduced and examined in Section 5.
Let be a vector of state variables that govern all prices. The vector might include interest rates, commodity prices, equity prices, and other underlying prices referenced by derivatives. Let be the filtration generated by . For use in discounting future cash flows, we denote by the value at time of $1 invested at time in a risk free money market account, i.e.,
The portfolio consists of positions. The price of position at time depends on , , and the contractual terms of the instrument.4 Position 0 represents the sub-portfolio of instruments for which there exist analytical pricing functions. Without loss of generality, we treat this as a single composite instrument. Among the contractual terms for an instrument is its maturity. We assume maturity is finite for . As in all risk measurement exercises, the portfolio is assumed to be held static over the model horizon.
Conditional on , the cashflows up to time are nonstochastic functions of time that depend on the contractual terms. Let be the cumulative cashflow for on . Note that increments to can be positive or negative, and can arrive at discrete time intervals or continuously. The market value of each position is the present discounted expected value of its cashflows under the risk-neutral measure :
The present time is normalized to 0 and the model horizon is . "Loss" is defined as the difference between current value and discounted future value at the horizon, adjusting for interim cashflows. Portfolio loss is
We now develop notation related to the simulation process. The simulation is nested: There is an "outer step" in which we draw histories up to the horizon . For each trial in the outer step, there is an "inner step" simulation needed for repricing at the horizon.
Let be the number of trials in the outer step. In each of these trials, we
Observe that the full dependence structure across the portfolio is captured in the period up to the model horizon. Inner step simulations, in contrast, are run independently across positions. This is because the value of position at time is simply a conditional expectation (given and under the risk-neutral measure) of its own subsequent cash flows, and does not depend on future cash flows of other positions. Intuition might suggest that it would be more efficient from a simulation perspective to run inner step simulations simultaneously across all positions in order to reduce the total number of sampled paths of on . However, if we use the same samples of across inner step simulations, pricing errors are no longer independent across the positions, and so do not diversify away as effectively at the portfolio level. Furthermore, when the positions are repriced independently, to reprice position we need only draw joint paths for the elements of that influence that instrument. This may greatly reduce the memory footprint of the simulation, in particular when the number of state variables () is large and when some of the maturities are very long relative to the horizon .
We have assumed that initial prices are already known and can be taken as constants in our algorithm. Of course, this can be relaxed.
In the following three subsections, we discuss estimation of large loss probabilities (§2.1), Value-at-Risk (§2.2), and Expected Shortfall (§2.3). For simplicity, we impose a single value of across all positions (i.e., for ). This restriction is relaxed in Section 2.4. In Section 2.5, we consider the asymptotic behavior of the optimal allocation of computational resources as the portfolio size grows large. Last, in Section 2.6, we elaborate on the trade-offs associated with simultaneous repricing.
We first consider the problem of efficient estimation of via simulation for a given . If for each generated , the mark-to-market values of each position were known, the associated would be known and simulation would involve generating i.i.d. samples and taking the average
Within the inner step simulation for repricing position , each trial gives an unbiased (but very noisy) estimate of . Let denote the zero-mean pricing error associated with the such sample for position , let denote the portfolio pricing error for the inner step sample, and finally let
We now examine the mean square error of . Let denote . The mean square error of the estimator separates into
Let so that has a non-trivial limit as . Then . Our asymptotic analysis relies on Taylor series expansion of the joint density function of and its partial derivatives. Assumption 1 ensures that higher order terms in such expansions can be ignored.
This assumption may be expected to be true in a large portfolio where there are at least a few positions that have a sufficiently smooth payoff. Alternatively, this assumption may be satisfied by perturbing and through adding to both of them mean zero, variance independent Gaussian random variables, also independent of and . For small this has a negligible impact on the tail measures. Then, if
Assumption 1 is sufficient to deliver a useful convergence property. Here and henceforth, let and denote the density and cumulative distribution function for , and let and denote the density and cumulative distribution function for . Now let be some sequence of real numbers that converges to a real number . In Appendix A.1, we prove the following lemma:
We now approximate in orders of . We define the function
For the distributions and large loss levels one might expect to appear in practice, the bias will be upwards (i.e., ). By construction, the distribution of differs from the distribution of by a mean-preserving spread, in the sense of Rothschild and Stiglitz (1970). Unless the two distributions have an infinite number of crossings, there will exist a such that for all .
Applying Proposition 1, the objective function reduces to finding that minimizes
It is easy to see that an optimal for this has the form
For large computational budgets, we see that grows with the square of . Thus, marginal increments to are allocated mainly to the outer step. It is easy to intuitively see the imbalance between and . Note that when and are of the same order , the squared bias term contributes much less to the mean square error compared to the variance term. By increasing at the expense of we reduce variance till it matches up in contribution to the squared bias term.
We now consider the problem of efficient estimation of Value-at-Risk for . For a target insolvency probability , VaR is the value given by
Under Assumption 1, is a continuous random variable so that . As before, our nested simulation generates samples where . We sort these draws as , so that provides an estimate of , where denotes the integer ceiling of the real number . Our interest is in characterizing the mean square error and then minimizing it. As before, we decompose MSE into variance and squared bias:
To approximate bias and variance, we use the following result:Gordy, 2004; Martin and Wilde, 2002). To avoid lengthy technical digressions, our statement of the proposition and its derivation in Appendix A.3 abstract from certain mild but cumbersome regularity conditions; see the appendix for details.
Our budget allocation problem reduces to minimizing the mean square error
Although Value-at-Risk is ubiquitous in industry practice, it is well understood that it has significant theoretical and practical shortcomings. It ignores the distribution of losses beyond the target quantile, so may give incentives to build portfolios that are highly sensitive to extreme tail events. More formally, Value-at-Risk fails to satisfy the sub-addivity property, so a merger of two portfolios can yield VaR greater than the sum of the two stand-alone VaRs. For this reason, Value-at-Risk is not a coherent risk-measure, in the sense of Artzner et al. (1999).
As an alternative to VaR, Acerbi and Tasche (2002) propose using generalized Expected Shortfall ("ES"), defined by
We begin with the more general problem of optimally allocating a computational budget to efficiently estimate for arbitrary . This is easier than the problem of estimating since here is specified while in the latter case is estimated. We return later to analyze the bias associated with the estimate of .
Again, our sample output from the simulation to estimate equals . Let denote . The following proposition evaluates the bias associated with this term.A.4.
Using similar analysis to the proof of Proposition 3, we can establish that
We return now to the problem of the bias of . We can write the difference between the Expected Shortfall of random variables and as
In this subsection we relax the restriction that is equal across . We focus on estimation of large loss probabilities. Similar analysis would allow us to vary across positions in estimating VaR and ES.
We redefine as the total number of inner step simulations. This aggregate is to be divided up among the positions by allocating simulations for position where each and . Suppose that the average effort to generate a single such inner step simulation for is . Then, total inner step simulation effort equals where
The analysis to compute the mean square error proceeds exactly as in Section 2.1. The resultant in this setting is
Recall from Section 2.1 that the mean square error at optimal equals
Intuition suggests that as the portfolio size increases, the optimal number of inner loops needed becomes small, even falling to for a sufficiently large portfolio. We formalize this intuition by considering an asymptotic framework where both the portfolio size and the computational budget increases to infinity. To avoid cumbersome notation and tedious technical arguments, we focus on the case of a portfolio of exchangeable (i.e., statistically homogeneous) positions. The arguments given are somewhat heuristic to give the flavor of analysis involved while avoiding the cumbersome and lengthy notation and assumptions needed to make it completely rigorous.
Consider an infinite sequence of exchangeable position indexed by , and let be the loss on position . Let be the average loss per position on a portfolio consisting of the first positions in the sequence, i.e.,
As before, instead of observing , we generate inner step samples for , so that our simulation provides an unbiased estimator for the probability
By the law of large numbers, , almost surely, so the cdf converges to a non-degenerate limiting distribution , which is the distribution of . Similarly, converges to . Under suitable regularity conditions, , where
We assume that the computational budget for and . The value of captures the size of computational budget available relative to the time taken to generate a single inner loop sample. Note that if then asymptotically even a single sample cannot be generated.
Recall that denotes the number of underlying state variables that control the prices of positions. Suppose that the computational effort to generate one sample of outer step simulation on average equals for some function , and to generate an inner step simulation sample on the average equals for a constant . Average effort per outer step trial then equals and average effort for such trials equals . We analyze the order of magnitude of and that minimize the resultant mean square error of the estimator.
The mean square error of the estimator equals
Up to this point, we have stipulated that the inner step samples for each position in the portfolio are generated independently (conditional on ) across positions. In application to derivative portfolios, there may be factors common to many positions (e.g., the prices of underlying securities) and it may be computationally efficient to generate these factors once for all positions rather than generating them independently for each position. While this reduces the computational effort required to generate a single sample of each position, it induces dependence across positions in the generated samples. If the dependence is such that the sum of the resultant noise from each position has lesser variance than if these samples were generated independently, as might be the case when there are many offsetting positions, then the former is a preferred method. However, typically the noises generated may have positive dependence and that may enhance the variance of the resulting samples thereby increasing the total number of samples required to achieve specified accuracy.
We now make this idea precise in a very simple setting. Consider the case where we want to find the expectation of via simulation. Suppose that average computational effort needed to generate a sample of by generating independent samples of equals for some constant . Let denote the variance of these 's (to keep the discussion simple we assume that all rv have the same variance). Then, the computational effort required to get a specified accuracy is proportional to the variance of the sample times the expected effort required to generate a single sample (see Glynn and Whitt, 1992), i.e., . We refer to this measure as the simulation efficiency.
Now consider the case where we generate by generating dependent samples of . Suppose that the computational effort to generate these samples on average equals for some . Further suppose that the correlation between any two random variables and for is . Then the variance of equals . So the simulation efficiency equals . We therefore prefer to draw dependent samples whenever
We illustrate our results with a parametric example. Distributions for loss and the pricing errors are specified to ensure that the bias and variance of our simulation estimators are in closed-form. While the example is highly stylized, it allows us to compare our asymptotically optimal to the exact optimal solution under a finite computational budget. We have used simulation to perform similar exercises on the somewhat more realistic example of a portfolio of equity options. All our conclusions are robust.
Consider a homogeneous portfolio of positions. Let the state variable, , represent a single-dimensional market risk factor, and assume . Let be the idiosyncratic component to the return on position at the horizon, so that the loss on position is per unit of exposure. We assume that the are i.i.d. . To facilitate comparative statics on , we scale exposure sizes by .
The exact distribution for portfolio loss is . We assume that the position-level inner step pricing errors are i.i.d. per unit of exposure, so that the portfolio pricing error has variance across inner step trials. This implies that the simulated loss variable is distributed . Figure 1 shows how the density of the simulated loss distribution varies with the choice of . For the baseline parameter values , and , we observe that the density of for is a close approximation to the "true" density for . Even for , the error due to inner step simulation appears modest., , .
We consider our estimator of the large loss probability for a fixed loss level . The expected value of is . Applying Proposition 1, bias in expands as where
In Figure 3, we plot the exact root mean square error of as a function of and with the computational budget held fixed. Here we make the assumption that the fixed cost of the outer step is negligible (i.e., ), so that . We observe that the optimal is increasing with the budget, but remains quite modest even for the largest budget depicted ( ).and RMSE for a fixed computational budget. Parameters: , , and .
The relationship between the computational budget and optimal is explored further in Figure 4. We solve for the that minimizes the (exact) mean square error, and plot as a function of the budget . Again, we see that is much smaller than and grows at a slower rate with . For example, when , we find is under 6 and is over 165. Increasing the budget by a factor of 64, we find roughly quadruples while increases by a factor of roughly 16. The figure demonstrates the accuracy of the approximation to given by equation (4). When , the relative error of the approximation is 8.3%. Increasing the budget to shrinks the relative error to under 2.5%.. Parameters: , , and .
The optimal and the accuracy of its approximation may depend on the exceedance threshold , and not necessarily in a monotonic fashion. This is demonstrated in the top panel of Figure 5. The budget here is large, and we see that the approximation to is accurate over the entire range of interest (say, for ). When the budget is smaller (bottom panel), the accuracy of the approximation is most severely degraded in the tail.marked in basis points. Budget is . Parameters: , and .
In Figure 6, we explore the relationship between portfolio size and optimal . For simplicity, we assume here that the budget grows linearly with . In the baseline case of , we find that is roughly 23. When we triple the portfolio size (and budget), we find that falls to under six. If we increase the portfolio size by a factor of 10, we find that is under two. These results suggest that the large portfolio asymptotics of section 2.5 may pertain to portfolios of realistic size.for , and grows linearly with . Parameters: , , and .
As Value-at-Risk is the risk-measure most commonly used in practice, we turn briefly to the estimation of VaR. The results of Section 2.2 show that the bias in vanishes with . This is demonstrated in Figure 7 for three values of . For each line, the y-axis intercept at is the unbiased benchmark, i.e., . The distance on the y-axis between (on the solid line) for any finite and the corresponding unbiased benchmark is the exact bias attributable to errors in pricing at the horizon. The dashed lines show the approximated VaR based on
Jackknife estimators, introduced over 50 years ago by Quenouille (1956), are commonly applied when bias reduction is desired. We show how jackknife methods can be applied in our setting to the estimation of large loss probabilities. Parallel methods would apply to VaR and ES.
We divide the inner loop sample of draws into non-overlapping sections ( and are selected so that is an integer). Section covers the draws . For a single outer step trial, represented by , let denote the sample output from the inner step as proposed in earlier sections. Let be the estimate of that is obtained when section is omitted, and define similarly, e.g., . Observe that the bias in is plus terms, and furthermore that we can construct different sample outputs this way.
We now propose the jackknife inner step simulation output
Define for . From the second representation of in (15), we see that the variance of the estimator is
Our jackknife estimator is the average of samples of , i.e., . The contribution of variance to its mean square error is
In most applications of jackknife methods, bias reduction comes at the price of increased variance. This trade-off applies here as well. To illustrate, we return to the Gaussian example of Section 3. For any parametric example, it is useful to re-write the terms in
equation (16) for the variance of
For the Gaussian example, let
Holding fixed , we plot the bias of the jackknife estimator as a function of the number of sections () in the top panel of Figure 8. The bias is decreasing in , but the sensitivity to is modest in both absolute terms and relative to the bias of the uncorrected estimator. The bias of the uncorrected is 9.04 basis points, whereas the bias of is -0.29 basis points when and -0.15 basis points when . The standard deviation of is plotted as a function of in the bottom panel of the figure. We find the standard deviation increases in roughly linear fashion with . For the uncorrected estimator, we find , whereas for and for ., , , and .
The optimal choice of for minimizing mean square error will depend on . The larger is , the smaller the contribution of variance to MSE, so the larger the optimal . As a practical matter, we advocate setting , which eliminates nearly all the bias at very little cost to variance. Setting has the further advantage of minimal memory overhead, in that one can estimate , , and for each outer step in a single pass through the inner loop. By contrast, to implement the jackknife with sections, we need to save each of the inner loop draws in order to calculate each of the estimates .
Through dynamic allocation of workload in the inner step we can further reduce the computational effort in the inner step while increasing the bias by a negligible controlled amount or even (in many cases) decreasing the bias. Consider the estimation of large loss probabilities for a given . Bias in our estimated is increasing with the likelihood that the indicator obtained from the inner step simulation is not equal to the true for a given realization of the outer step. Say we form a preliminary estimate based on the average of a small number of inner step trials. If this estimate is much smaller or much larger than , it may be a waste of effort to generate many more samples in the inner simulation step. However, if this average is close to , it makes sense to generate many more inner step samples in order to increase the probability that the estimated is equal to the true value . The variance of is dominated by , so dynamic allocation in the inner step will have little effect on the variance. As in Section 4, we develop our dynamic allocation methods for the estimation of large loss probabilities. Similar methods would apply to VaR and ES.
Our proposed dynamic allocation (DA) scheme is very simple. For each trial of the outer step, we generate inner step trials (selecting so that is an integer), and let the resulting preliminary loss estimate be denoted . If for some well-chosen then we terminate the inner step and our sample output is zero.7 Otherwise, we generate a second estimate based on an additional conditionally independent samples. Defining
With dynamic allocation, the bias in the estimated exceedance probability is
If then dynamic allocation will increase absolute bias. Even in this general case, we can derive an upper bound on the increase in absolute bias. As shown in Appendix B, for every one can find an so that the resultant increase in bias is within a specified tolerance. The issue then is to select a that most reduces the computational effort.
For numerical illustration, we return to our Gaussian example of Section 3. The exact expected value of in this example is
More complex dynamic schemes are possible. For example, instead of a single stopping test at , one can have a series of stopping tests after inner step trials. The bandwidths would narrow successively, i.e., . It may also be possible to design adaptive strategies for choosing and that will converge to the optimal values as the outer step simulation progresses.
We have shown that nested simulation of loss distributions poses a much less formidable computational obstacle than it might initially appear. The essential intuition is similar to the intuition for diversification in the long-established literature on portfolio choice. In the context of a large, well-diversified portfolio, risk-averse investors need not avoid high-variance investments so long as most of the variance is idiosyncratic to the position. In our risk-measurement problem, we see that large errors in pricing at the model horizon can be tolerated so long as the errors are zero mean and mainly idiosyncratic. In the aggregate, such errors have modest impact on our estimated loss distribution. More formally, we are able to quantify that impact in terms of bias and variance of the resulting estimator, and allocate workload in the simulation algorithm to minimize mean square error. Simple extensions of the basic nested algorithm can eliminate much of the bias at modest cost.
Our results suggest that current practice is misguided. In order to avoid the "inner step" simulation for repricing at the model horizon, practitioners typically rely on overly stylized pricing models with closed-form solution. Unlike the simulation pricing error in our nested approach, the pricing errors that arise due to model misspecification are difficult to quantify, need not be zero mean, and are likely to be correlated across positions in the portfolio. At the portfolio level, therefore, the error in estimates of Value-at-Risk (or other quantities of interest) cannot readily be bounded and does not vanish asymptotically. Our results imply that practitioners should retain the best pricing models that are available, regardless of their computational tractability. A single trial of a simulation algorithm for the preferred model will often be less costly than a single call to the stylized pricing function,8 so running a nested simulation with a small number of trials in the inner step may be comparable in computational burden to the traditional approach. Despite the high likelihood of grotesque pricing errors at the instrument level, the impact on estimated VaR is small. In the limit of an asymptotically large, fine-grained portfolio, even a single inner step trial per instrument is sufficient to obtain exact pricing at the portfolio level.
Our methods have application to other problems in finance. Nested simulation may arise in pricing options on complex derivatives (e.g., a European call option on a CDO tranche). When parameters in an option pricing model are estimated with uncertainty, nested simulation may also be needed to determine confidence intervals on model prices for thinly-traded complex derivatives. Similar problems arise in the rating of CDOs and other structured debt instruments when model parameters are subject to uncertainty. These applications will be developed in future work.
In this section we prove Lemma 1 and Proposition 1. We then derive the expression for the mean square error of Value-at-Risk discussed in Section 2.2. We then provide a proof of Proposition 3. In our notation, we suppress the dependence on and wherever it is visually convenient.
Similarly we see that as .
Note that the first term above equals
We derive the bias and variance approximations of Proposition 2. Our treatment in places is heuristical to avoid lengthy technical issues.
We first develop an asymptotic expression for the bias . To evaluate , it is useful to consider the order statistics of random variables uniformly distributed over the unit interval. Within this appendix only, let denote the inverse of the cdf of . Observe that has the same distribution as .
Consider the Taylor's series expansion
Let denote the quantile corresponding to the random variable . If we can show that
To see that is , note that due to Assumption 1, and have bounded derivatives, which implies
We now show (28). Using the Taylor's series expansion,
can be written as
We apply a Taylor series expansion to the first term in (32):
For the second term in (32),
We decompose the variance of as . The expectation of is
We now argue that is , ignoring some minor technical issues. From expanding the square and taking expectations, observe that
We now show that these two probabilities are equal in their dominant term and are
. Some notation is needed for this purpose. Let
We proceed in exactly the same fashion for the second term in (33), and find that this term is also .
We can bound the absolute bias under dynamic allocation via the triangle inequality:
From here we can take two different approaches:
We illustrate this computation using our simple Gaussian example with baseline parameter values , , and . This gives us as the standard deviation of the . Let , so that . Say we fix , and . The upper bound in equation (35) for the increase in bias for over the static estimator is under . The probability of stopping with the preliminary sample is