1.1 The use of proxies represents a potential single point of failure of internal models

While the use of proxies makes the estimation of capital in simulation-based capital models tractable, their use introduces a possibility of model failure: the capital approximation may not be close to the capital requirement implied by the financial assumptions, stress calibrations, and model design. Proxies therefore represent a possible single point of failure of internal models.

Careful justification and validation of the use of proxies allow internal model firms to demonstrate their use is appropriate for estimating capital requirements, or for other use cases. However, there is interest in improving proxy modeling and their validation. In the public letter to internal model firms issued by the UK’s Prudential Regulation Authority (PRA),Footnote ¹ they recognize that proxy modeling is an area where thinking and techniques continue to evolve.

It is apparent then that the justification and validation of proxies have significant importance to both the estimation of capital and to widening possible use cases into areas such as new business pricing, liquidity management, and stress testing, among others.

1.2 Using proxies to calculate capital requirements does not make sense without error analysis

Measuring or estimating the possible error in the capital estimate, introduced by proxies, removes proxies from being a potential single point of failure. Practitioners having a measurement of potential error can assess whether the accuracy is appropriate for the use case at hand. For example, the use of proxies may be acceptable if the error analysis indicated they introduce appropriate prudence or accuracy.

Here we consider an approach to measuring proxy errors that may assist firms in remedying proxies from being considered a potential single point of failure within their internal model.

In our setting, a loss scenario is calculated as the change in own funds, defined as the change in value of assets less liabilities. We assume that the liability valuation is given by the market value of a matching asset portfolio, defined by a minimum cost optimization, where eligibility conditions that define a possible portfolio are posed as constraints to the optimization.

Krah et al. (Reference Krah, Nikolić and Korn2018) report that the main computational burden in simulation-based capital models arises from the calculation of projected liability cashflows. We say that the calculations of asset values and projected cashflows are computationally expensive, or “heavy”. Once asset values and cashflows have been calculated for a given risk scenario, the related optimization calculation to establish the liability value is computationally inexpensive, or “light”.

When proxies are used to approximate the optimization data, the approximate loss scenario is formed from the optimal value function of optimization problem perturbed by approximation error. This motivates the consideration of variational analysis of optimization problems as a tool to estimate how error propagates through the loss calculation.

When the optimal value function defining the liabilities is differentiable, explicit expressions for the derivatives can be stated, indicating that a calculus-based approach to error analysis within simulation-based capital models may be applicable in some circumstances.

1.3 Proxies represent a source of systematic error

Consideration of potential errors arising from approximations used to estimate capital requirements is of fundamental importance. Establishing whether an estimate is prudent or appropriate is not possible without consideration of potential error. To further establish whether an approximation is good requires quantification of the potential error.

Once potential sources of error have been recognized, and their potential size quantified, error analysis describes the process of quantifying their potential impact as they spread into subsequent calculations. Hughes & Hase (Reference Hughes and Hase2010) provide a thorough and elementary introduction to error analysis in calculations. They introduce a taxonomy of three sources of errors that have natural parallels to the setting of proxy modeling within simulation-based capital models:

• Random errors – e.g., values being approximated by averages from Monte Carlo sampling.

• Systematic errors – e.g., values approximated by a proxy function.

• Mistakes – e.g., bad data points, coding errors.

They observe that error analysis is often an exercise that requires different approaches for different situations. Kirkup & Frenkel (Reference Kirkup and Frenkel2006) note that the terms “error” and “uncertainty” are often used interchangeably. Here we use the term error to refer to the difference between the calculated value and the true value. This usage follows the guidelines of International Organization for Standardization (ISO) (1993).

In our application, approximation errors arising from the use of proxies are systematic and repeatable and are not necessarily random when considered as functions of the risk factor space. We assume that data are accurate, and no errors are due to mistakes. Where an underlying model itself uses sampling or Monte Carlo simulation, the model’s output can be considered random. In our study, we assume in such cases that the model error is bounded by known values, and at the discretion of the practitioner, can be reduced to an arbitrary small amount – for example, by increasing the number of simulations used, or by refinement of approximation schemes.

We use the calculus-based approach to error analysis, since we wish to assert candidates for loss error bounds. Alternatively, a distributional approach could be taken in the analysis. This approach does not form part of this study, since we are interested in establishing analytic bounds on approximated loss scenarios. Crispin & Kinsley (Reference Crispin and Kinsley2022) show that once analytic bounds on approximated loss scenarios are known, the potential error in the corresponding capital estimates may be calculated. The calculation of error bounds on capital estimates allows the practitioner to pursue error reduction and, possibly, error elimination.

1.4 Error analysis and attribution creates a mechanism for error reduction and removal

The loss error bound can be attributed to individual proxy functions, leading to a mechanism for improvements to the approximations, and when combined with the proxy error elimination technique of Crispin & Kinsley (Reference Crispin and Kinsley2022), gives the possibility of completely removing proxy errors from capital estimates.

Christiansen (Reference Christiansen2008) describes examples of uses of first order sensitivity analysis within life insurance applications, referred to as the sensitivity concept. In our application, we consider the full derivative and assert the tangent plane as approximating the function at nearby points.

Here, a mathematical framework for the analysis of errors in simulation-based capital models is developed. The techniques are demonstrated with a stylized example and associated numerical study. The purpose is to demonstrate that formal error analysis may be feasible for some internal model designs. Formal methods of error analysis can have a role in the validation of capital estimates and in the communication of capital uncertainty.

1.5 Wider applications beyond error analysis

Beyond error analysis within simulation-based capital modeling, the presented methods can be applied to general sensitivity analysis and the calculation of risk within the setting of classical asset-liability-management. Many studies have presented the notion of matching as a technique to reduce exposure to interest rate changes through the formulation of a portfolio optimization problem where matching conditions are represented as optimization constraints. Examples of such studies include: Tilley (Reference Tilley1980), Wise (Reference Wise1984a, Reference Wise1984b), Kocherlakota et al. (Reference Kocherlakota, Rosenbloom and Shiu1988), Daykin & Hey (Reference Daykin and Hey1990), and Conwill (Reference Conwill1991).

Consider, for example, the optimization problem formulated by Kocherlakota et al. (Reference Kocherlakota, Rosenbloom and Shiu1988). They consider a minimal cost portfolio of bonds chosen to cover liability cashflows. They allow for cash deposits and for borrowing, and define matching by the requirement that the terminal balance of payments is non-negative. Positive cash balances are assumed to accrue at a cash rate, and borrowing, represented by negative cash balances, is assumed to accrue at a higher borrowing rate. The trick used in their formulation to represent the matching portfolio as a solution to a linear optimization problem is to note that matching constraints can be written as linear expressions of separate positive and negative cash balances, with these placed as decision variables within the optimization.

The novelty of our presentation within this general application lies within the consideration of Lipschitz continuity (Lemma 4) to motivate the likely existence of the full derivative with respect to all data components of the linear optimization problem (Lemma 5).

2.1 Direct loss approximation

The setting of this paper is the analysis of error propagation arising from the use of proxies in simulation-based capital models. One approach to using proxies is to pose a proxy for the loss calculation directly as a function of the risk factor space. Let $\boldsymbol{{r}}$ be a risk factor scenario and let $x(\boldsymbol{{r}})$ denote the associated exact loss, and let $x^*(\boldsymbol{{r}})$ denote the associated proxy value. In this study, positive values of $x$ represent gains, and negative values represent losses. The difference in computational burden between the heavy and light models can be shown schematically as

(2a) \begin{align} \boldsymbol{{r}}\xrightarrow{\text{heavy}} x(\boldsymbol{{r}}), \end{align}

(2b) \begin{align} \boldsymbol{{r}}\xrightarrow{\text{light}} x^*(\boldsymbol{{r}}). \end{align}

However, directly approximating loss as a function of the risk factor space may be problematic. The potentially high dimensionality of the risk factor space means it can be computationally difficult, or even impossible, to sample the risk factor sufficiently to ensure goodness of fit, and to perform validation. Problems associated with performing data analysis in high dimensions is often referred to as the curse of dimensionality. Liu & Özsu (Reference Liu and Özsu2009) describe the curse of dimensionality as the phenomena whereby the number of samples needed to approximate an arbitrary function with a given level of accuracy grows exponentially with respect to the number of input variables (i.e., dimensionality) of the function. They note the expression was introduced by Bellman (Reference Bellman1961). Debie & Shafi (Reference Debie and Shafi2019) describe the curse of dimensionality as a phenomenon that arises when analyzing data in high-dimensional spaces, challenging algorithm performance and accuracy.

The curse of dimensionality is associated with the direct approximation approach when the dimensionality of the risk factor space is high. Therefore, in such a circumstance, it can be advantageous to investigate alternative approaches involving functions that are, somehow, low-dimensional. There are many ways that a function of high dimensions could be considered low dimensional. For simplicity, we consider a function as being low dimensional if it is dependent only on a small number of risk factor components.

Hejazi & Jackson (Reference Hejazi and Jackson2017) report that the primary computational difficulty in computing capital requirements often lies within the computation of liability cashflow scenarios. In alignment with this observation, here we consider the situation where asset values and projected cashflows, considered as functions of the risk factor space, are low dimensional and can be easily approximated with proxies. We further assume that once these have been approximated, the onward computation of the approximate loss scenario is light. In this setting, we may consider an indirect loss approximation.

2.2 Indirect loss approximation

Suppose, as before, that $\boldsymbol{{r}}$ represents the risk-factor scenario. Denote by $\boldsymbol{\pi }$ a computationally expensive (heavy) function of the risk factor that determines all required information to calculate the loss scenario. This includes asset values (possibly aggregated), asset and liability cashflows and all information required to specify the matching portfolio within the loss calculation. The loss function, considered as function of the data $\boldsymbol{\pi }$ , is denoted $\mathcal{X}(\boldsymbol{\pi })$ . The composition of computational burdens, in terms of heavy and light models, can be shown schematically as

(3a) \begin{align} \boldsymbol{{r}}\xrightarrow{\text{heavy}}\boldsymbol{\pi }(\boldsymbol{{r}})\xrightarrow{\text{light}} \mathcal{X}\left (\boldsymbol{\pi }(\boldsymbol{{r}})\right ). \end{align}

The expression $x(\boldsymbol{{r}})$ in (2a) evaluates identically to the expression $\mathcal{X}\left (\boldsymbol{\pi }(\boldsymbol{{r}})\right )$ in (3a), that is, $x(\boldsymbol{{r}}) = \mathcal{X}\left (\boldsymbol{\pi }(\boldsymbol{{r}})\right )$ . The notational difference between (2a) and (3a) is used to emphasize that the loss scenario can be calculated directly as a function of the risk factor, or indirectly via intermediate calculations such as cashflows and asset values.

The computational burden from risk factor to loss scenario is heavy due to the processing of data under the risk factor scenario. Once these have been calculated, the computational burden of calculating the corresponding loss scenario is assumed light. In order to make the computation tractable under Monte Carlo simulation, a proxy function for the data $\boldsymbol{\pi }$ , denoted $\boldsymbol{\pi }^*$ is introduced, with corresponding schematics of the computational burden given by:

(3b) \begin{align} \boldsymbol{{r}}\xrightarrow{\text{light}}\boldsymbol{\pi }^*(\boldsymbol{{r}}) \xrightarrow{\text{light}} \mathcal{X}\left (\boldsymbol{\pi }^*(\boldsymbol{{r}})\right ). \end{align}

If the data proxy $\boldsymbol{\pi }^*$ approximates the exact data $\boldsymbol{\pi }$ sufficiently well, a practitioner may hope that the approximate loss $\mathcal{X}\left (\boldsymbol{\pi }^*(\boldsymbol{{r}})\right )$ closely approximates the exact loss $\mathcal{X}\left (\boldsymbol{\pi }(\boldsymbol{{r}})\right )$ sufficiently well for practical applications. However, this approach does not address the important question of how closely the approximation matches the (possibly unknown) exact value.

A key feature of the indirect approach is the requirement to develop proxies for each data element of $\boldsymbol{\pi }$ . The number of proxies that may need to be developed could be significant due to the potentially large number of data elements. However, for a given data element, the number of risk factors that it is sensitive to is likely to be low, and therefore the development of proxies is more easily achieved due to the issues of dimensionality not being present.

Here, we investigate estimation of the approximation error $\left | \mathcal{X}\left (\boldsymbol{\pi }^*(\boldsymbol{{r}})\right ) - \mathcal{X}\left (\boldsymbol{\pi }(\boldsymbol{{r}})\right )\right |$ given knowledge of the proxy errors $\left | \pi ^*_i(\boldsymbol{{r}}) - \pi _i(\boldsymbol{{r}})\right |$ at each component of the data $\boldsymbol{\pi }^*(\boldsymbol{{r}}) - \boldsymbol{\pi }(\boldsymbol{{r}})$ . First, we clarify the definition of proxy functions and their assumed computational properties in the context of their role in simulation-based capital models.

The use of the full optimization for liability valuation does propagate fitting errors from the proxies to the loss scenario, but does not itself introduce any new sources of uncertainty. The use of a potentially high-dimensional loss proxy is avoided. Through a variational analysis of the optimization problem used to define total liability, we show in a realistic abstract setting that the potential size of errors propagating from data proxies through to the loss proxy may be estimated analytically.

Looking forward, we find two categories of expressions required to form the total approximate error bounds:

• the derivative of the (indirect) loss function with respect to each data component $\frac{\partial \mathcal{X}}{\partial \pi _i^*}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)$ ,

• the component-wise error bounds $\varepsilon (\boldsymbol{{r}},\pi _i^*)$ of the data proxy components $\pi _i^*\left(\boldsymbol{{r}}\right)$ .

In our study and analysis, an important assumption is made that component-wise data proxies can be considered as low-dimensional functions of the risk factor space and can be formed with known error bounds.

Androschuck et al. (Reference Androschuck, Gibbs, Katrakis, Lau, Oram, Raddall, Semchyshyn, Stevenson and Waters2017) connect the activity of proxy model validation with proxy model error bounds. For example, they observe the common approach of measuring the maximum absolute error observed on points used in the construction of the proxy and at other points throughout the function’s domain. They also highlight the usefulness of visual inspection of a proxy’s fit, not only for evidence used to support validation but also as part of the communication of the validation work. They note that one of the drawbacks of visualization techniques is that they naturally tend to be useful for functions of one or two dimensions. When a function is high dimensional, effective exploration of the function’s domain for out-of-sample testing also becomes increasingly impractical. Therefore, common approaches to proxy validation may only be valid in low dimensions, motivating our assumption within this analysis.

While the activity of estimating error bounds of data proxies is necessarily model dependent, it is important to note that the effectiveness of error propagation analysis may be affected by the quality of the error bounds of the data proxies. The relationship between the quality of the (approximate) loss error bounds and the (approximate) error bounds of the data proxies is explored in a sensitivity analysis in Table 1, demonstrating possible failure of approximate error bounds on losses may occur even when error bounds for data proxies are chosen prudently.

Table 1. The sensitivity of the success of approximate error analysis to the estimation of error bounds of data proxies within the stylized example of a simulation-based capital model.*

* The table shows the sensitivity of the success of the approximated loss error bound with respect to a scaling factor $\gamma$ applied to the estimate of the data proxy error bounds. Loss and approximate bound data are from a Monte Carlo of 1,000,000 simulations from the model described in Section 5.1 with example data from Example 5.1. The proportion of successful unordered bounds is measured by the proportion of times within the Monte Carlo simulation the posed bounds in (32a) with $l_i = x^*(\boldsymbol{{r}}_{\boldsymbol{{i}}})-\gamma \varepsilon (\boldsymbol{{r}}_{\boldsymbol{{i}}};\,x^*)$ and $u_i = x^*(\boldsymbol{{r}}_{\boldsymbol{{i}}})+\gamma \varepsilon (\boldsymbol{{r}}_{\boldsymbol{{i}}};\,x^*)$ hold with mathematical (nonstrict) inequality. Similarly, the analogous sensitivity to $\gamma$ for ordered approximate bounds to hold with mathematical (nonstrict) inequality is shown. The success of the approximate bounds is more successful in the setting of the ordered bounds, suggesting that sorting provides additional prudence. It is important to note that, in this example, the unordered bounds fail (in the mathematical sense of inequality) in about 4.6% of cases in the base case of $\gamma =1$ and is only fully successful when prudence to the data proxy error bounds is applied, represented by $\gamma =1.45$ . Numbers are shown rounded to 1d.p. with the underlying data for the entries marked 100.0% being exactly 100% showing for these entries all proxy losses lie within the bounds.

2.3 Posing error bounds through linear approximations

It is commonly known that a differentiable function can be approximated by its tangent plane in a neighborhood of the point that the derivative is taken. Here, we use the tangent plane as an approximating function and consider the potential approximation error that may arise as errors propagate through onward calculations. Once derivatives have been established, and the tangent plane proposed as an approximating function, the triangle inequality can be used to propose approximate bounds on the potential error. We show that the approximate error bound may have potential application, although we also demonstrate that the bound is still approximate and can fail. Therefore, careful validation and prudence is required before this approach is used for capital requirements modeling.

Here, we use the mean value theorem to create an analytical bound between the values of a function between two points, in order to compare the bound to an approximate bound developed though using the tangent as an approximation. Fig. 1 illustrates that approximate bounds developed this way can be effective, but importantly, may also fail.

Figure 1. An illustration of analytical error bounds (Panel A) and approximate error bounds (Panel B) based on Example 2.1. The function $\mathcal{X}\,:\,\mathbb{R}\mapsto \mathbb{R}$ is defined to be $\mathcal{X}(s)=\exp (s)$ , $s^*=0.8$ and $\varepsilon =0.5$ with $s$ satisfying $|s-s^*|\leq \varepsilon$ . The function $s\mapsto \mathcal{X}(s)$ (blue line) and the point $(s^*,\mathcal{X}(s^*))$ (blue dot) are shown identically in Panels A and B. The left and right boundaries of the (green) rectangles of Panels A and B are identical and given by $s^*\pm \varepsilon$ . Panel A: The (green) rectangle depicts the feasible region for $(s,\mathcal{X}(s))$ defined by analytical error bounds in (12a) depicted as horizontal lines (green). The use of green indicates that the bounds are effective. Panel B: The (green and hatched-red) rectangle depicts the approximated feasible region for $(s,\mathcal{X}(s))$ defined by approximate error bounds in (12b). Regions where the approximated upper and lower approximate bounds fail to hold are shown in hatched-red. Note the approximate lower bound is effective while the approximate upper bound fails for values of $s$ near to its maximum value $s^*+\varepsilon$ .

In order to compare a concrete example of exact and approximate error bounds, recall the mean value theorem from elementary calculus. Here, we state the theorem in its high-dimensional form applicable to our application, where there are many data components forming input to the loss calculation. In the following, the integer $d$ denotes the number of components of the data $\boldsymbol{\pi }$ .

Lemma 1. (Mean value theorem for functions of several variables) Suppose that $\mathcal{X}\,:\,\mathbb{R}^d\mapsto \mathbb{R}$ is differentiable at each point on an open set $S\subset \mathbb{R}^d$ . If $\boldsymbol{{a}}, \boldsymbol{{b}}\in S$ and $L(\boldsymbol{{a}},\boldsymbol{{b}})\,:\!=\,\{\boldsymbol{{a}}t+(1-t)\boldsymbol{{b}}\,:\,t\in (0,1)\}\subset S$ , then there exists $\boldsymbol{{z}}\in L(\boldsymbol{{a}},\boldsymbol{{b}})$ such that

(4) \begin{align} \mathcal{X}(\boldsymbol{{b}}) = \mathcal{X}(\boldsymbol{{a}}) + \sum _{i=1}^d \frac{\partial \mathcal{X} (\boldsymbol{{z}})}{\partial x_i}(\boldsymbol{{b}}_i - \boldsymbol{{a}}_i). \end{align}

Under the additional assumption that the partial derivatives are continuous on the closure of the interval $L(\boldsymbol{{a}},\boldsymbol{{b}})$ , written $L[\boldsymbol{{a}},\boldsymbol{{b}}]$ , there holds:

(5) \begin{align} \left | \mathcal{X}(\boldsymbol{{b}}) - \mathcal{X}(\boldsymbol{{a}}) \right | \leq \max _{\boldsymbol{{z}}\in L[\boldsymbol{{a}},\boldsymbol{{b}}]} \sum _{i=1}^d \left | \frac{\partial \mathcal{X} (\boldsymbol{{z}})}{\partial x_i}\right ||\boldsymbol{{b}}-\boldsymbol{{a}}|. \end{align}

The analytic bound (5) is considered for illustrative and comparative purposes only. It has limited use in practical applications since the evaluation point $\boldsymbol{{z}}$ is unknown. To understand the potential size of the analytical error bound, all derivatives must be known along the line $L[\boldsymbol{{a}}, \boldsymbol{{b}}]$ . Apart from offering no computational advantage to simply using the heavy model directly, even estimation of the bound may be infeasible due to the computational limitations of using the heavy model in a Monte Carlo setting. These factors motivate the consideration of approximate methods of error analysis.

Consider again the indirect loss calculation (3a) with the loss function, $\mathcal{X}\,:\,\mathbb{R}^d\mapsto \mathbb{R}$ , defined on a high-dimensional domain representing the large number of data components used in the loss calculation. The data components are dependent of the risk factor scenario so can be considered functions of the risk factor space: $\pi _i\,:\,E\mapsto \mathbb{R}$ with $\boldsymbol{\pi }(\boldsymbol{{r}})=(\pi _1(\boldsymbol{{r}}),\pi _2(\boldsymbol{{r}}),\ldots,\pi _d(\boldsymbol{{r}}))$ . The loss scenario can then be written $\mathcal{X}(\pi _1(\boldsymbol{{r}}),\pi _2(\boldsymbol{{r}}),\ldots,\pi _d(\boldsymbol{{r}}))$ or simply $\mathcal{X}(\boldsymbol{\pi }(\boldsymbol{{r}}))$ .

Suppose now that proxies $\pi _i^*$ have been developed for $\pi _i$ with corresponding errors bounds as known functions (Definitions 2 and 3). That is, suppose that there are functions $\varepsilon \left(\cdot ;\, \pi _i^*\right)$ of the risk factor space satisfying:

(6) \begin{align} |\pi _i^*(\boldsymbol{{r}})-\pi _i(\boldsymbol{{r}})| \leq \varepsilon \left(\boldsymbol{{r}};\,\pi _i^*\right). \end{align}

The reasonableness of this assumption relies on two aspects. First, the data-components depend, when considered individually, only on a small number of risk factors. That is, the curse of dimensionality is not encountered. Second, the computational resources are available to design, calibrate, and validate the proxies and their error bounds.

We are interested in understanding how well $\mathcal{X}(\pi _1(\boldsymbol{{r}}),\pi _2(\boldsymbol{{r}}),\ldots,\pi _d(\boldsymbol{{r}}))$ is approximated by $\mathcal{X}\left(\pi _1^*(\boldsymbol{{r}}),\pi _2^*(\boldsymbol{{r}}),\ldots,\pi _d^*(\boldsymbol{{r}})\right)$ . Since a differentiable function is approximated by its tangent plane at nearby points, we pose the tangent plane as an approximating function:

(7) \begin{align} \mathcal{X}\left(\boldsymbol{\pi }(\boldsymbol{{r}})\right)- \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right) \approx \sum _{i=1}^d \frac{\partial \mathcal{X}}{\partial x_i}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)\left(\pi _i(\boldsymbol{{r}})-\pi _i^*(\boldsymbol{{r}})\right). \end{align}

Here, the symbol $\alpha \approx \beta$ is used to denote that an expression $\beta$ has been posed as an approximation to $\alpha$ , and importantly, without any claim to its accuracy. Through a basic application of the triangle inequality, and applying expression (6), there holds:

(8) \begin{eqnarray} \left | \sum _{i=1}^d \frac{\partial \mathcal{X}}{\partial x_i}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)\left(\pi _i(\boldsymbol{{r}})-\pi _i^*(\boldsymbol{{r}})\right) \right | \leq \sum _{i=1}^d \left |\frac{\partial \mathcal{X}}{\partial x_i}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)\right | \varepsilon \left(\boldsymbol{{r}};\,\pi _i^*\right). \end{eqnarray}

Following the idea that a differentiable function is approximated by its tangent plane, the value $\sum _{i=1}^n \left |\frac{\partial \mathcal{X}}{\partial x_i}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)\right | \varepsilon \left(\boldsymbol{{r}};\,\pi _i^*\right)$ is posed as a potential bound for $\left |\mathcal{X}(\boldsymbol{\pi }(\boldsymbol{{r}}))- \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)\right |$ .

Notation 1. (Posing approximate bounds) For $A,B\in \mathbb{R}$ , the notation $A\approx B$ means that $B$ is posed as an approximation to $A$ , and the notation $A\lessapprox B$ means that $B$ is posed as a potential upper bound for $A$ . In either usage, any of $A\leq B$ or $A\gt B$ may hold in practice. For example, we write

(9) \begin{eqnarray} \left | \mathcal{X}(\boldsymbol{\pi }(\boldsymbol{{r}})) - \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right) \right | \lessapprox \sum _{i=1}^d \left |\frac{\partial \mathcal{X}}{\partial x_i}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)\right | \varepsilon \left(\boldsymbol{{r}};\,\pi _i^*\right) \end{eqnarray}

where, in steps (6), (7) and (8), the expression $\sum _{i=1}^d \left |\frac{\partial \mathcal{X}}{\partial x_i}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)\right | \varepsilon \left(\boldsymbol{{r}};\,\pi _i^*\right)$ is posed as an approximate upper bound for $\left | \mathcal{X}(\boldsymbol{\pi }(\boldsymbol{{r}})) - \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right) \right |$ .

It is important to note that the use of the symbol $\lessapprox$ highlights that the bound (9) has not been mathematically proven and should therefore be subject to further validation before its use in capital requirements modeling.

The approximate bound (9) is posed as having potential practical applications. The approximate bound may be used as a starting point in developing prudent bounds that can be defended through validation activities. It is important to note that approximate error bounds may fail in practical conditions, for example, when the function exhibits significant curvature or when the valuation point differs significantly from the point used to establish the tangent plane, see Fig. 1. Therefore, it is important to validate approximate error bounds, or ones posed from these as prudent, before their use within capital requirement modeling. Next we consider an elementary example that illustrates how approximate error bounds may fail.

2.4 An elementary example illustrating success and failure of approximate error bounds

Here, we compare the analytic bound (5) with the approximate bound (9) in a simple setting in order to illustrate how approximate error bounds developed in this way may fail.

For this purpose, consider a one-dimensional loss function, $\mathcal{X}\,:\,\mathbb{R}\mapsto \mathbb{R}$ , and its value at two points $s,s^*\in \mathbb{R}$ . The value $s^*$ represents a known value derived through a proxy, and $s$ represents an exact value that is unknown unless computational resources are used to establish it. Suppose, mimicking (6), that there exists a proxy bound $\varepsilon \gt 0$ , so that $|s-s^*|\leq \varepsilon$ . Assuming sufficient continuity and differentiability, by (5), there holds:

(10) \begin{align} \text{Analytic:}\qquad \left | \mathcal{X}(s) - \mathcal{X}(s^*) \right | \leq \max _{s\in [s^*-\varepsilon,s^*+\varepsilon ]}|\mathcal{X}'(s)|\varepsilon. \end{align}

The approximate bound, proposed in (9), becomes:

(11) \begin{align} \text{Approximate:}\qquad |\mathcal{X}(s)-\mathcal{X}(s^*)| \lessapprox \left | \mathcal{X}'(s^*) \right |\varepsilon. \end{align}

The following example illustrates a comparison between the analytic expression (10) and the quantity in (11) representing an approximate error bound.

Example 2.1. Let $\mathcal{X}\,:\,\mathbb{R}\mapsto \mathbb{R}$ be given by $\mathcal{X}(s) =\exp (s)$ for all $s\in \mathbb{R}$ . Let $s^*=0.8$ , and $\varepsilon =0.5$ , so that $|s-s^*|\leq \varepsilon$ . In this example, the analytic bound (10) and approximate bound (11) are given for, $s\in [s^*-\varepsilon,s^*+\varepsilon ]$ , by

(12a) \begin{align} \text{Analytic:}\qquad \left | \mathcal{X}(s) - \mathcal{X}(s^*) \right | \leq \exp (s^*+\varepsilon )\varepsilon, \end{align}

(12b) \begin{align}\text{Approximate:}\qquad |\mathcal{X}(s)-\mathcal{X}(s^*)| \lessapprox \left | \exp (s^*) \right |\varepsilon. \end{align}

The approximate error bound fails for large $s\in [s^*-\varepsilon, s^*+\varepsilon ]$ , but is effective for other values, see Fig. 1.

The success and failure of approximate bounds in Example 2.1 are illustrated in Fig. 1. Regions of feasibility are differentiated by colors. Green indicates success, as is guaranteed in the analytical case, and hatched-red indicated failure. The use of the triangle inequality (8) has meant, in this example, that the approximate bound is successful over the majority of the range of $s$ , though fails when $s$ is near its upper boundary. In higher dimensions, where errors arising from different components may be offsetting, the triangle inequality can be expected to yield a more prudent bound. However, this example illustrates that failure can still be observed. In these instances, the approximate upper bound was close but not strict. While the bounds are posed as approximate, there may be circumstances where suitable additional prudence ensures that the bounds are effective for their use case. In these circumstances, bounds posed as prudent would still require suitable validation before their use in capital estimation.

In order to use a calculus-based approach to error analysis, we must consider the differentiability of the underlying model.

2.5 Differentiability phenomena

Classical calculus-based error analysis requires differentiability. In the setting of the indirect loss approximation, where data are proxied, the derivative of a loss scenarios with respect to its model data is a key input into a calculus-based error analysis. The question of whether a model is differentiable, considered as a function of its input data, naturally arises. In simulation-based capital models, calculation steps are complex, and analytical expressions for loss scenarios are not necessarily available. Also, it may not be straightforward to prove differentiability at any given risk factor scenario. Related to this, it may be easy to identify possible situations where differentiability may not be expected. For example, in cases where contractual cashflows depend discontinuously on risk factor values. However, when the risk factor is chosen randomly, as in our setting of simulation-based capital models, a general differentiability phenomena may arise. A striking and emblematic example of a differentiability phenomena is Lebesgue’s Theorem on monotone functions.Footnote ³

Consider Lebesgue’s theorem in a Monte Carlo setting. Suppose the function $f\,:\,\mathbb{R}\mapsto \mathbb{R}$ is monotone, and suppose $R$ is a random variable with distribution $\mu _R$ that is absolutely continuous with respect to Lebesgue measure. Then, by Lebesgue’s Theorem, there is zero probability that $R$ takes a value where $f$ is nondifferentiable. Similarly, suppose $\{r_i\}_{i=1}^N$ is an independent sample from $R$ . Then Lebesgue’s theorem implies $f$ is differentiable at each point $r_i$ with probability one.

Rademacher (Reference Rademacher1919) established differentiability almost everywhere for a large class of functions, applicable to our study, having a locally Lipschitz property.

Definition (Locally Lipschitz) A function $f$ from $\mathbb{R}^n$ (or a subset of $\mathbb{R}^n$ ) into $\mathbb{R}$ is called locally Lipschitz if for any bounded set $B$ from the interior of the domain of $f$ there exists $K\gt 0$ such that

(13) \begin{align} |f(x)-f(y)|\leq K\|x-y\|, \text{ for all }x,y\in B. \end{align}

In the one-dimensional setting, it can be readily observed that Lebesgue’s theorem implies locally Lipschitz functions are differentiable almost everywhere. Consider, for example, the function $f\,:\,\mathbb{R}\mapsto \mathbb{R}$ assumed locally Lipschitz. Then for each bounded open interval $B\subset \mathbb{R}$ , there exists $K\geq 0$ such that $|f(x)-f(y)|\leq K|x-y|$ for all $x,y\in B$ . Then $x\mapsto f(x) + Kx$ is nondecreasing on $B$ , and so, by an application of Lebesgue’s Theorem, it is differentiable almost everywhere on $B$ . Therefore $x\mapsto f(x)$ is differentiable almost everywhere on $\mathbb{R}$ .

In what follows, we establish that our loss function $\mathcal{X}$ may exhibit the locally Lipschitz property when considered as a function of the data forming the loss calculation. Whenever the locally Lipschitz property holds, Rademacher’s theorem implies the loss model is differentiable almost everywhere with respect to the data. Differentiability then allows us to consider the calculus-based approach to proposing proxy error bounds.

In summary, computationally heavy models may be approximated by tangent planes in Monte Carlo settings, when it can be shown that functions exhibit the differentiability phenomenon. Looking ahead, this approach can be used to form approximate loss calculations with computationally tractable approximate error bounds.

4.1 Computational assumptions

Androschuck et al. (Reference Androschuck, Gibbs, Katrakis, Lau, Oram, Raddall, Semchyshyn, Stevenson and Waters2017) and Crispin & Kinsley (Reference Crispin and Kinsley2022)Footnote ¹¹ describe that practical computational limitations motivate the use of approximations in capital requirements modeling. To reflect the role computational limitations play in motivating our study, they are introduced within our framework through the following assumption.

A consequence of Assumption 2 and Lemma 5 is that the derivatives of the loss function with respect to data, shown in (25), are feasible to compute for a large number of risk factor scenarios, whenever the scenario data are available.

4.2 Definition of the proxy loss function

Here, we define the data proxy $\boldsymbol{\pi }^*$ and proxy loss function $x^*$ of the indirect approach (3b). When considering proxies for data elements of the loss calculation, we recall that we are motivated to consider this approach, since when taken individually the components of the data are only dependent on a small number of components of the risk factors. The data components can therefore be considered as low-dimensional functions, and therefore development of proxies for these are not exposed to the curse of dimensionality. This motivates our assumption that proxies for data components can be developed with appropriately validated proxy error bounds.

Assumption 3. (Data proxy) Consider the loss model (Definition 4) and suppose there exists a proxy $\boldsymbol{\pi }^*$ (Definition 2) with error bounds (Definition 3) for the scenario-data function $\boldsymbol{\pi }$ defined in (15a). That is, suppose there exists computationally light function (Definition  1 ) $\boldsymbol{\pi }^*\,:\,E \mapsto \mathbb{R}^{n+mn+m}$ satisfying:

(26) \begin{align} \text{Data proxy:}\qquad \boldsymbol{\pi }^*(\boldsymbol{{r}}) = [\boldsymbol{{A}}^*(\boldsymbol{{r}}),\boldsymbol{{b}}^*(\boldsymbol{{r}}),\boldsymbol{{c}}^*(\boldsymbol{{r}})] \end{align}

for all $\boldsymbol{{r}}\in E$ , where components of the data proxies are written: $A_{i,j}^*(\boldsymbol{{r}})$ , $b_j^*(\boldsymbol{{r}})$ and $c_i^*(\boldsymbol{{r}})$ . Suppose further that the components of the data proxy have computationally light error bounds. That is, there exist computationally light functions $\varepsilon \left(\cdot ;\, A^*_{i,j}\right),\varepsilon \left(\cdot ;\, b^*_j\right),\varepsilon \left(\cdot ;\, c^*_i\right)\,:\,E\mapsto \mathbb{R}$ , satisfying:

(27a) \begin{align} \left|A_{i,j}(\boldsymbol{{r}})- A^*_{i,j}(\boldsymbol{{r}})\right| \leq \varepsilon \left(\boldsymbol{{r}};\,A^*_{i,j}\right), \end{align}

(27b) \begin{align} \left|b_j(\boldsymbol{{r}})- b^*_{j}(\boldsymbol{{r}})\right| \leq \varepsilon \left(\boldsymbol{{r}};\, b^*_{j}\right), \end{align}

(27c) \begin{align} \left|c_i(\boldsymbol{{r}})- c^*_{i}(\boldsymbol{{r}})\right| \leq \varepsilon \left(\boldsymbol{{r}};\, c^*_{i}\right), \end{align}

for all $\boldsymbol{{r}}\in E$ .

The practicality of Assumption 3 will depend on whether proxies and their bounds can be asserted and validated for all the required scenario-data inputs. In our exposition, it is assumed that the computational demand to fit and validate the data proxies are within practical computational limitations. Assumption 3 may not be reasonably expected to hold in situations where data components are essentially high dimensional and depend on many dimensions of the risk-factor space.

Definition 5. (Proxy loss model) Consider the loss model (Definition 4) and recall the (exact) loss function $x\,:\, E\mapsto \mathbb{R}$ defined in (14). The function $x^*\,:\, E\mapsto \mathbb{R}$ defined by

(28) \begin{align} x^*(\boldsymbol{{r}}) \,:\!=\, \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right), \end{align}

is posed a proxy function for $x$ , where $\mathcal{X}$ is defined in (15b) and $\boldsymbol{\pi }^*$ is defined in (26).

The fact that the proxy $x^*$ is computationally light follows from Assumptions 2 and 3. The utility of $x^*$ as a proxy function will depend on whether error bounds can be established. In forming error bounds, we note that the differentiability phenomena may also occur when using proxied data. In such situations, analytical derivative expressions can be established in a similar manner to Lemma 5.

Lemma 6. (Derivatives of the proxy loss function) Consider the loss model (Definition 6). Suppose that the primal problem (16) and dual problem (17) have finite optimal values, attained at unique points $\boldsymbol{\alpha }^*$ and $\boldsymbol{\lambda }^*$ , respectively. Suppose further that the function $\mathcal{L}$ of (15d) is differentiable at $\boldsymbol{\pi }^*$ . Then the approximate loss value $\mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)$ , given by (15b), is differentiable with respect to the problem data $\boldsymbol{\pi }^*(\boldsymbol{{r}})$ with:

(29) \begin{align} \frac{\partial \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)}{\partial A_{i,j}^*} = \alpha _i^*(\boldsymbol{{r}})\lambda _j^*(\boldsymbol{{r}}),\quad \frac{\partial \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)}{\partial b_{j}^*} = -\lambda _j^*(\boldsymbol{{r}}), \quad \frac{\partial \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)}{\partial c_{i}^*} = 1 -\alpha _i^*(\boldsymbol{{r}}). \end{align}

Before moving onto proposing approximate error bounds for the proxy loss function, it is useful to consider some limitations of the approach so far discussed. We have appealed to Rademacher’s theorem in order to justify a calculus-based approach to error analysis. This required establishing the Lipschitz property for the optimal value function defining the liabilities. However, Rademacher’s theorem only establishes differentiability almost everywhere in a neighborhood. Specifically, for a given model $\boldsymbol{{r}}\mapsto \boldsymbol{\pi }(\boldsymbol{{r}})$ , it is not directly implied that the loss function $\mathcal{X}(\boldsymbol{\pi }(\boldsymbol{{r}}))$ is differentiable with respect to $\boldsymbol{\pi }(\boldsymbol{{r}})$ for almost all $\boldsymbol{{r}}$ . For future research, it may be interesting to investigate whether an extension of Stepanov’s theoremFootnote ¹² to functions of the extended reals could be used to infer differentiability for almost all $\boldsymbol{{r}}$ under the conditions of Lemma 4 applied to $\boldsymbol{\pi }(\boldsymbol{{r}})$ .

In the following, we appeal to the locally Lipschitz property to motivate the use of derivative expressions for posing of error bounds. The expression for the partial derivatives of $\mathcal{X}$ with respect to components $A_{i,j}$ of $\boldsymbol{{A}}$ required attainment and uniqueness of the primal problem (16) and dual problem (17). When primal optimality is not attained, it may indicate that improvements to the model design could be made, as nonattainment indicates no optimal matching portfolio exists. However, uniqueness may not hold in practical applications, even when a linear optimization has been used in the model specification. Practitioners faced with nonuniqueness may consider the following options:

• In certain circumstances, it may be possible to pose an economically equivalent formulation of the model that exhibits uniqueness.

• Consider adding a small positive-definite quadratic penalization term of the decision variables to the objective function. This achieves uniqueness and the Lipschitz property of the optimal value function, but at the cost of having to reformulate equations for the derivatives used within the error analysis.

• Motivated by Theorem 5.1 of De Wolf & Smeers (Reference De Wolf and Smeers2021), concerning the generalized derivative of the optimal value function with respect to $\boldsymbol{{A}}$ , practitioners may explore accepting nonuniqueness and treat derivative calculations as elements of a generalized subgradient derivative. In this setting, feasible solutions can be explored to find maximum approximate bounds.

Next, we consider the specific form of the loss model, with its associated derivatives, to pose approximate error bounds on the proxy loss.

4.3 Posing approximate error bounds for the proxy loss

Suppose the loss function $\mathcal{X}$ is differentiable with respect to the approximate data $\boldsymbol{\pi }^*(\boldsymbol{{r}})$ for almost all $\boldsymbol{{r}}\in E$ . In the following development of approximate error bounds, we recall that the symbols $\approx$ and $\lessapprox$ , defined in Notation 1, mean that the approximations are being posed as having potential practical application, and that strict mathematical equality or inequality is neither claimed nor proven. By definition (28), the exact, but unknown, approximation error is given by

(30a) \begin{align} |x^*(\boldsymbol{{r}}) - x(\boldsymbol{{r}})| = \left | \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right) - \mathcal{X}(\boldsymbol{\pi }(\boldsymbol{{r}})) \right |. \end{align}

By posing the tangent plane as an approximating function, we may equivalently write,

(30b) \begin{align} |x^*(\boldsymbol{{r}}) - x(\boldsymbol{{r}})| &\approx \left | \sum _{i,j} \frac{\partial \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)}{\partial A^*_{i,j}}\left(A^*_{i,j}-A_{i,j}\right) \right. \nonumber \\& \left .+\sum _j \frac{\partial \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)}{\partial b^*_{j}}(b^*_{j}-b_{j}) + \sum _i \frac{\partial \mathcal{X}\left(\boldsymbol{\pi }^*(\boldsymbol{{r}})\right)}{\partial c^*_{i}}(c^*_{i}-c_{i}) \right |. \end{align}

By bounding the approximation through the triangle inequality, using the expressions (29) for the partial derivatives, and applying known bounds (27a), (27b) and (27c) on the data proxies from Assumption 3, we may pose an approximate error bound for the loss function proxy:

(31) \begin{align} |x^*(\boldsymbol{{r}}) - x(\boldsymbol{{r}})| \lessapprox \sum _{i,j} \alpha _i^*(\boldsymbol{{r}})\lambda _j^*(\boldsymbol{{r}})\varepsilon \left(\boldsymbol{{r}};\,A^*_{i,j}\right) +\sum _j \lambda _j^*\varepsilon \left(\boldsymbol{{r}};\,b^*_{j}\right) +\sum _i |1-\alpha _i^*|\varepsilon \left(\boldsymbol{{r}};\,c^*_{i}\right). \end{align}

Observe also that the error bound (31) is local – that is, it changes across the risk factor space $E$ . The local refinement of the error estimate is a desirable feature of the method. However, in some circumstances, the approximation may not be effective. These may include: cases where data error bounds have not been chosen prudently, cases where the actual errors in the data approximation are significant, and in cases where the matching portfolio has been misidentified.

4.4 Approximate errors bounds for percentile estimators

We now consider the approximate error bound in the Monte Carlo context of a simulation-based capital model with the use case of estimating capital requirements. Take, for example, a simulation-based capital model that has been calibrated to form the one-year horizon loss distribution. Under Solvency II, the capital requirement is then given by the 99.5% percentile loss. Given an independent sample of the risk factor space $\{\boldsymbol{{r}}_i\}_{i=1}^N$ , we may consider forming approximate error bounds on capital estimates as follows. For each risk factor scenario, calculate the loss proxy and the estimated error bound and use (31) to pose approximate lower and upper bounds on the (unknown and exact) loss scenarios:

(32a) \begin{align} l_i \,\,\lessapprox \,\, x_i \,\, \lessapprox \,\,u_i \end{align}

where the approximate lower and upper bounds on the exact loss are given by

(32b) \begin{align} x_i \,:\!=\, x(\boldsymbol{{r}}_i ) \end{align}

(32c) \begin{align} l_i \,:\!=\, x^*(\boldsymbol{{r}}_i) - \varepsilon \left(\boldsymbol{{r}}_i;\,x^*\right) \end{align}

(32d) \begin{align} u_i \,:\!=\,x^*(\boldsymbol{{r}}_i) + \varepsilon (\boldsymbol{{r}}_i;\,x^*) \end{align}

with $\varepsilon (\boldsymbol{{r}};\,x^*)$ , for $\boldsymbol{{r}}\in E$ , denoting the upper bound in (31) so that:

(32e) \begin{align} \varepsilon (\boldsymbol{{r}};\,x^*) \,:\!=\, \sum _{i,j} \alpha _i^*(\boldsymbol{{r}})\lambda _j^*(\boldsymbol{{r}})\varepsilon \left(\boldsymbol{{r}};\,A^*_{i,j}\right) +\sum _j \lambda _j^*\varepsilon \left(\boldsymbol{{r}};\,b^*_j\right) +\sum _i |1-\alpha _i^*|\varepsilon \left(\boldsymbol{{r}};\,c^*_i\right). \end{align}

Consider the lists $\{l_i\}_{i=1}^N$ , $\{x_i\}_{i=1}^N$ and $\{u_i\}_{i=1}^N$ individually and sort them each into increasing order, written $\{l_{(i)}\}_{i=1}^N$ , $\{x_{(i)}\}_{i=1}^N$ and $\{u_{(i)}\}_{i=1}^N$ where $l_{(i)}\leq l_{(i+1)}$ , $x_{(i)}\leq x_{(i+1)}$ and $u_{(i)}\leq u_{(i+1)}$ for all $i=1,\ldots,N-1$ .

Following the approach of Lemma 2 of Crispin & Kinsley (Reference Crispin and Kinsley2022), we pose the ordered approximate bounds as having potential practical application in estimating bounds on the unknown ordered losses:

(33) \begin{align} l_{(i)} \,\,\lessapprox \,\, x_{(i)} \,\, \lessapprox \,\,u_{(i)}, \end{align}

for $i=1,\ldots,N$ . In particular, consider the basic percentile L-estimator $\xi =x_{(k)}$ , where $k=\lceil 0.005\times N\rceil$ is the index associated with the 99.5% loss, due to our convention that positive values of $x$ are gains in own funds. Under our sign convention, an estimate of the capital requirements is given by $-\xi$ . The value of the estimator is unknown, since $\{x_i\}_{i=1}^N$ is unknown and is infeasible to calculate (Assumption 2). However, under the assumption that the approximate error bounds are mostly effective, or represent bounds to first order, we may pose approximate bounds on the percentile estimator:

(34) \begin{align} l_{(k)}\,\, \lessapprox \,\, \xi \,\,\lessapprox u_{(k)}. \end{align}

Note, as before, that the symbol $\lessapprox$ does not imply that the bounds are mathematically true, only that the expressions are being posed as potential bounds.

4.5 Error attribution creates a mechanism for proxy model refinement and error reduction

Consider, as an example, the situation where a practitioner has identified that the error bounds calculated in (32a) and (33) and aims to further improve the estimate of the 99.5% loss through refinement of the data proxies.

For single loss scenario $i$ , the summation form of the error term in (32e) can be used in a natural manner to attribute the approximate error $\varepsilon (\boldsymbol{{r}}_{\boldsymbol{{i}}};\,x^*)$ across each of the data proxies $A_{i,j}^*$ , $b_j^*$ , and $c_i^*$ : the proxy $A_{i,j}^*$ can be attributed the error $\alpha _i^*(\boldsymbol{{r}}_{\boldsymbol{{i}}})\lambda _j^*(\boldsymbol{{r}}_{\boldsymbol{{i}}})\varepsilon \left(\boldsymbol{{r}}_{\boldsymbol{{i}}};\,A_{i,j}^*\right)$ , and similarly for the proxies $b_j^*$ and $c_i^*$ . This attribution can be used to identify proxies with disproportionate attributed error for improvement.

When considering improvements to error estimates of ordered losses $x_{(k)}$ both upper and lower bounds must be considered separately, since the index that sorts lower and upper approximate bounds may not be equal. One approach to refinement is to consider attribution and refinement to the approximate error interval formed from the approximate lower and upper bounds as follows. Suppose $u_{(k)}=u_{i_k}$ for some $i_k$ and $l_{(k)}=l_{j_k}$ for some $j_k$ . Then from (32a) there holds:

(35) \begin{align} u_{(k)} - l_{(k)} = \underbrace{x^*\left(\boldsymbol{{r}}_{\boldsymbol{{i}}_{\boldsymbol{{k}}}}\right) - x^*\left(\boldsymbol{{r}}_{\boldsymbol{{j}}_{\boldsymbol{{k}}}}\right)}_{\text{Non-attributable}} + \underbrace{\varepsilon \left(\boldsymbol{{r}}_{\boldsymbol{{i}}_{\boldsymbol{{k}}}};\,x^*\right)}_{\text{Attributable}}- \underbrace{\varepsilon \left(\boldsymbol{{r}}_{\boldsymbol{{j}}_{\boldsymbol{{k}}}};\,x^*\right)}_{\text{Attributable}}. \end{align}

The first term, $x^*\left(\boldsymbol{{r}}_{\boldsymbol{{i}}_{\boldsymbol{{k}}}}\right) - x^*\left(\boldsymbol{{r}}_{\boldsymbol{{j}}_{\boldsymbol{{k}}}}\right)$ , contains no direct proxy error, and so cannot be improved or attributed to individual proxies by direct methods; proxy error manifests in this expression indirectly through the possible misidentification of the true risk factor index corresponding to $x_{(k)}$ . The second and third expressions, given by $\varepsilon \left(\boldsymbol{{r}}_{\boldsymbol{{i}}_{\boldsymbol{{k}}}};\,x^*\right)$ and $ \varepsilon \left(\boldsymbol{{r}}_{\boldsymbol{{j}}_{\boldsymbol{{k}}}};\,x^*\right)$ respectively, may however be attributed to individual proxies using their summation-form as above. With a (partial) attribution of approximate error to each data proxy in place, proxies with outsized error attribution can be chosen for improvement. The approximate error bound on the proxy loss function in (31) exhibits some intuitive properties that assist refinement by attributing zero error to proxies in many circumstances:

• Assets assigned to the matching portfolio satisfy $\alpha _i^*=1$ . For such assets, errors in their valuation data $c_i^*$ are not propagated. The approximation errors in asset values of assets assigned to the matching portfolio offset since they occur on both asset and liability sides of the balance sheet.

• Assets not assigned to the matching portfolio satisfy $\alpha _i^*=0$ . For such assets, the approximation errors $A_{i,j}^*$ are not propagated.

• Nonbinding constraints are identified by having a zero dual value: $\lambda _j^*=0$ . Therefore, whenever the $j$ th constraint is not binding, errors associated with $b_j^*$ and $A_{i,j}^*$ are not propagated.

Improvements to proxies may be made in different ways. Analysis of an individual data proxy may identify that a different proxy technique is desirable. For example, a proxy may be chosen to exhibit a certain discontinuity to accommodate discontinuous relationships between data and risk factors. Alternatively, the family of functions the proxy is drawn from may be suitable, and the proxy improved with the use of extra fitting points, lowering the associated proxy error estimate. Whether refinement of the chosen data proxies increases computational burden will depend on the form of improvement chosen.

4.6 Use of approximate error bounds in practice

Practitioners may face several situations where approximate bounds on capital estimates could be of assistance:

• Approximate bounds may give evidence that existing capital estimates are effective by strengthening the body of evidence supporting their effectiveness in their use cases.

• Prudence can be applied to the approximate bounds. If these prudent bounds are suitably validated, exact methods of error analysis can be performed. Lemma 1 and Theorem 1 of Crispin & Kinsley (Reference Crispin and Kinsley2022) imply that the proxy error in the capital estimate can be bounded, and the method of targeted exact computation may apply resulting in the elimination of proxy error from the estimate.

• Approximate bounds may be attributed across the individual data proxies to identify where improvements to data proxies may be most useful, and then to use improved proxies to further refine the approximate capital bounds (Section 4.5).

• The proxy loss model and approximate error bounds could be used to cross-validate separately developed proxies, including those aimed at directly approximating the loss function whose development and validation may have encountered the curse-of-dimensionality.

If approximate error bounds are not in agreement with existing capital estimates, this may justify further validation work on production proxy models. Next we consider a stylized example, chosen for its simplicity to illustrate posing approximate error bounds.

5.1 Model definition

Let $\boldsymbol{{r}}$ denote a risk factor scenario, a realization from the random variable $\boldsymbol{{R}}$ . Let $T$ denote the year-end corresponding to the maximum cashflow date arising from assets or liabilities, and let $\{t_j\}_{j=0}^m$ be the year-end dates such that $t_0\lt t_1\lt \cdots \lt t_m=T$ . Period $j$ refers to the time interval $[t_{j-1},t_j)$ . Assets are indexed by $i=1,\ldots,n$ , with market value $v_i(\boldsymbol{{r}})$ , and have risk-adjusted expected cashflows in period $j$ denoted $C_{i,j}(\boldsymbol{{r}})$ . Let $l_j$ be the expected liability cashflow for period $j$ , where positive values represent monies that must be paid (negative liabilities represent monies to be received). Denote the discount curve for tenor $t$ by $DF(0,t)(\boldsymbol{{r}})$ .

A matching portfolio is formed from assigning it a portion of each asset. Denote by $\alpha _i\in [0,1]$ the proportion of asset $i$ assigned to the matching portfolio. Then the matching portfolio can be represented by the vector $[\alpha _1,\alpha _2,\ldots,\alpha _n]^T$ and has market value given by $\sum _{i=1}^n \alpha _i v_i(\boldsymbol{{r}})$ .

The notion of matching is captured by placing constraints on the allowable portfolio. Here we follow the example of Berry & Sharpe (Reference Berry and Sharpe2021), where the PRA’s Test 1 and Test 3 matching constraints are formulated as follows.

First, denote the present value of the accumulated asset and liability cashflows, up to period $j$ , by $C_{i,j}^{\rightarrow }(\boldsymbol{{r}})$ and $l_j^{\rightarrow }(\boldsymbol{{r}})$ , respectively:

(36) \begin{align} C_{i,j}^{\rightarrow }(\boldsymbol{{r}}) = \sum _{k=1}^j C_{i,k} DF(0,t_k)(\boldsymbol{{r}}), \end{align}

(37) \begin{align} l_j^{\rightarrow }(\boldsymbol{{r}}) = \sum _{k=1}^j l_k DF(0,t_k)(\boldsymbol{{r}}), \end{align}

where $DF(0,t_k)(\boldsymbol{{r}})$ is a (stochastic) discount factor, specified later in (47), and the arrow notation denotes the cumulative present values. Then the Test 1 and Test 3 matching constraints are represented by

(38) \begin{align} \text{Test 1:} \quad \sum _{i=1}^n \alpha _i C_{i,j}^{\rightarrow }(\boldsymbol{{r}}) - l_j^{\rightarrow }(\boldsymbol{{r}}) \geq -0.03 l_m^{\rightarrow }(\boldsymbol{{r}}),\quad j \in \{1,2,\ldots,m-1\}, \end{align}

(39) \begin{align} \ \ \text{Test 3:} \quad \sum _{i=1}^n \alpha _i C_{i,m}^{\rightarrow }(\boldsymbol{{r}}) - l_m^{\rightarrow }(\boldsymbol{{r}}) \geq 0. \qquad\qquad\qquad\qquad\qquad\qquad\qquad\end{align}

Note that an $m$ -th Test 1 constraint would be redundant due to the Test 3 constraint, so is omitted. Note also that this example is further stylized by the omission of the PRA’s Test 2, and by its simplification of the Test 1 to one-sided inequality constraints. The insurer minimizes the cost of the matching portfolio by solving the corresponding linear optimization problem:

(40a) \begin{align} \text{minimise}\quad \sum _{i=1}^n \alpha _i v_i(\boldsymbol{{r}}) \end{align}

(40b) \begin{align} \text{ subject to}\quad \begin{cases} \sum _{i=1}^n \alpha _i C_{i,j}^{\rightarrow }(\boldsymbol{{r}}) - l_j^{\rightarrow }(\boldsymbol{{r}})\geq -0.03 l_m^{\rightarrow }(\boldsymbol{{r}}), j\in \{1,2,\ldots,m-1\}, \\[4pt] \sum _{i=1}^n \alpha _i C_{i,m}^{\rightarrow }(\boldsymbol{{r}}) - l_m^{\rightarrow }(\boldsymbol{{r}}) \geq 0, \\[4pt] 0\leq \alpha _i \leq 1, i\in \{1,2,\ldots,n\}. \end{cases} \end{align}

This may be written in standard form as

(41a) \begin{align} \text{minimise}\quad \boldsymbol{{c}}^T(\boldsymbol{{r}}) \boldsymbol{\alpha } \end{align}

(41b) \begin{align} \text{subject to}\quad \boldsymbol{{A}}(\boldsymbol{{r}})\boldsymbol{\alpha } \geq \boldsymbol{{b}}(\boldsymbol{{r}}), \boldsymbol{\alpha }\geq \boldsymbol{0} \end{align}

with $\boldsymbol{{A}}(\boldsymbol{{r}})\in \mathbb{R}^{m+n,n}, \boldsymbol{{b}}(\boldsymbol{{r}})\in \mathbb{R}^{m+n},\boldsymbol{{c}}(\boldsymbol{{r}})\in \mathbb{R}^{n}$ and $\boldsymbol{\alpha }\in \mathbb{R}^{n}$ where

(41c) \begin{align} \boldsymbol{{A}}(\boldsymbol{{r}}) & = \begin{bmatrix} C_{1,1}^{\rightarrow }(\boldsymbol{{r}}) & \quad C_{2,1}^{\rightarrow }(\boldsymbol{{r}}) & \quad \cdots & \quad C_{n,1}^{\rightarrow }(\boldsymbol{{r}}) \\[3pt] C_{1,2}^{\rightarrow }(\boldsymbol{{r}}) & \quad C_{2,2}^{\rightarrow }(\boldsymbol{{r}}) & \quad \cdots & \quad C_{n,2}^{\rightarrow }(\boldsymbol{{r}}) \\[3pt] \vdots & \quad \vdots & \quad & \quad \vdots \\[3pt] C_{1,m-1}^{\rightarrow }(\boldsymbol{{r}}) & \quad C_{2,m-1}^{\rightarrow }(\boldsymbol{{r}}) & \quad \cdots & \quad C_{n,m-1}^{\rightarrow }(\boldsymbol{{r}}) \\[3pt] C_{1,m}^{\rightarrow }(\boldsymbol{{r}}) & \quad C_{2,m}^{\rightarrow }(\boldsymbol{{r}}) & \quad \cdots & \quad C_{n,m}^{\rightarrow }(\boldsymbol{{r}}) \\[3pt] -1 & \quad 0 & \quad \cdots & \quad 0 \\[3pt] 0 & \quad -1 & \quad \cdots & \quad 0 \\[3pt] & \quad & \quad \ddots & \quad \\[3pt] & \quad & \quad & \quad -1 \end{bmatrix}, \, \boldsymbol{{b}}(\boldsymbol{{r}}) = \begin{bmatrix} l_1^{\rightarrow }(\boldsymbol{{r}}) - 0.03l_m^{\rightarrow }(\boldsymbol{{r}}) \\[3pt] l_2^{\rightarrow }(\boldsymbol{{r}}) - 0.03 l_m^{\rightarrow }(\boldsymbol{{r}}) \\[3pt] \vdots \\[3pt] l_{m-1}^{\rightarrow }(\boldsymbol{{r}}) - 0.03l_m^{\rightarrow }(\boldsymbol{{r}}) \\[3pt] l_m^{\rightarrow }(\boldsymbol{{r}})\\[3pt] -1 \\[3pt] -1 \\[3pt] \vdots \\[3pt] -1 \end{bmatrix} \nonumber \\[3pt] &\boldsymbol{{c}}(\boldsymbol{{r}}) = [v_1(\boldsymbol{{r}}),v_2(\boldsymbol{{r}}),\ldots, v_n(\boldsymbol{{r}})]^T \text{ and } \boldsymbol{\alpha } = [\alpha _1, \alpha _2, \ldots, \alpha _n]^T. \end{align}

Up to now, the model closely follows that given by Berry & Sharpe (Reference Berry and Sharpe2021). We now extend the example by specifying how the asset values, cashflows, and liabilities may depend on risk factor components. The exact forms have been chosen for simplicity, rather than realism, to aid the explanation of the techniques.

Denote the components of the risk factor $\boldsymbol{{r}}$ by $[r_0,s_1,\ldots,s_n]^T$ . The first component $r_0$ represents an interest rate risk factor, and the subsequent components represent spread risk factors, in the following sense. Let the asset prices $v_i(\boldsymbol{{r}})$ be given by

(42) \begin{align} v_i(\boldsymbol{{r}}) = \sum _{j=1}^m C_{i,j}\exp \left (-(\bar{s}_i+s_i)j\right ) \end{align}

where the initial spread $\bar{s}_j$ is implicitly defined by

(43) \begin{align} v_i = \sum _{j=1}^m C_{i,j}\exp \left (-\bar{s}_i j \right ) \end{align}

where $v_i$ is the current value of asset $i$ . The currently observed discount curve $t\mapsto DF(0,t)$ can be expressed through implicitly defined continuously compounded annualized rates $\bar{r}_j$ through the following relationship:

(44) \begin{align} DF(0,t) = \exp \left ( -\int _0^t r(s)\,\mathrm{d}s \right )\quad \text{where}\quad r(s)=\bar{r}_j \quad \text{whenever}\quad s\in \big(t_{j-1},t_j\big]. \end{align}

For year-end dates $t_j$ there holds $t_{j}-t_{j-1} = 1$ , so the expression simplifies to give:

(45) \begin{align} DF\left(0,t_j\right) = \exp \left (-\sum _{k=1}^j \bar{r}_k\right ). \end{align}

Therefore, the rates $\bar{r}_j$ are given by

(46) \begin{align} \bar{r}_j = \log \frac{DF\left(0,t_{j-1}\right)}{DF\left(0,t_j\right)}. \end{align}

The risk factor $r_0\in \mathbb{R}$ is defined to be an additive spread to the current risk-free annualized rates $\bar{r}_k$ so that the discount curve, considered as a function of the risk factor, is given by

(47) \begin{align} DF\left(0,t_j\right)(\boldsymbol{{r}}) = \exp \left ( -\sum _{k=1}^j \left(\bar{r}_k+r_0 \right)\right ). \end{align}

Under a risk factor scenario the present values of the cumulated cashflows are assumed to be given by

(48) \begin{align} C_{i,j}^{\rightarrow }(\boldsymbol{{r}}) = \sum _{k=1}^j C_{i,k} \exp \left (-\sum _{k=1}^j \left(\bar{r}_k + r_0\right) \right ), \end{align}

(49) \begin{align} l_j^{\rightarrow }(\boldsymbol{{r}}) = \sum _{k=1}^j l_k \exp \left (-\sum _{k=1}^j \left(\bar{r}_k + r_0\right) \right ). \end{align}

Finally, we assume the risk factor $\boldsymbol{{r}}$ is distributed as a multivariate normal $N(0,\Sigma )$ , where the covariance matrix $\Sigma \in \mathbb{R}^{n+1,n+1}$ is given by

(50) \begin{align} \Sigma _{i,j} =\begin{cases} \sigma _i^2, & 0\leq i,j \leq n, i=j,\\[3pt] \sigma _0\sigma _j \rho _a, & i=0,1\leq j \leq n, \\[3pt] \sigma _i\sigma _0 \rho _a, & 1\leq i \leq n,j=0, \\[3pt] \sigma _i \sigma _j \rho _b, & 1\leq i,j \leq n, i\neq j, \end{cases} \end{align}

with $\rho _a,\rho _b\in [-1,1]$ specified such that $\Sigma$ is positive definite. With the stylized example now defined, we next calculate a numerical example demonstrating that data proxies and associated error analysis may have potential application.

5.2 Conclusions from the numerical example

Perhaps surprisingly, the approximate error bound on ordered losses (34) is found to hold with equality for all ordered losses in the numerical example. In the results of the numerical example, depicted in Fig. 2, we observe that:

Figure 2. The panels show the result of two Monte Carlo loss simulations of the stylized model (Section 5.1) with example data (Example 5.1). Each row of panels depicts the same data. The right-hand column of panels shows the full ordered loss data with the left-hand column scaled to show detail of the ordered loss data around the 0.5th loss percentile. On the panels, blue lines show the ordered exact loss, and the green shaded region is indicates ordered loss values between the approximate lower and upper error bounds. The blue dot is the exact loss at the $0.005\times N$ ordered loss, representing the 0.5th loss percentile. The green horizontal lines below and above the dot show the associated approximate lower and upper bounds. That the blue line always lies within the shaded region shows that, in this example, the approximated bounds hold mathematically. Panels A1 and A2 show a Monte Carlo simulation with $N=10,000$ with data proxies chosen to be Chebshev interpolation with 4 points. The error attribution process is applied (Section 4.5) and all proxies attributed with more than 1% of the upper or lower approximate error bounds are refined – they are rebuilt with 7 Chebyshev points. Panels B1 and B2 show the Monte Carlo loss simulation with the refined proxies and $N=1,000,000$ . The reduced approximate error bounds show that error measurement and attribution can be used as a mechanism for error reduction in percentile estimates.

• Approximate error bounds can be calculated at percentiles relevant for use cases of the proxy model. Using error attribution, practitioners may target improvements to individual proxies to achieve acceptable levels of approximation error across use cases.

• The approximate bound for unordered losses (32a), denoted $\lessapprox$ , may be observed to hold with mathematical inequality in practical settings. Sensitivity analysis shown in Table 1 indicates the importance of choosing data proxies prudently.

• The approximate ordered bounds for ordered losses (33) are wider than the unordered bounds, therefore, these may also hold with mathematical inequality. While sensitivity analysis indicates these bounds are less sensitive to prudence, it remains important that error bounds for data proxies are chosen carefully.

• The approach to attribution identified in Section 4.5 and (35) results in the identification of proxies whose associated error contributes significantly to the error bounds. When these proxies are improved, a reduction in the error bound is observed. This can be seen as a systematic mechanism for proxy design improvements and for the reduction in approximation errors propagating into percentile estimators and capital estimates.

• When designing data proxies, practitioners need to consider the computational cost of calculating interpolation and validation points. The error attributed to an individual proxy may be used to decide whether further computational resources should be applied to improvements in the proxy.

• It may be feasible to improve proxy error bounds through making a different choice of approximation or interpolation method, without further increasing the computational cost of the proxy’s construction or validation.

In the numerical example, two distinct Monte Carlo loss simulations are performed. The loss calculations are based on the stylized model (Section 5.1) with example data (Example 5.1). In the initial simulation, data proxies are constructed with four interpolation points. The choice to use so few interpolation points is to create proxy errors that will propagate into percentile and capital estimates, and reflects that in practical settings, computational considerations may impose a choice of which proxies most interpolation points are assigned. In the first simulation, only 10,000 scenarios are used, representing a setting where a practitioner may be interested in testing and refining the initial design of data proxies without the use of large Monte Carlo runs. The initial simulation is depicted graphically in Panels A1 and A2 of Fig. 2. Panel A2 shows the exact loss calculation lies within the approximate proxy error bounds at all ordered loss scenarios. The specific 0.5th percentile loss is highlighted with the approximate bounds. Panel A1 depicts a subset of the same scenario data around this 0.5th percentile loss to aid inspection of the approximate error bounds around this percentile.

The approach to partial error attribution to individual proxies, as outlined in Section 4.5, is applied with the basic (and somewhat arbitrary) choice to use seven interpolation points whenever the attribution of the lower or upper approximate error bounds to an individual proxy exceeds 1%. The exact choice of improvement to the proxy is arbitrary for our example, except to illustrate how improvements to individual proxies, and their associated error bounds, can be shown to propagate into improved percentile error bounds.

Panels B1 and B2 of Fig. 2 show a loss simulation with $N=1,000,000$ and confirm improvements to the approximate error bounds of ordered losses arising from the refinements to certain proxies as an outcome of the error attribution above. Both panels confirm a notable improvement to the approximate error bounds on ordered losses, and further, show that the approximate ordered error bounds still hold mathematically when compared against the exact loss calculations. A sensitivity analysis of the mathematical success of approximate ordered and unordered bounds with respect to scaling of the initial data proxy error bounds is shown in Table 1. It demonstrates increased success as estimates of the data proxy error bounds are made more prudent with the approximate ordered bounds holding in all cases without the need in this example for extra prudence.

As a computational assumption, we assume that full Monte Carlo with exact loss calculations are not feasible. Therefore, practitioners may choose to validate the performance of approximate error bounds on computationally feasible subset of scenarios representative of the loss distribution. In practical settings, the attribution of approximate error bounds to proxies may be performed iteratively to achieve the desired computational performance and  error bounds across percentiles suitable for the various use cases at hand. In the numerical example, no extrapolation points are encountered due to the significant interpolation domain chosen.

Table 2. Data for the stylized example of the analysis of error propagation and attribution in simulation-based capital models.*

* Data for the stylized example are from Berry & Sharpe (Reference Berry and Sharpe2021) except for the annualized rates $\bar{r}_j$ , shown rounded to 4 decimal places, that are derived from the discount factors $DF\left(0,t_j\right)$ through the expression (46). Positive liability cashflow values represent (expected) monies to be paid, and positive asset cashflows represent (risk adjusted) monies to be received.

Table 3. Asset price data for the stylized example of the analysis of error propagation and attribution in simulation-based capital models.*

* The stylized example market data is from Berry & Sharpe (Reference Berry and Sharpe2021). Asset spreads $\bar{s}_i$ are derived from (43) and are shown rounded to 4 decimal places.