Model and Assumption Uncertainty in ORSA

Why Uncertainty Matters in an ORSA

An ORSA projection is a reasoned estimate of how an insurer’s financial position may develop, not a guaranteed account of the future. Its outputs depend on information available at the valuation date, methods used to transform that information, judgments made where evidence is limited, and assumptions about future conditions. If those foundations are uncertain, a precise-looking capital figure may convey more confidence than the evidence warrants.

The task is not to eliminate uncertainty or to replace an assessment with a range of numbers. It is to identify material uncertainty, explain how it could affect projected solvency, and make clear which conclusions are robust and which rely on particular choices. This tutorial is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

For clarity, let \(C_t\) denote projected available capital at time \(t\), and let \(K_t\) denote the internal capital need on the same basis and at the same time. The projected surplus is \(M_t=C_t-K_t\). When \(K_t>0\), the ratio \(C_t/K_t\) is one way to summarize the relationship between resources and need. These measures are useful only if their definitions, scope, and timing are consistent. Uncertainty can affect \(C_t\), \(K_t\), or both.

Where Uncertainty Enters

A practical assessment separates uncertainty by source. The categories can overlap, but separating them helps explain what evidence is missing and what kind of analysis could reveal the effect.

• Data uncertainty concerns completeness, accuracy, relevance, or consistency of recorded information. Examples include missing exposure characteristics, uncertain claim status, or data collected under definitions that changed over time.
• Parameter uncertainty concerns the values used within a selected model, such as an estimated frequency, severity, trend, or correlation. Even with the model structure fixed, estimates based on limited or changing experience may be imprecise.
• Model uncertainty concerns whether the model structure is an adequate representation of the relevant relationships. Two reasonable structures can produce different projections from the same data because they treat development, diversification, or other relationships differently.
• Expert-judgment uncertainty arises when informed judgment is used because evidence does not uniquely determine an input. The range of reasonable judgments, their rationale, and the authority for selecting one should be visible.
• Scenario-assumption uncertainty concerns the specification of a hypothetical future condition: its severity, timing, duration, or interaction with other conditions. A scenario is conditional; it is not a forecast that the assumed event will occur.

These categories should not be treated as interchangeable. For instance, a data correction may change an estimate without showing that the model structure is wrong. A model comparison may reveal structural sensitivity, but it does not prove which model is correct. A scenario sensitivity shows the consequence of changing a specified condition; without a probability model, it does not assign a likelihood to that consequence.

Tracing Uncertainty Into Solvency Measures

A useful assessment links each uncertainty to the output it could affect. A reserve input might change projected liabilities and therefore available capital. A change in risk calibration might affect internal capital need. A judgment about the timing of a management action might affect both, depending on how the projection is constructed. Record the direction and approximate magnitude where they can be supported, and state where the effect cannot be quantified reliably.

The arithmetic behind a solvency sensitivity is direct. If an uncertain input \(\theta\) changes available capital and capital need, then the change in surplus is the change in capital less the change in capital need:

This identity explains why examining only available capital can be misleading. If a revised assumption reduces capital but also reduces capital need, the net effect on surplus is smaller than the capital change alone. Conversely, a modest capital reduction accompanied by a larger increase in capital need can make the overall solvency position materially weaker. The ratio must also be recalculated using both revised quantities; it is not generally changed by the same percentage as either one.

Comparing Models and Judgments

When model structure is uncertain, a disciplined comparison holds the valuation date, projection horizon, economic scenario, and output definitions constant as far as practicable. Differences in results can then be traced to structural choices rather than to unrelated changes in inputs. The comparison should describe why the alternative is credible, what evidence supports each approach, and which components of the result differ.

A simple average of two model results is not automatically a better estimate. Averaging is meaningful only if there is a defensible basis for treating the models as components of an estimate and for assigning weights. Otherwise, report the results side by side, explain the material drivers, and avoid implying that the midpoint has a special evidential status.

Scenario and Expert-Judgment Uncertainty

Scenario assumptions can be uncertain even when the arithmetic conditional on them is straightforward. A scenario may specify a loss, a recovery, or a timing assumption. The ORSA should distinguish the assumed condition from the evidence supporting its severity and from any judgment used to translate it into financial effects. If several assumptions are changed together, it may be difficult to tell which one drives the result; isolated sensitivities can help when they are appropriate to the question.

What a Sensitivity Can Establish

A sensitivity is a conditional calculation: it answers what the selected measures would be if a specified input changed while the stated remaining assumptions were held fixed. It can demonstrate materiality, identify which output is exposed, and help prioritize further investigation. It does not, by itself, show how likely the alternative is, whether the input range is complete, or whether interactions with other inputs are negligible.

For a small change in a continuous input \(\theta\), a local approximation may be useful if the output is sufficiently smooth near the selected value:

The derivative describes local sensitivity around the current input; it is not a probability or a guarantee that the same response applies to a large change. For nonlinear models, thresholds, or interactions, calculate the alternative scenario directly rather than extrapolating a local sensitivity without checking it.

A useful uncertainty record connects the source to the decision-relevant output. It can include the uncertain item, its evidence and limitations, the selected assumption, credible alternatives, affected measures, the method used to quantify the effect, and any unresolved limitation. The analysis should also distinguish a quantified range from a qualitative concern. If a limitation cannot be measured reliably, stating that explicitly is more informative than presenting a false-precision estimate.

Common Traps and Practical Implications

• Reporting a point estimate as certainty. A projected capital figure is conditional on inputs, methods, and assumptions. State material dependencies alongside the number.
• Confusing sensitivity with likelihood. An alternative scenario’s result does not imply that the scenario is probable unless a probability basis has been established.
• Changing several items without attribution. If data, model structure, and scenario severity all change at once, the effect of each may be obscured. Explain the combined nature of the comparison or separate the changes where feasible.
• Using inconsistent bases. Ratios and surplus comparisons require capital and capital need to share timing, scope, and definitions.
• Ignoring uncertainty because it is hard to quantify. A limitation can still be material. Describe its direction if known, the reason quantification is difficult, and how much confidence it should remove from the conclusion.

Uncertainty analysis matters because solvency assessment is used to understand resilience, not merely to produce a single projected figure. A conclusion is more credible when its important dependencies are visible and when the analysis distinguishes evidence-based estimates from judgments and conditional scenarios. The objective is not to claim certainty, but to make the basis and limits of the assessment understandable.