1.1. Introduction

1.1.1 This paper is a key read for those who work in reserving and who work around reserving teams. It outlines frameworks to use for reserving validation and gives the reader an overview of current techniques being employed. In the experience of the authors, many companies lack an embedded reserve validation framework and reserve validation can appear piecemeal and unstructured.

1.1.2 Although the paper is aimed at practitioners in the field of general insurance, the validation approaches outlined in section 3 may well be relevant to actuaries working in other fields and in particular those working in health insurance who may well apply general insurance reserving techniques as part of their work. Moreover, the application of machine learning fitting techniques such as those described in section 6 of this paper may well aid validation in other areas where sufficient data are available, such as to the fitting of mortality curves to crude mortality data or the extrapolation of past mortality improvements for modelling future improvements.

1.1.3 The paper then outlines a case study demonstrating how successful machine learning techniques will become and then goes on to discuss the implications of machine learning to the future of reserving departments, processes, data and validation techniques.

What Will You Get from this Paper?

1.1.4 Reserving validation can take many diverse forms, from simple checks to full independent reviews; if done well, it adds value to the reserving process, enhances governance and increases confidence and reliability in results.

1.1.5 This paper discusses:

• The significance and importance of reserving validation

• Common weaknesses and their solutions

• Suggestions of a framework in which to apply validation tools

• Current developing techniques used in reserve validation

• Specific and unique impacts of the COVID-19 pandemic on reserving validation

What Does the Future Hold for the Reserving Actuary?

1.1.6 Reserving is entering a new era as machine learning techniques take hold. No doubt this will include a trial-and-error phase during which much dialogue will occur regarding failing and succeeding projects. The result may well be improvements in predictive capabilities the likes of which we have not seen in decades. To add to this, these methods will produce these predictive capabilities in much faster timeframes than a human could achieve leading to much needed automation of parts of the process.

1.1.7 The final section of the paper is focused on a case study on machine learning together with discussion of the implications of this technology for reserving departments, processes and validation techniques, we also discuss data implications for these processes.

Why is Validation Important?

1.1.8 Setting correct reserves is critical to insurance businesses. The results of the reserve review form a key component of the management and statutory accounts. They form the starting point for capital calculations and a crucial feedback loop into the pricing and planning process. It is critically important that management and non-executive directors have confidence in the judgements and processes underpinning the reserve booking decision and reserving actuaries need to ensure that this confidence is not misplaced. The topic of reserve risk was also highlighted in the November 2020 Dear CRO letterFootnote ¹ from the PRA which itself followed a Dear CEO letterFootnote ² and Dear Chief Actuary letterFootnote ³ on the same theme a year earlier. Clearly the area of reserve risk and hence validation within the process is of great importance to both regulators and to Boards – a focus which has been increased by the COVID-19 pandemic and consequent recession.

2.1. Reserving Validation

2.1.1 In the case of reserving validation, we are interested in ensuring the results from the reserving analyses are reasonable and appropriate. Given that perfect accuracy in the estimation of reserves is an impossible target, reserving validation looks to ensure reserves are unbiased and not inaccurate and that there is a thorough understanding of what the estimates represent and the uncertainty around these.

2.1.2 Certain validation techniques can also be used to highlight when reserves are becoming invalid in an automatic way to focus the Actuary’s attention in the relevant areas. It can take many forms, from simple checks to full independent reviews. Done well it adds value to the reserving process, enhances governance and increases confidence in the results.

2.1.3 This paper discusses the importance of reserving validation, common weaknesses in the current process and how to overcome them. We also suggest a framework in which to apply validation tools. Subsequent sections of this paper set out the current techniques used in reserving validation and the developing techniques that could be used by the reserving actuary of the future to embed validation techniques in the reserving process.

2.2. The Importance of Reserving Validation

2.2.1 Setting adequate reserves is critical to the continued solvency of insurance companies. Previous studies have shown that inadequate reserves are among the primary causes for insurer insolvency.

2.2.2 The results of reserve reviews are used extensively throughout insurance companies. They form a key component of the management and statutory accounts, the starting point for capital calculations and a crucial feedback loop in portfolio management, pricing and planning processes.

2.2.3 A robust reserving validation process can help ensure stable and unbiased reserves are set and that key stakeholders have a clear understanding of any point estimates (e.g. best estimate) as well as the key risks and uncertainties in the figures presented. It can provide the tools and information upon which to challenge and can ultimately help the board make informed decisions about the results presented and the level of reserves to hold.

2.2.4 Key areas that validation can assist with are:

1. a. During the reserve review:

   • Identifying assumptions at an early stage that are no longer valid.

   • Detecting emerging trends and risks that may need to be allowed for.

   • Identifying the most material assumptions and the sensitivity of those assumptions.

   • Selection of best-fit models.

   • Highlighting areas of concern requiring additional attention.

   • Providing an overview of areas of uncertainty in the review.

2. b. When signing off reserves:

   • Enabling users to challenge more effectively.

   • Identifying the key assumptions that impact the overall results.

   • Demonstrating that key assumptions have been challenged and plausible alternatives considered.

   • Giving clarity around point estimates and uncertainty.

   • Providing evidence of good governance and appropriate challenge.

3. c. Out of cycle

   • Providing ongoing monitoring of experience to ensure assumptions remain appropriate.

   • Providing evidence to respond to regulatory queries (e.g. Dear Chief Actuary/CRO/CEO letters).

   • Supporting/feeding into other processes.

   • Providing meaningful comparisons to market.

   • Policing of new “black box methods”.

2.3. How Validation Can Help Reserving Actuaries Overcome Current Challenges

2.3.1 This section sets out the key challenges that current reserving processes face and how a strong validation framework can help address them.

Increased Time-Pressure to Deliver Results

2.3.2 Reserving processes across the market are facing pressure to move more quickly from the cut-off date for data to delivering results. This means actuaries can be faced with difficult decisions as to where to prioritise their time. There can be a trade-off between performing a broad review of all reserving classes and methods, against the need to perform a deep dive in key problem areas. It is also desirable to spend time on forward-looking proactive analyses to pre-empt potential future issues.

2.3.3 Reserving validation can help in two ways:

1. 1. First, it can help the actuarial team determine how to prioritise their resource by helping identify problem classes and those where the previous review’s assumptions still stand. Time spent testing and then setting appropriate diagnostics and triggers can be outside of peak-review time, making the overall process more efficient.

2. 2. A similar diagnostic suite can allow a small risk team or 2nd line of defence to monitor the activity of a larger actuarial function effectively. Where previously 2nd line reviews might follow a significant deterioration, a strong reserving validation framework can allow a risk team to engage pro-actively with the actuarial function.

Perceptions of Stakeholders

2.3.4 Stakeholders, such as Finance Directors and Non-Executive Directors, may have perceptions about the reserving process which influence how they view the results of a review. A common stereotype is the assumption that actuaries include an inbuilt margin or element of caution into reserves, perhaps because giving good news is easier than bad.

2.3.5 Sometimes management concerns in this field are justified, but they can also arise when they are not justified. For example, a class that sees improvements 9 out of 10 years in succession might be prudently reserved, but equally the distribution of losses may be skewed (the mean being far above the median).

2.3.6 It is essential that stakeholders have a clear understanding of the reserves, as well as the main risks and uncertainties within them.

2.3.7 A strong reserving validation framework will help detect trends in the reserving strength, pre-empt challenge from stakeholders and to provide clarity as to where pre-conceptions are valid or not.

Increased Regulatory Scrutiny

2.3.8 There has been increased focus on general insurance companies over recent years as demonstrated by the PRA’s Dear Chief Actuary, Dear CEO (both Nov 2019) and Dear CRO (Nov 2020) letters.

2.3.9 Dear Chief Actuary letter (November 2019): Covered a variety of reserving issues including possible biases in reserve assessment (upside is allowed for but downside risks less so). Concerns over weakening of case reserving across several classes, for example,Medical Malpractice and US General Liability. The PRA also raised concerns that there was inadequate allowance for claims inflation and concern that information provided to Boards may not be understood due lack of transparency.

2.3.10 Dear CEO letter (November 2019): Concern that Boards and management are not receiving sufficient data or information to fully understand key uncertainties to facilitate challenge of the reserving assumptions before sign-off. The letter also highlighted that some actuarial functions remained undeveloped and so reserving, pricing and claims departments were not always consistent in setting of assumptions and that detection of trends are also slower. There was also a general concern that underwriting remediation was being allowed for prematurely in the reserving process even if there was a lack of evidence to suggest the re-underwriting was having a positive impact. General concern that poor infrastructure and data systems meant that firms are also unable to identify trends or react to emerging risks in a timely manner.

2.3.11 Dear CRO letter (November 2020): Although targeted at risk functions, reserving processes continued to be flagged as a key concern in the paper. In particular indications of possible weakening of case reserve estimates and biases in reserve estimates (concerns that worsening claims experience is not being incorporated to setting of reserves). The PRA set out expectations for risk functions to provide evidence of how they challenge the reserving processes and results, and how the risk function is helping to ensure that stakeholders such as the Board understands the uncertainty in the reserves before being signed off.

2.3.12 Regulators expect companies to consider their reserves carefully, and by including it in both the “Dear CEO” and “Dear CRO” letters, it is clear they are looking for better reserving engagement from outside of the Actuarial Function. Reserving validation forms an important part of this engagement.

Commercial Pressures

2.3.13 Reserving actuaries have always had the pressure of balancing regulatory and commercial pressures. This is not a new problem but one which is becoming more prevalent. For example, Lloyd’s Decile 10 has put poor performing classes in the spotlight whereby underwriters will be keen to improve these classes and want to see benefit taken in the reserving process quickly. However, this is in direct conflict with regulatory pressures faced by reserving actuaries who need to ensure deterioration in experience is being allowed for in the latest reserving exercises.

2.3.14 This is also a concern highlighted by the PRA as outlined above in the Dear CEO letter, with the regulator hinting that firms should expect reserving to remain “important area of supervisory focus for us in the months ahead, particularly for firms with material exposure to longer-tail casualty lines, firms that show a poor track record or reserving developments compared with initial assumptions and those who have shown rapid growth”

2.3.15 With the increase in availability of data, enhanced computing power and development of new techniques, there is competitive advantage to be had by automating processes, focusing effort in the value-add areas and ultimately understanding your business better. Given how important reserving analyses are to the business, reserving validation is an essential part of this.

2.4. COVID Provisions as an Example of Validation under Stress

2.4.1 The global pandemic had a significant impact on the insurance market as a whole in 2020 and the impact varied by type of firm (personal lines/London market) and by class of business (e.g. motor has performed well whereas contingency and financial lines e.g. Directors & Officers (D&O) or Political and Financial Risk have seen deteriorations).

2.4.2 This made reserving for affected classes difficult given the sensitivity and subjectivity of assumptions (e.g. duration of lockdown and multiplicity of lock downs), and it tested our standard validation techniques. For example, where reported claims were lower than expected (typically a measure of improved experience), it was not clear whether this was simply a delay in claims notification as a result of lockdown or a genuine improvement in claims experience. The impact of COVID on some of our standard validation techniques will continue long after the reserving assumptions return to a “normal” level as it has disrupted the prerequisite assumption that data sets are relatively stable.

2.4.3 Reserving actuaries are under the spotlight more than ever to effectively communicate key uncertainties and assumptions to both the Board as well as external stakeholders such as auditors and regulators and validation forms an important part of this.

2.4.4 The pandemic has impacted other departments, for example, capital, pricing of future policies and wordings. This has put additional pressure on the actuarial function to ensure a consistent approach in methodology and assumptions in determining COVID provisions and validation of the reserves set for these losses. Overall, the impact of COVID has brought into focus the limitations of our current validation techniques.

2.5. Validation Framework

Current Status of Validation in the Industry

2.5.1 Reserving validation is not a new topic, most firms have been completing some form of validation on their reserves for many years, using techniques such Actual versus Expected analysis and comparing their reserves to market benchmarks. The wide range of validation approaches used is described in detail in section 3. The challenge is that these are often applied piecemeal and have significant limitations when only looked at individually.

2.5.2 Overall, there is little standardisation to the validation of reserves in the insurance market, and there is often no formal validation framework used. This contrasts with many firms’ approaches to capital validation, where the methodology is reasonably homogenous across the industry and the objectives are well defined and agreed upon.

2.5.3 If a firm does not yet have a formal validation framework, we recommend developing one.

Why Adopt a Validation Framework

2.5.4 A robust validation framework for reserving is important to ensure the reserves booked by an insurance company are reasonable and appropriate for the risks written by the insurance company This is more true today than ever before given factors such as

• Increased regulatory scrutiny (Dear CEO, CRO and Chief actuary letters)

• The impact of COVID-19 and increased economic uncertainty in the aftermath

• Climate change and other ESG risks

2.5.5 Each validation method (set out in section 3) has distinct strengths and weaknesses. Applying techniques in a scatter-gun approach risks duplicating effort in some areas of reserving validation, whilst potentially missing other missing signs of reserve inadequacy elsewhere.

2.5.6 Where validation is unstructured, the emphasis is on the reviewing actuary to judge where validation tools are most needed and then apply them. For example, there might only be time to conduct scenario analysis or backtesting of assumptions on one or two specific classes and limited other validation information to assist with the selection of those classes.

What Makes a Good Validation Framework?

2.5.7 A good validation framework provides appropriate assurance that the reserves are appropriate in the most efficient way possible. A well-designed validation framework is set up in advance and includes a full “toolkit” to assess the validity of reserves. This ensures the right tools are available at the right time to both:

• Monitor the portfolio in aggregate

• Perform detailed validation in high-risk areas.

2.5.8 We classify validation techniques into three categories:

1. 1. Numerical validation

2. 2. Contextual validation

3. 3. Governance validation

2.5.9 A good validation process will include aspects of all three to ensure complete coverage of the reserves. This is an example of the “swiss cheese model” where each layer of validation has flaws (holes in the cheese), but each slice of cheese has the holes in difference places, so there is no direct path through the multiple layers.

2.5.10 Figure 1 sets out a suggested structure for a reserving validation process, with key methods added to each stage.

Figure 1.

Suggested structure for reserving validation process

3.1. Reserving Validation Toolkit

3.1.1 Validation activity is an integral part of the actuarial control cycle and forms an indispensable part of the reserving process. Reserving validation is not therefore a new topic; most firms have been completing some form of validation on their reserves for many years, irrespective of whether or not these activities are called “validation” or carried out as part of a structured reserving validation framework. A wide range of validation methods already exist and have wide use across the market, assisting firms with the following:

• Validity of data

• Appropriateness of the data segmentation into projection classes

• Suitability of the projection methodology

• Reasonableness of the key assumptions and judgements

• Reasonableness of movements in ultimate between reviews

• Assessment of uncertainty in best estimate

• Appropriateness of the booked reserve results (if different to the best estimate).

3.1.2 Figure 1 in section 2.5.9 sets out some of the main approaches used by reserving teams to assist with reserving validation following the structure of a proposed optimal validation framework. This section discusses these current approaches to validating reserves, together with their strengths and limitations.

3.2. Numerical Validation Methods

3.2.39 Ultimate Tracking

Method Overview

3.2.40 The ultimate liability estimated at a preceding valuation date may be revised at the current valuation date. This revision can arise due to any of the following reasons:

• Updating data for the current period (for, e.g. adding the diagonal to the Chain ladder method)

• Change in the method

• Change in the assumptions

3.2.42 The impact of the first reason will be captured in AvE diagnostics. For instance, a reduction in the actual claims compared to expected can be inferred as a positive movement resulting in reduction of the ultimate liability. This shows an underlying link from the AvE to the movement in ultimate.

3.2.40 However, the other reason also needs to be monitored and their impact on the ultimate liability should be estimated and explained to stakeholders. The results of Movement Analysis can aid in discussion with stakeholders, where they can provide additional feedback and corroborate with our findings. We can explain the movement in form of a waterfall chart where the reader can understand the most important component leading to change in ultimate:

3.2.43 Figure 2 shows the change in the reserving method has mainly contributed to an increase in ultimate from 90 to 100. This change in the ultimate is monitored on an accident year/underwriting year basis for different classes of business. There can be higher volatility in ultimate liability for the less mature years or the long-tailed classes of business.

Figure 2.

Movement analysis on movement in ultimate claims

3.2.44 Helping users to understand the movements helps them to understand the drivers of the model which allows them to validate the model more easily and sense check the movements against how they expect the figures to move based on experience.

3.3. Contextual Validation Methods

3.4. Governance Validation

3.4.1 This section provides a brief examination of the role of the governance process in validating the reserves. From the perspective of the reserving team, these validation activities are downstream activities after the team’s internal validation is complete. However, the ultimate purpose of validation is to ensure appropriate reserves are set and thus help protect policy holders. From this standpoint, the three lines of defence model and governance-based validation remains important. As such, we discuss it at a high level for completeness although the main focus of this paper is on more technical validation.

We consider the following:

1. 1. Assumptions challenge

2. 2. Reserving committee

3. 3. Three lines of defence model (including the role of 2^nd line review, internal audit and external reviews)

Assumptions Challenge

3.4.2 A powerful form of validation is the process for agreeing and challenging key assumptions with the wider business during the reserving process. An effective process ensures that the views of all stakeholders are captured, ensuring the reserving team does not operate in a silo.

3.4.3 Business challenge can take many forms, from informal conversations with interested parties through to formal assumption setting committees.

3.4.4 From a validation perspective, it can be useful to carry out this type of internal review in at least a semi-structured way. This might include

• Having a key assumptions log and a set review schedule to review and debate key assumptions (eg inflation rates) with the business.

• Formalising discussions around key assumptions in regular working groups or committees.

• Documenting the discussions to ensure that appropriate evidence in support of parameter setting is captured. This is standard practice for capital model validation, but not yet fully adopted in reserving.

3.4.5 Some of the challenges associated with assumption setting groups are

• Participants often lack a consistent basis to compare views or data. For example, claims teams considering inflation on a settlement basis, pricing actuaries on an underwriting basis and reserving actuaries on an accident basis.

• It can be difficult to validate an individual assumption out of context of the overall reserving process.

• There may be the perception that some participants have a vested interest in setting assumptions higher or lower than best estimate.

Reserving Committee

3.4.6 The reserving committee is a key stage in the challenge of reserves for any insurer. The purpose is to bring together expertise and knowledge from across the business to review and agree the reserving results. Many firms’ reserving committees include or are chaired by Non-Executive Directors, but this is not universal.

3.4.7 The effectiveness of validation and challenge provided by a reserving committee depends on both the range of skills represented on the committee and also on the quality of reporting.

3.4.8 In terms of committee composition, it is normal practice to bring together senior stakeholders from underwriting, claims, and finance. These function areas overlap the most with reserving and their representatives are well placed to provide detailed challenge or debate of key assumptions or areas of concern within the reserves. The role of the NEDs (non-executive directors) is to challenge the overall strength of the reserves and to ensure that it is appropriate. This challenge is important as it involves taking a step back from intricacies of individual classes, movements and assumptions.

3.4.9 Key to any successful reserving committee is the quality of reporting provided to the committee. The content and structure of reserving committee papers is, on the whole, beyond the scope of this review. However, the inclusion of good quality exhibits from a strong validation process can help stakeholders get comfortable with reserves or to challenge assumptions as appropriate.

Three Lines of Defence

3.4.10 Almost all firms operate a do/check or do/check/review process within the actuarial function itself. This is not surprising, given the IFOA’s APS X2 requirements that actuarial work should be reviewed by an appropriately skilled person. Most firms also operate a form of the 3-lines of defence model. However, there is wide variation in who acts as the 3 lines, and what the remit of each line is. Consider the two contrasting examples in Table 1:

Table 1.

Different Roles within Three Lines of Defence

Risk Team

3.4.11 The risk team often acts as the second line of defence for reserve validation. In smaller firms its role may be blended with that of internal audit (see below), with the principal form of oversight via a control framework. For larger firms a more hands-on approach is more commonplace, with the risk team forming its own independent opinion on all or part of the reserves or performing deep dives to validate the first line assumptions in key areas.

3.4.12 The risk team is perhaps the team that can benefit most from the adoption of a technical validation framework that includes many of the technical validation methods discussed already in this section. Adopting a strong validation framework can help an (often small) risk function fulfil its second line of defence role more effectively by helping focus its efforts on at-risk classes or the most uncertain assumptions. This can be much more proportionate than simply re-projecting key classes using similar methods and judgements to the first line.

Internal Audit

3.4.13 Third line independent testing of the reserving processes provides assurance on controls and governance. It is an additional safeguard against material risks that could be missed due to process related oversights such as data quality. It can also assess whether reserve results are presented in a way that is appropriate for the audience to make the best use and judgement of reserving information.

3.4.14 However, the scope of a third line internal audit is not particularly detailed, where the focus is usually on reserving process and operation rather than reserve selections. This is because internal auditors tend to have an accountancy and audit background rather than being actuaries.

3.4.15 Actuarial resource may be called upon for some Internal Audit reviews, but this may be an actuary from another part of the business (as an internal co-source arrangement) or an external actuarial consultant. They are also generally asked to focus on the compliance with the relevant Reserving Policy and Control Validation testing (with line 1), rather than challenging the reserve model assumptions and selections.

3.4.16 Such reviews would tend to be infrequent and may be conducted only when issues are identified or at the request of Board, Audit and Risk Committees or when required to by a regulator. So, it is more reactive rather than proactive at identifying and dealing with issues. Typically, an internal audit review could focus more on the reserving process rather than the reserving results themselves – that is, the remit being to check the robustness of process – with such an approach reflecting the background and viewpoint of internal auditors in practice.

External Audit

3.4.17 Compared with internal audit, the scope of an external audit will typically include independent projections or methodology and assumptions review on company’s reserves. External audit engagements generally span several years, and the auditor may focus on different areas each year, but the insurer can discuss with the auditor the classes that should be reviewed in detail. The auditor will provide an independent view for comparison especially where independent projections are performed and can also provide comparison to an alternative approach (e.g. reserving at a less granular level). The auditor has access to market views, which provide points of comparison against the market.

3.4.18 The goal of the external auditor may be to only form a view as to whether the methodology and assumptions are unreasonable and may not always perform an independent projection and provide an alternative view. However, the auditor can make suggestions and request results based on alternative scenarios, which can be helpful for the insurer to understand its reserve risk better. Auditors may not always consider the specific issues faced by the insurer and may not go into as much detail which may make their independent estimates less useful.

3.4.19 The auditor’s market view can be particularly useful for market wide loss events to compare to market wide estimate, such as for losses related to COVID, to understand if the insurer’s approach and results are in line with the market.

Independent Actuarial Reviews

3.4.20 An external actuarial review may be requested by the Board, Audit, Risk or Reserving Committees. It is like an external audit, where an independent actuarial team performs independent projections. Because the scope is requested by the company, there is the flexibility to set the scope such that it is aligned to the company’s aims, which can include deep dives in targeted areas.

3.4.21 Unlike external audit, where the same team handles the audit engagement for several years, the actuary performing the reserve review can be selected using a tender process. This means the company can select the reviewer that has the most extensive experience in the area reviewed.

3.4.22 The actuary performing the reserve review usually does not normally have knowledge of the reserves set by the company, so should give a truly independent review of the reserving process and reserve selections. The independent actuary can provide benchmarking of both process (to best in class in market) and reserving selections (loss ratios etc.) to other insurers operating in the country, internationally or specific market segment.

3.4.23 When reserve review results differ significantly, it is commonplace to hold discussions to understand the reasons for the deviation between the internal company view and that of the external reviewer. These discussions may lead the insurer to revise the reserving assumptions in some situations. At other times deviations may be identified as providing areas for deep dives in future reserving exercises. An agreed approach or policy should be put in place prior to the review to outline what approach will be taken if the reviewers’ numbers are significantly different from the internal view. This will enable decisions to be impartial and minimise the impact of commercial pressures.

Regulatory Actuarial Reviews

3.4.24 Reserve reviews can also be conducted by regulator, which may follow a regular review cycle or be ad-hoc if concerns are identified. The regulator may use their own actuaries or commission firms of consulting actuaries to conduct the review. The regulator tends to use experienced actuaries who have wide market experience given they review the reserving of multiple insurers.

3.4.25 The Regulator usually has access to both public and private reporting (e.g. Solvency II reporting), so will also usually benchmark company metrics (loss ratio, paid to incurred ratio, reserve to premium ratio, etc.) against the whole market and other individual insurers, and provides a percentile of where the company’s metrics lie in the range of outcomes for all insurers. This can show if the company is out of line with the market. However, these measures may not always be comparable due to the reporting segmentation not being granular enough or various company specific factors.

3.4.26 The limitations identified above in the section on external reserve reviews applies here in this context as well. The review may only be done at a high level and smaller classes may be excluded from the review. Actuaries conducting the review may not have explicit knowledge of the insurer and its products. This is important when products or policy coverages are non-standard in the market as this can give rise to unusual features impacting the claims reserving which reduces the usefulness of external reserve reviews in this context.

5.1. Impact of IFRS 17 on Reserving Validation

5.1.1 IFRS 17 is the revision of IFRS 4 standard issued by the IASB (The International Accounting Standards Board) that will drastically change how business performance is measured and reported by (re)insurers globally. These changes are intended to bring standardisation and transparency within financial reporting by insurance companies.

5.1.2 IFRS 17 measures the liabilities in terms of fulfilment cashflows arising from insurance contracts. The fulfilment cashflows include estimate of future cashflows including premium, claims and expenses; besides, an adjustment for the time value of money and provision for the non-financial risks are included. Moreover, the netting-down approach utilised for reinsurance cashflows is no longer compliant and should be replaced with independent processes to source and form assumptions for reinsurance contracts held.

5.1.3 These requirements will drastically change the present processes and validations for IFRS 4 and GAAP reporting. Consequently, this will impact the work of the reserving actuary in various ways – including the approach towards validation as shown in Figures 5 and 6 below.

Figure 5.

Pre-IFRS 17 implementation example reserving workflow

Figure 6.

Post IFRS 17 implementation example reserving workflow

Working Day Acceleration Impact

5.1.4 IFRS 17 imposes additional requirements for cashflows modelling, discounting and complex allocations. All these additional steps in the workflow will put pressure on the actuarial resources and processes. Typically, Solvency II processes are completed in the last half of month of closing. But IFRS 17 requires more work than Solvency II and these have to be completed in the first week of any reporting entity. Consequently, IFRS 17 implementation will put pressure on the validation. Solvency II continues and IFRS 17 builds on these foundations.

5.1.5 The response to this challenge is yet to be finalised by companies. A strategy including the mix of automated checks and roll-forwards, reduction in manual processes, continuous validation (as compared to expost-facto (after the event) validation) appears to be the best way forward.

IFRS 17 Adds More Layers to Validation

5.1.6 IFRS 4 and GAAP focuses on point estimate of reserves within liability. In contrast, IFRS 17 aims to provide transparency regarding financial and non-financial risk elements within financial statements and these will need validation. Consequently, the below apply.

5.1.7 Validation has multiple layers and is no longer linear like in IFRS 4. The AvE actual versus expected net claims experience is the key driver of profitability under IFRS 4. In contrast, IFRS 17 liability and income perspectives are dynamic in that it considers cashflows view factoring the time value of money, financial and non-financial risks, often acting together. Also, IFRS 17 unpacks additional elements, and these will have P&L impact and more public disclosure – for example, unwind of discount rate.

5.1.8 IFRS 17 includes assumptions regarding compensation for uncertainty, that is, the risk adjustment. Therefore, the current validation metrics will need changes to support a risk and time value adjusted view aligned to IFRS 17 and will need validation.

5.1.9 Most of the current validation techniques rely on gross and net of reinsurance metrics. However, IFRS 17 unpicks this approach and requires separate validation and tracking of inwards business and outwards reinsurance.

Renewed Focus on Unearned Exposures

5.1.10 The current validation for IFRS 4 focuses on earned business. This is because the unearned business liability is simply the unearned premium with any provision for the additional unexpired risk result (AURR). The AURR provisions are typically nil because they are calculated at an aggregated level that offsets profitable and unprofitable business.

5.1.11 But under IFRS 17 there are specific requirements around how the groups of contracts are identified and the losses in one group of contracts cannot be offset against the other. The losses at initial measurement have to be recognised at day zero within the liabilities and flows directly into the P&L. Consequently, IFRS 17 puts unearned business under spotlight and new validation metrics for unearned assumptions will become important and prevalent.

Analysis of Change

5.1.12 IFRS 17 disclosures and presentation requires the break-down of the movement of various components of liability and traces their movement into P&L. For example, the liability for remaining coverage (LFRC), liability for incurred claims (LFIC), contractual service margin (CSM), loss components (LC) are all broken down in terms of movements within components. The drivers of changes in liability are explained as experience against expected, economic variances and premium variances. Therefore, IFRS 17 will require the creation of additional validation metrics to support with the analysis of change and to explain them to external and internal stakeholders. For example, coverage units under General Measurement Model are dependent on the level of benefits and extent of coverage provided by the insurance company. There is monitoring, tracking and validation of the coverage assumed (for example, sum assured, PML) to be provided – all these will be new requirements for insurance companies and will require provision of additional data fields and corresponding meta-data. Some of the current validation techniques like movement analysis should be transformed to be line with IFRS 17 view of movements.

Hybrid KPIs

5.1.13 IFRS 17 reporters are likely to amend metrics to an IFRS 17 basis, like insurance service, revenue and expenses, CSM, risk adjustment and loss component within their internal reporting and planning. This will facilitate a good link between how results are reported externally and how performance is measured internally. Alongside these metrics and KPIs, many IFRS 4 metrics will be retained because of demand from business and investors. For example, the measures like gross written premium, combined ratios provide a sense of continuity. However, they are likely to be adapted for IFRS 17. The expenses within combined ratios will be the IFRS 17 view of attributable expenses.

Validation

5.1.14 Under IFRS 4, there was no strict requirement for reserving validation. Historically, the validation has often been completed expost-facto, at specific points within the reserving workflow, and sometimes after important decisions had been taken that had a bearing on the valuation of liabilities. However, because of the integrated structure of IFRS 17 described above, this approach to validation may not be feasible. IFRS 17 may need controls that can predict and are more responsive to an evolving risk within the broad environment of the company. Consequently, the emerging view of validation under IFRS 17 has the features below:

• Automated: Because of the working day pressures, the sheer number of assumptions and complex hand-offs between teams, automation of validation is recommended.

• Continuous: Rather than an expost-facto view of validation, a continuous approach is required because there are a number of off-cycle processes that have an important bearing on the quarter-end/annual results.

• Adaptive: There are non-linear impacts arising within IFRS 17 because the cumulative impact of several risks acting together are just as important as single stress events; and

• Predictive: The validation metrics, measures and diagnostics have a forward-looking view – taking into consideration the future impacts arising from the current play of risk factors.

5.1.15 More generally, we consider these to be desirable characteristics of a validation process, whether reporting under IFRS 17 or not.

5.2. Early Warning Indicators for Adverse Development

5.2.1 With growing demands for explanation of variability around actuarial best estimate results, there is significant interest in identifying the business drivers of adverse development, and thereby understanding leading indicators of future adverse development as well to calibrate reserve estimates accordingly.

5.2.2 The following KPIs serve as leading indicators to signal worse than expected claim frequency and/or IBNR. This is important for reserving validation as an indicator or areas for further analysis. This section is here because of the increasing use of feedback loops to set forward-looking assumptions is becoming increasingly important in the validation tool box:

Mix of Business in a Leading Indicator of Claims, and Claims are Leading Indicators of Loss Ratio Performance

• Worse than expected mix of business (skewed more towards high-risk segments)

  • ○ For example, in Workers’ Compensation (in the US) or Employers Liability, if the mix of business is skewed towards high-risk industries/class codes

  • ○ Similarly, in Motor Liability or Motor Physical Damage, if the mix of business is skewed towards high-risk territories/age groups

• Actual versus expected policy counts and premiums in different risk segments (low/medium/high)

  • ○ Risk is defined based on predictive model scores, where high risk = high expected loss ratio

• Higher than expected number of claims during initial earned exposure months

  • ○ For example, in Professional Indemnity, or Cyber Liability, if a particular trend in claims leads to a consistent upward trend in reported claims, relative to historical years

  • ○ Similarly, in personal/commercial auto, if the number of reported claims within a policy year is higher than the “seasonality-adjusted” average of prior historical years

• Actual versus Expected large losses, early in the policy life cycle

  • ○ Ultimate claims incurred amounts are lagging indicators due to the tail on development

  • ○ Claims predictive models are used to predict the claim severities

5.2.3 Figure 7 shows an example of how the combined effect of a worse-than expected mix of business and a higher-than-expected number of claims during initial exposure months was used to forecast IBNR loss ratios (accounting for adverse development), with a margin of uncertainty using Bayesian predictive modelling methods. An example set of different insurers from different industry segments is presented in Figure 7.

1. 1. A multivariate Bayesian frequency model (using size, industry, territory, limit and deductible) was developed to estimate the number of claims per unit of exposure (measured in terms of revenue)

2. 2. That multivariate model was applied to future earnings of exposures to forecast claim counts for each policy, based on future Earned Revenues

3. 3. A multivariate Bayesian severity model (using industry, territory, limit and deductible) was developed to forecast the ultimate severity for claims forecasted in step 2 above. This gave loss forecasts for each policy

4. 4. The orange bars represent the predicted range of unadjusted loss ratios, calculated by dividing the total losses forecasted, by the unadjusted premiums. The green bars represent the adjusted loss ratios, after the premiums were adjusted to account for prospective rate adjustments based on rating plan changes

Figure 7.

Predicted unadjusted loss ratio v loss ratio (adjusted for rate change)

5.2.4 The following KPIs serve as leading indicators to signal worse than expected claim severity IBNR:

• Worse than expected mix of claims (skewed more towards high-risk claim types)

  • ○ For example, in Workers’ Compensation (US) or Employers’ Liability, if the mix of claims is skewed towards high-risk injuries (head injuries, motor vehicle accidents, etc.)

  • ○ Similarly, in Motor Liability or Auto Physical Damage, if the mix of claims is skewed towards high-risk claims such as multi-party claims, or legally represented claims

• Worse than expected claim severities during initial earned exposure months

  • ○ For example, in Motor Liability or Professional Indemnity, if the average settlement amounts in initial development periods indicate a consistent severity uptrend driven by social inflation

5.2.5 Figures 8, 9, 10 and 11 are the results of the multivariate Bayesian severity model mentioned in step 3.

• X axis represents the age of the accident year (in months)

• Y axis represents the severity (actual and forecasted) where the divergent coloured lines represent forecasted severity

Figure 8.

2018 Severity charts

Figure 9.

2019 Severity charts

Figure 10.

2020 Severity charts

Figure 11.

2021 Severity charts

6.1. Analysis of Trends in Reserving Validation: Future Outlook

Why Validate

6.1.1 At the core of validation is the analysis of expectation from actuarial method that requires this to be best estimate; unbiased; stable; responsive and independent. It has to be comprehensive in order to provide sufficient coverage.

Questions to Answer While Validating Reserve Results

6.1.2 In earlier sections, we have reviewed the questions to address when validating reserve results. These include measurement of estimation error; appropriateness to specific circumstances and provision of insights into relative strengths and weakness of methods under circumstances such as through the COVID-19 pandemic. However, we would like to add the following goals based on our experience across commercial and personal lines in both a Lloyds and non-Lloyd’s context:

• Diagnostics of results against experience such as the table proposed in section 3.2.42

• Avoidance of methodological bias and other systematic errors

• Early warning for appropriate action – these can be developed using development factor heat aps such as that shown in section 4.1.6 in the context of COVID-19

• Visualisation of key risks and insights

• Striking the right balance in terms of estimation bias; variance and accuracy

• The implementation of methods that as simple to explain, easy to measure and understand

• The adoption of a goal of the avoidance of cognitive bias

• Clarity regarding appropriate levels of prudence

• Adoption of clear assumptions regarding stationarity

What Stages to Apply Validation

6.1.3 We note that there has been great progress in development of validation methods over recent years, and that this has of necessity taken a significant jump forward in response to the COVID-19 crisis and the inability of reserving models to respond dynamically to the data and development challenges of COVID-19. Although it should be noted that the topic of validation was raised in the PRA Dear Chief Actuary letter of November 2019 (which pre-dated the emergence of COVID-19 by several months). Validation should not be an ex-post facto exercise and look at, for example:

• Performance testing on statistical predictions, and

• Performance testing on (for example, clinical predictions) (including the actuary’s judgment) within the validation process.

How to Address the Weakness of Current Validation Methods

6.1.4 In order to develop a view of the future of validation, the weaknesses of current validation methods need to be identified and addressed. This could be achieved by techniques such as enhanced movement analysis to avoid anchoring and biases, embedded stress and scenario testing, and the use of sensitivity test libraries from reserve reviews and stress testing of reserves on a quarterly basis. This will highlight the uncertainty around the best estimate reserves and provide business insight into reserve adequacy.

6.1.5 As discussed in section 3.2.3, it is also important to look at approaches such as scenario and sensitivity testing around the key assumptions being used on both a gross and net basis. This will include development patterns, initial expected loss ratios (IELRs) and RI (return on investment) ratios. It is also important to look at scenario testing, and work with the business to determine key scenarios that may impact the reserves of a company. This should include large losses and catastrophe (cat) losses. This will bring more focus on reinsurance and net ultimate incurred losses with the use of the capital model to remodel complicated reinsurance structures and how changes in gross positions change RI, and to investigate key percentiles for assessing reserve risk. It is important to address then known weaknesses of actual versus expected (AvE) tracking, and also to recognise that constantly adding additional diagnostics may lead to a point of diminishing returns being reached.

Leveraging New Tools for Expanding the Actuarial Repertoire for Validation

6.1.6 Cross learning and best practices from other fields can help reserving actuaries move forward in terms of validation. Machine learning (ML) predictive models are increasingly being used to validate assumptions that provides an approach to improve validation by introducing a degree of automation into the validation process. Typical validation questions can then become whether the mean of implied frequency and severity models from the reserving process fall in the ballpark of a ground up claim-level frequency/severity ML model?

6.1.7 We can transplant validation techniques or use ML as diagnostics. This allows cross-validation techniques to become important, with a need to consider that these are commonly used in statistics and ML. It is very useful to measure model, parameter and process risks as part of the validation process. This can be used to enhance the following techniques in use: AvE; back testing and variants of cross validation (K-Fold Covariance analysis etc.). At present, these only measure the error of statistical predictions and are difficult to perform for clinical predictions (for example), and there are challenges with application of serial correlation.

6.1.8 The use of clustering has been suggested to identify homogeneous groups and have AvE at the granularity of homogenous risk groups – this helps to avoid Simpson’s paradox (2021) (i.e. where a trend appears in several groups of data but reverses or disappears when they are combined). Auto roll-forward approaches and dynamic weightings are growing in popularity as part of the trend to use clustering to select homogenous risk groups for AvE analysis.

6.1.9 Bootstrapping, sensitivity and scenario analysis are mainstays of validation and can be supplemented by more advanced visualisation. Such advanced visualisation approaches to validation include partial dependence plots, receiver operating characteristics, shapley value and SHAP dependence plots.

6.1.10 Looking to the future is also important. Big data (i.e., extremely large data sets that can be analysed computationally to reveal patterns) and IFRS 17 are topics of great interest in respect of their impact on validation (and IFRS 17 and validation processes was covered in detail in section 5). These lead to the need for the ability to perform reserve validation at a much more granular level (e.g., claim level, market segment level) and to ask whether particular cohorts of claims or particular market segments (geographies, industries, etc.) are driving any observed reserve deterioration.

6.1.11 Companies and consultants are both investigating techniques such as ML. This creates new opportunities for validation and cross learning opportunities. Advances in natural language processing bring the automation of the following into scope:

• Checks and balances between assumption setting for reserving, planning, capital, and pricing:

  • ○ Frequency-severity assumptions implied by reserving (at a granular level) are being compared against frequency-severity assumptions implied by pricing

• Dynamic dashboards for reserving:

  • ○ Big data can now be aggregated real-time thanks to much better processing power, and this is being used to feed dynamic dashboards that break the down the implied reserving assumptions and allow for better assumption tracking

  • ○ Are the implied loss ratios (say at an industry/geography level) moving away from the pricing assumptions, and is this movement within an acceptable threshold?

6.1.12 The design of early warning indicators of under reserving and deterioration is important. These include

• Use of big data/granular data and ML to provide early warning indicators

  • ○ For example: if number of litigated claims is higher than expected, even if the case reserves may not have developed yet, it is an early warning to increase (incurred by not enough reported) IBNER

• Claims/UW ML predictive models predict a higher-than-average number of “complex” claims, or project higher-than-average frequency, the IBNR should reflect those projections even if the loss experience does not reflect yet

• There should also be a quarterly feedback loop between the loss experience and the ML models

6.2. Why Machine Learning is Important to the Future of Reserving

In this section, we do not seek to explain how Machine Learning methods work or give a guide as to how they should be used. Instead, we use an example to demonstrate that they are capable of successfully being as predictive as humans to demonstrate to the reader our conclusion that they are important to the reserving actuary of the future.

In this section, we discuss two developments that we foresee will impact reserving individually but to a greater extent together.

More Predictive Reserving Methods from Machine learning (ML) Algorithms

6.2.1 The first development is on the reserving methodology side. Within the next ten years several carriers will develop new reserving methods through the application of ML techniques that will see improvements in predictive capabilities that we have o’t seen in the last 40 years. Not only that, but these methods will also be automated, meaning that the time taken to make these predictions will be greatly reduced. The following section aims expand on this.

6.2.2 The Arch actuarial team have kindly allowed us to publish the results of an exercise that they carried out in conjunction with a data science specialist consultancy called Augmize. Again, we do not give all the details as to how this work was implemented – we merely use the results as a case study to prove our point. Alastair Lauder oversaw this project while working at Arch and this is how it was introduced to the working party. If you would like an original copy of the paper, you should contact Steven Loyens the Chief Actuary at Arch Insurance.

6.2.3 The work involved carrying out an experiment using Lloyd’s risk code data. The Arch Insurance actuarial team provided a database of historic risk code projections going back to 2016 carried out by actuaries.

6.2.4 The accuracy of the human projections was tested against that of ML algorithms built to predict the ultimate premiums and claims for risk code data. For this machine learning software with interpretable AI for granular loss modelling platform for creation and maintenance of reserving and pricing models was used (supplied by Augmize).

6.2.5 The purpose of this section is not to give an in-depth view of the methodology but to demonstrate that these algorithms provide a similar level of accuracy to that of a human actuary.

Dataset

6.2.6 Risk code data which spanning 1999–2019 with 2100 unique classes of business and account years. Models were trained on a core bed of data spanning 1999 to (end of) 2015. Training data were then added incrementally for more recent origin years to compare to those estimates made by actuaries.

6.2.7 Figures 12 and 13 demonstrate what information was used by the ML algorithms.

Figure 12.

Comparison of actuary versus machine learning

Figure 13.

Breakdown of predicted ULR

6.2.8 Models were based on premiums, paid and incurred triangles along with the risk code mapping text data in order to assign labels to the data.

6.2.9 The simple error metric is taken as % error of the prediction versus actual ultimate loss from history. Two different calculations of actual ultimate loss were used. However, there are many, and it is one area that requires more analysis and discussion as to the pros and cons of each way of assessing the error.

6.2.10 Error method 1: Chasing known ultimates from history. This method filters for examples of historical classes that have fully developed, which form ground truth examples of Ultimate Loss. Here we may challenge both the ML and actuary to predict these ultimates and measure absolute errors of human and machine. Note that overall this limits the number of data points because only fully developed years are relied upon.

6.2.11 Table 2 shows average errors of the ML are equivalent to the actuary, within +/− figures showing the ML to be more consistent in the ultimate predictions. At low development the ML performs very well, with significant out performance at the development ranges 0–3 years.

Table 2.

Errors for Machine Learning versus Actuary

6.2.12 The difference in performance at high development is relatively small with a plus or minus tolerance roughly equal to the error introduced in filtering for historical ultimates in the data selection for the backtest. The final diagonal for the actuary performs well, brought about again by the backtest filtering process, where we define the ultimate in the same way that an actuary would, adding bias towards the humans estimates.

6.2.13 We highlight that the ML generated results are completely automated, forming compelling opportunities for scalable and fast business-user analysis.

6.2.14 Figure 14 shows the ML average error performance, with win-lose-same frequency visualised along the bottom row. Figure 15 shows the actual average errors. The low development portion of the triangle is circled showing the larger errors from the actuary in deep red.

Figure 14:

Machine Learning average errors

Figure 15:

Actual selected average errors

Figure 16:

Machine Learning average errors

Figure 17:

Actual selected average errors

6.2.15 At low development the ML outperforms the actuary. Also note that the shaded standard deviation is higher for the actuary than ML suggesting that the ML deals better with the difficulty faced in predicting ultimates at low development.

6.2.16 Error method 2: Chasing the most recent actuarial ultimate. This method takes the actuary’s latest selected ultimate as the value to measure error against. This gives more data points to measure than error method 1; however, it is unknown if the actuaries ultimate is accurate or not.

Table 3:

Errors for Machine Learning v Actuary

6.2.17 The results are similar to error method 1 where overall the ML error is similar to the actuaries’ error, however, it is better at lower development and worse in the middle development.

6.2.18 The same pattern emerges where the ML is better at lower development and not as good in the middle development. Note that this means that the ML predictions converge to the most recent actuarial estimate faster at lower development periods.

6.2.19 The above result is slightly counterintuitive. One suggestion, however, is that when the actuary models a single risk code at low development, they do not really factor in any data from other risk codes. However, the ML will make up for the lack of data by using data from all other risk codes. This increases accuracy and lowers standard deviation. This could also explain why the actuary is better in the middle of the development. The actuary is focused on that particular triangle, however, the ML is still trying to use information from all other classes. If the actuary more successfully predicts the result, it implies that in practice that particular triangle is more important than the ML is giving it credit for. This theory is still under investigation.

Implications for the Reserving Process

Table 4:

Implications of Machine learning to the reserving process

6.2.22 Within the reserving process, there are multiple uses possible:

• Use as a validation tool

• Replace actuarial estimates

• Fully automated projections with Actuarial review

6.2.23 There is an obvious link to validation and in the first instance these methods will likely be compared to older methods in order to validate them or used as a validation tool for the old methods. Over time people will gain more comfort that these methods actually are more predictive in some circumstances than allowing humans to make judgements.

6.2.24 It will not be possible to carry out fully automated or unsupervised reserving for the vast majority of classes of business due to the complexity of the reserving problem and the inadequate data that actuaries are using to carry out their reserving processes.

6.2.25 Instead, the real value from these ML methods will be in aiding the actuary to be more predictive. It will be a combination of machine supervised by human which will succeed in this space. This will mean that understanding these methods will be a must within reserving teams. Moreover, regulators and auditors will need solid explanation of how the algorithms are making decisions.

6.2.26 Note that these methods favour bigger carriers with larger data sets rather than small. This provides large insurers with another advantage relative to smaller insurers – when arguably solvency 2 has already created an advantage for larger carriers through economies of scale (including but not limited to diversification benefits).