Policy Significance Statement
The current study supports the claim that extortion payments by ransomware victims should not be considered insurable by cyber insurance providers that use assumptions of classical ruin theory in their solvency determinations. Cyber insurers and re-insurers should examine and consider reforming the practice and policy of covering extortion demands as part of insurance claims related to ransomware attacks. This study also encourages an evidence-based approach for creating and applying cyber risk solvency standards by insurance regulators to ransomware-related extortion payments.
1. Introduction
The already $16 billion yearly demand for insurance coverage for cybersecurity incidents is only expected to increase in the coming years, with more frequent incidents and evolving regulatory requirements worldwide (FBI, 2024; Fortune Business Insights, 2025). The incidents driving losses are varied, including, but not limited to, distributed denial of service (DDoS) attacks, business email compromises (BECs), network loss, data breaches, financial fraud, and ransomware attacks. In particular, ransomware—i.e. malicious software or “malware” used for extortion—actors continue to target critical sectors, including healthcare and public health, government facilities, and information technology organizations (FBI, 2024; Verizon, 2024). In 2023, 32% of data breaches involved some type of extortion technique, including ransomware, and over the last five years (2019–2024), reported ransomware attacks have continued to increase and accounted for around 20–30% of insurance claims (FBI, 2024; SouthPoint Risk, n.d.; Verizon, 2024).
General insurance has two primary tasks: pricing and solvency. While insurance premiums are designed to reflect portfolio-wide risk, insurance providers may still face challenges if capital reserves and reinsurance strategies are insufficient to absorb tightly aggregated events over time. This study is concerned with solvency and issues that arise for cyber insurance providers and modelling catastrophic events, particularly for ransomware attacks and extortion payments. In insurance, solvency is assessed using numerous risk factors and determinations for regulatory capital and economic capital and is a balancing act between initial capital reserves, premiums collected, and first-party and third-party liability claims paid (Eling and Schnell, Reference Eling and Schnell2019). For cyber insurance providers, this means solvency is assessed based on risks, including cyber risk, first-party liability for insured (data recovery and restoration, customer notification, public relations assistance, business interruption reimbursements, extortion payments, etc.), and third-party liability for lawsuits from the insured’s client(s) (Eling and Schnell, Reference Eling and Schnell2019; Romanosky et al., Reference Romanosky, Ablon, Kuehn and Jones2019; Woods et al., Reference Woods, Böhme, Wolff and Schwarcz2023). Thus, economic capital from collected premiums and regulatory capital based on mandated requirements, along with initial surplus, must not be vulnerable to insolvency or ruin after payment of claims.
It should be noted, premium pricing adequacy is not a sufficient proxy for avoidance of ruin if it ignores temporal parameters such as inter-arrival rates (Hult and Lindskog, Reference Hult and Lindskog2011). In classic ruin theory, mathematical models for insolvency evaluate the risk of ruin based on the theory that premiums are collected at a constant rate while claim payments occur with values that are also stochastic and are independent and identically distributed (i.i.d.) (Hult and Lindskog, Reference Hult and Lindskog2011; Lundberg, Reference Lundberg1903). If an insurance provider must pay a claim prior to collecting sufficient premiums or finding additional customers, reserves may have to be exhausted immediately. With the diverse frequency of ransomware attacks and heavy-tailed distribution of extortion payments, complete depletion of reserves is potentially a realistic outcome, especially without sub-limits for extortion payments.
Given the increasing pattern for cybersecurity incidents to result in severe, widespread, and even systemic consequences, it is essential to better understand cyber risk to mitigate technical and operational risks and appropriately prepare for financial risks (Eling and Loperfido, Reference Eling and Loperfido2017; Eling et al., Reference Eling, McShane and Nguyen2021; World Economic Forum, 2022). This is important not only for stakeholders involved in designing safeguards and shaping standards policies but also for those assessing organizational cyber risk, such as cyber insurance providers (Eling et al., Reference Eling, McShane and Nguyen2021). More informed risk assessments can enable a more accurate estimation of risk exposure and further inform evaluations of aggregate risk across insurance portfolios, particularly where interdependencies may lead to simultaneous failures across multiple systems or organizations (Axon et al., Reference Axon, Erola, Agrafiotis, Uuganbayar, Goldsmith and Creese2023).
To accurately understand the cost of ransomware attacks for society, we must be able to identify and measure many of the dynamic issues for extortion payments—a challenge compounded by data limitations. Unlike traditional property and casualty insurance, cyber risk underwriting must contend with a rapidly evolving set of threats, making risk assessment inherently more complex and uncertain. We make three contributions to the existing literature. First, the present study provides additional statistical analyses of aspects of cyber risk, specifically the distributions of extortion payments clustered by ransomware strain/variant, using novel data on ransomware attacks and payments. Second, our findings indicate that extortion payments are not an i.i.d. process and thus violate an underlying principle of solvency in classical ruin theory, suggesting that extortion payments are not insurable by firms using such solvency models. Third, our work positions fluctuating dynamics in extortion demands from ransomware attacks as a consideration in predator-prey frameworks for cybersecurity. Lastly, our work highlights an urgent need for reconsideration of solvency standards by insurance regulators for firms that provide coverage for ransomware extortion payments.
2. Literature review
2.1. Heavy-tailed distributions and cyber insurance
Heavy-tailed distributions, i.e. distributions with tails that take longer to approach zero than exponential distributions, appear frequently in statistical analyses regarding cybersecurity threats (e.g., Bryson, Reference Bryson1974; Edwards et al., Reference Edwards, Hofmeyr and Forrest2016; Eling and Wirfs, Reference Eling and Wirfs2016; Florêncio and Herley, Reference Florêncio and Herley2011; Hult and Lindskog, Reference Hult and Lindskog2011; Maillart and Sornette, Reference Maillart and Sornette2010; Romanosky, Reference Romanosky2016; Shevchenko et al., Reference Shevchenko, Jang, Malavasi, Peters, Sofronov and Trueck2023). Such distributions can appear when investigating different types of cybersecurity threats, e.g. estimating infected devices amongst certain Internet of Things (IoT) manufacturers, as well as in studies related to actuarial assessments, insurance, and models of risk for cybersecurity threats (Biener et al., Reference Biener, Eling and Wirfs2015; Dacorogna et al., Reference Dacorogna, Debbabi and Kratz2023; Jung, Reference Jung2021; Rodríguez et al., Reference Rodríguez, Noroozian, van Eeten and Gañán2021). While there is some research that attempts to fit a single distribution to a loss model for cybersecurity threats, there is less peer-reviewed work that comprehensively examines the foundational underpinnings of why cybersecurity threats are heavy-tailed (August et al., Reference August, Dao and Niculescu2022; Kolesnikov et al., Reference Kolesnikov, Markov, Smagulov and Solovjovs2022; Laszka et al., Reference Laszka, Farhang and Grossklags2017; Santini et al., Reference Santini, Gottardi, Baldi and Chiaraluce2019). Some studies have sought to determine if different driving processes are independent and can be accounted for in modelling, though these have primarily used existing datasets related to data breaches and not ransomware incidents (e.g., Eling and Wirfs, Reference Eling and Wirfs2016; Xu et al., Reference Xu, Schweitzer, Bateman and Xu2018; Zeller and Scherer, Reference Zeller and Scherer2022). Looking at ransomware attacks in particular, prior literature includes modelling of losses for ransomware attacks with traditional assessments of aggregate cyber risk across various cybersecurity threats and has not sufficiently explored evaluating risk based on the specific type of cybersecurity threat and any unique underlying driving processes (Axon et al., Reference Axon, Erola, Agrafiotis, Uuganbayar, Goldsmith and Creese2023; Caporusso et al., Reference Caporusso, Chea and Abukhaled2019).
Heavy-tailed distributions also appear in losses associated with financially motivated cybersecurity threats over time. Over the last two years, BECs and extortion breaches (including ransomware attacks) accounted for approximately one-fourth and one-third of financially motivated attacks, respectively. Additionally, BECs had a median transaction amount of $50,000 USD in both years, while according to the Federal Bureau of Investigation’s Internet Crime Complaint Center (IC3) ransomware incident complaint data, “the median loss associated with ransomware and other extortion breaches has been $46,000 [USD], ranging between $3 [USD] and $1,141,467 [USD] for 95% of cases” (FBI, 2024; Verizon, 2024). Extortion amounts vary by ransomware strain or variantFootnote 1 due to differences in the attack behaviours and choices of targets (perhaps specific to a sector). Extortion payments also vary over time, with these temporal factors having ramifications for risk management.
Previously, Leverett et al. (Reference Leverett, Jardine, Burns, Gangwal and Geer2020) used power laws to study the characterization of ransomware-related extortion payments by their averages. In particular, the authors highlighted that data from recent years suggests that the exponent of the power laws might be testing the boundaries of the insurability of paying ransoms with insurance policies. Their findings indicated this was the result of no stable, defined first moment under recent exponents. Additionally, Leverett et al. concluded that power laws perform a good distribution fit to ransom payments due to the diversity induced by multiple sub-population distributions. This is consistent with other research on mixture models and power law tails (Patriarca et al., Reference Patriarca, Heinsalu, Marzola, Chakraborti and Kaskin.d.). Furthermore, the work of Cartwright and Cartwright predicts that a mixed distribution would be an emergent property of a diversity of differential pricing by different groups (Hernandez-Castro et al., Reference Hernandez-Castro, Cartwright and Cartwright2020).
General liability insurance to cover losses involves some type of risk assessment and related pricing. While some providers assess an overall risk, others use more advanced methods (e.g. spatial analysis, AI-driven modelling, etc.) to determine insurance tiers and premiums based on different subsets of risks. For cyber insurance providers, such granularity does not currently exist, and many providers focus coverage on cybersecurity threats as a monolith (i.e. treating DDoS, fraud, ransomware, etc. as one) despite the reality that the insureds incur different losses driven by different processes. For example, a qualitative analysis of cyber insurance policies from 5 years ago reports that the average premium for most insured parties is between $10,000 and $25,000 (USD), with some providers setting limits up to $10–$50 million (USD) and others offering extreme loss coverage for certain industries with limits in the hundreds of millions of dollars (Romanosky et al., Reference Romanosky, Ablon, Kuehn and Jones2019). At present, there is still limited publicly available information regarding how cyber risk is determined, how cyber risk changes based on industry and sector, and how premiums are ultimately determined (CISA, 2021; Romanosky et al., Reference Romanosky, Ablon, Kuehn and Jones2019; Woods et al., Reference Woods, Agrafiotis, Nurse and Creese2017).
Fundamentally, cybersecurity threats are merely technologically elevated forms of adversarial actions already in existence, such as theft, fraud, sabotage, and extortion/ransom. It follows then that cyber risk is no different than other types of uncertain adversarial risks that can be assessed as a straightforward function of threat, vulnerability, and impact (Geer et al., Reference Geer, Jardine and Leverett2020). However, because of the ubiquity of technological systems and potential for wide-ranging and compounding consequences, cybersecurity threats can not only be thought of as their own new risk class but can also require providers to reassess many other risk classes (Wolff, Reference Wolff2022). A 2023 study found that providers were nearly split as to whether cyber risk is treated as its own unique class versus incorporated into other existing risk classes (Romanosky and Petrun Sayers, Reference Romanosky and Petrun Sayers2023).
Nearly 25 years ago, Dr. Ross Anderson aptly detailed the risk issues of cyber insurance that still ring true today: “Around 2000, the end of the dotcom boom created a downswing that coincided with the Millennium bug scare. Even now that markets are returning to normal, some kinds of cover are still limited because of correlated risks. Insurers are happy to cover events of known probability and local effect, but where the risks are unknown and the effect could be global (for example, a worm that took down the Internet for several days), markets tend to fail (Anderson, Reference Anderson2001).” Additionally, previous research has demonstrated how two tiers of cyber risk consideration creates potential barriers to a healthy cyber insurance market: “the first tier is the correlation of cyber-risks within a firm (i.e. correlated failure of multiple systems on its internal network) and the second tier is the correlation in risk at a global level (i.e. correlation across independent firms in an insurer’s portfolio)” (Anderson, Reference Anderson2001; Böhme and Kataria, Reference Böhme and Kataria2006). The second tier—of interest in our study—is often known as ‘accumulation’ or ‘aggregation’ of risk in a portfolio.
Alongside a split consensus as to how cyber risk is handled, cyber insurance providers are considerably varied on exclusions to coverage. While many providers offer policies that cover aspects of first-party liability with little to no variance in costs across customers, there is a pronounced division as to whether extortion payments are covered under the same policies (Woods et al., Reference Woods, Agrafiotis, Nurse and Creese2017; Romanosky et al., Reference Romanosky, Ablon, Kuehn and Jones2019). Cyber insurance firms that choose to exclude extortion payments from coverage can simplify considerations for solvency and premium pricing, though their offerings may not be as enticing to customers who are hoping for protection on extortion payments. However, insurance providers who do offer coverage of these payments face a challenging predicament of how to model high variability in losses incurred, including inconsistent extortion amounts demanded by different groups for different victim companies within different industries and sectors, varied clustering in time, and frequency of payments.
Pricing models in insurance are designed to have an underwriting loss ratio (ULR), and in the cyber insurance industry, it is not uncommon for that ratio to be 40–45 percent. For example, “although insurers experienced a range of outcomes in 2023…overall performance held strong in the face of market softening and rebounding frequency; the 2023 loss ratio of 42 percent [wa]s the lowest since calendar year 2018” (AON, 2023). This translates to premium pricing that leads to 42 cents of loss for every dollar generated in a given year. However, the ULR for some years can be worse than others, wherein losses may even exhaust the premium reserves. In these cases, the volatility of cyber risk has a cost outwith the pricing model, and the role of capital reserves and solvency determinations helps cover this (Mildenhall and Major, Reference Mildenhall and Major2022). Solvency can be achieved with four components: 1) the evaluation of liabilities; 2) the evaluation of assets; 3) the level of the premiums of long-term policies; and 4) reinsurance (Pentikäinen, Reference Pentikäinen1967). We also see similar statements when we examine only insurance pricing literature: “Premium covers only the cost of risk transfer…In practice capital costs are split between on-balance sheet costs, estimated via cost-of-capital, and the cost of reinsurance” (Mildenhall and Major, Reference Mildenhall and Major2022).
This then leads to the fundamental question of the present study: should extortion payments be included in cyber insurance policies? Model risk—risk of potential loss an institution may incur due to decisions based on inadequacies in internal models (e.g. financial risk measurement and valuation models)—is an acknowledged and long-standing issue in the field of financial risk modelling (e.g., Sibbertsen et al., Reference Sibbertsen, Stahl and Luedtke2008; Klüppelberg et al., Reference Klüppelberg, Straub and Welpe2014; Blanchet and Murthy, Reference Blanchet and Murthy2019). As simplified representations of complex phenomena, models are inherently prone to inaccuracies. Understanding the contexts and mechanisms of model failure is crucial for effective decision-making and policy design. For cyber insurance providers, determining if extortion payments are or are not i.i.d., the latter violating the classical model of ruin theory, would reduce model risk in certain solvency determination circumstances. Thus, our null hypothesis for this study is that extortion payments made to threat actors are independent and identically distributed within and across ransomware strains and variants.
2.2. Dynamics of ransomware attacks and ransom demands
As discussed, traditional model risk calculations and solvency determinations assume a certain degree of environmental stability and predictability—conditions that do not hold for cybersecurity threats. The adversarial nature of such threats creates a feedback loop of escalation that undermines static risk models. As a result, cyber insurers face significant challenges in accurately assessing pricing coverage and capital reserve requirements in a market where risk conditions can shift rapidly and unpredictably. The predator–prey paradigm offers a compelling framework for conceptualizing the dynamic interplay between threat actors and defensive systems and has been investigated in prior research (e.g., Ford et al., Reference Ford, Bush and Bulatov2006; Furnell, Reference Furnell2008; Kumar et al., Reference Kumar, Mishra and Panda2016; Ud Din et al., Reference Ud Din, Masood, Samar, Majeed and Raja2017; Ding et al., Reference Ding, Zhang and Chen2019; Axon et al., Reference Axon, Erola, Agrafiotis, Uuganbayar, Goldsmith and Creese2023; Gorman et al., Reference Gorman, Kulkarni and Schintler2004).
Analogous to biological ecosystems, where evolutionary pressures drive reciprocal adaptations between predators and their prey, the cybersecurity domain is characterized by continuous cycles of offensive innovation and defensive response (Nye, Reference Nye2010; Whyte, Reference Whyte2015; Axon et al., Reference Axon, Erola, Agrafiotis, Uuganbayar, Goldsmith and Creese2023). Threat actors, equipped with increasingly sophisticated tools and techniques, exploit vulnerabilities across digital infrastructures, prompting defensive countermeasures. These countermeasures, while temporarily reducing the efficacy of attacks, often induce adversaries to modify tactics, thereby perpetuating an evolutionary arms race. The resulting co-adaptive cycle mirrors ecological population fluctuations, where increases in attack prevalence are typically met with heightened defensive posture, and lulls in threat activity may inadvertently foster vulnerability through strategic or resource-driven complacency (Nye, Reference Nye2010; Axon et al., Reference Axon, Erola, Agrafiotis, Uuganbayar, Goldsmith and Creese2023).
The Lotka–Volterra model (Lotka, Reference Lotka1925; Volterra, Reference Volterra1926), a foundational construct in mathematical biology for describing predator–prey population dynamics, can be repurposed to simulate the interaction between cyber attackers and defenders within digital ecosystems. In this adapted framework, attackers are analogized to predators whose success is contingent upon the availability and vulnerability of targets (prey), while defenders represent the adaptive capacity of systems to mitigate and withstand these threats (Gorman, Reference Gorman, Kulkarni and Schintler2004; Axon et al., Reference Axon, Erola, Agrafiotis, Uuganbayar, Goldsmith and Creese2023). The coupled differential equations describe how the “prey” population—symbolizing the robustness or density of defended assets—grows in the absence of attack, while the “predator” population—denoting the intensity or frequency of cyber threats—increases through successful exploitation (Gorman et al., Reference Gorman, Kulkarni and Schintler2004):
where H(t) and P(t) represent the magnitude of prey and predator populations, respectively, a is the growth rate of the prey population in the absence of predators, and b is the rate at which a predator population will decrease without a sufficient amount of prey to feed upon. When transposed to cybersecurity threats, this model facilitates the quantitative examination of adversarial dynamics, enabling researchers to predict emergent threat behaviours, evaluate the temporal efficacy of defence strategies, and optimize proactive interventions in complex and evolving threat environments.
Ransomware attacks provide a particularly salient manifestation of predator–prey dynamics, wherein threat actors operate as strategic predators targeting digital ecosystems that often lag in defensive evolution. The cyclical nature of these attacks mirrors ecological interactions: as vulnerabilities proliferate through expanded digital infrastructure, inconsistent patching, stolen credentials, and human error, attackers adapt with increasing sophistication. This predator–prey relationship also has significant implications for the economics of ransomware. As defenders have improved response capabilities (i.e. stronger backup protocols, endpoint protections, and regulatory oversight) and refuse to pay extortion demands, ransomware threat actors have been forced to adapt. This can be seen in threat actors shifting to double and triple extortion techniques, including threats to leak stolen data or contact victims’ customers directly.
The predator–prey framework also invites a deeper examination of the empirical patterns these interactions produce. As discussed, unlike in other insurance areas where past loss experience provides a relatively stable foundation for forecasting future claims, cyber risk is shaped by a volatile and fast-changing threat environment that complicates predicting changes in cyber risk (Marotta et al., Reference Marotta, Martinelli, Nanni, Orlando and Yautsiukhin2017). The challenges posed by this dynamic nature for underwriting are additionally compounded by data availability issues. Existing literature has highlighted the diminishing value of historical data as cyber threats evolve more rapidly than underwriting models can (Eling and Wirfs, Reference Eling and Wirfs2016). Thus, the relevance of historical data for quantifying cyber risk and informing premium-setting is inherently short-lived (Shetty et al., Reference Shetty, McShane, Zhang, Kesan, Kamhoua, Kwiat and Njilla2018). This temporal limitation challenges insurers’ ability to differentiate between levels of cyber maturity across customers and may contribute to pricing strategies based more on market competition than on evidence-based risk segmentation—making reliance on premium pricing for solvency considerations even more hazardous.
2.3. Solvency and modelling considerations
The probability of ruin refers to the likelihood that an insurer’s liabilities will exceed its assets in present value terms at a specified future date, resulting in insolvency. It serves as a key metric for evaluating an insurance company’s risk of insolvency and overall risk exposure. Solvency determinations in classical ruin theory rely on a key assumption of independent and identically distributed (i.i.d.) claims based on a compound Poisson process characterized by a Poisson distribution (Lundberg, Reference Lundberg1903). The i.d.d. assumption has been well-documented in existing literature (e.g., Gerber, Reference Gerber1988; Dickson, Reference Dickson2005; Hult and Lindskog, Reference Hult and Lindskog2011), with more recent work also examining the assumption regarding heavy-tail distributions (Beck et al., Reference Beck, Blath and Scheutzown.d.). For a Poisson distribution, the assumption of i.i.d. is satisfied when random events in different time segments are independent and identically distributed. Additionally, a Poisson distribution also assumes that events in the sample are non-simultaneous, i.e. two or more events do not occur at the same time, and that the probability of an event occurring is not conditional on the number of previous events that have occurred within a small time period (Gill and Bao, Reference Gill and Bao2024).
In classical ruin theory, the Cramér-Lundberg model is the foundational mathematical model for testing solvency against a stochastic claims process (Lundberg, Reference Lundberg1903). Specifically, claims γ with i.i.d. inter-arrival times at time t according to a Poisson process N(t) with claim frequency γ can be defined as
$$ \Upsilon (t)=\sum \limits_{i=1}^{N(t)}{X}_i, $$
where i.i.d. claim amounts Xi occur based on a compound Poisson process with distribution Γ and positive mean μ with finite variance. For an insurance provider that starts with initial capital reserves
$ \phi \ge 0 $
and collects premiums at a constant rate c > 0, total assets
$ \Phi $
at time
$ t $
for
$ t\ge 0 $
can be defined as
$$ \Phi (t)=\phi + ct-\sum \limits_{i=1}^{N(t)}{X}_i. $$
With this model, the primary objective is to evaluate the probability of ultimate ruin
$ \Psi $
, i.e. bankruptcy, denoted as
which occurs when
$ \Phi (t)<0 $
at some
$ t $
(Dufresne and Gerber, Reference Dufresne and Gerber1989; Constantinescu and Thomann, Reference Constantinescu and Thomann2005; Wüthrich and Merz, Reference Wüthrich and Merz2013).
This classical model of ruin theory can also be adapted to account for i.i.d. claim amounts
$ {X}_i $
occurring based on other distributions, including exponential distributions and other variations. The Sparre–Anderson model expands the Cramér–Lundberg model to consider the i.i.d. assumption where claim inter-arrival times follow arbitrary distribution functions (Andersen, Reference Andersen1957). Essentially, instead of assuming exponentially distributed inter-arrival times, the Poisson process allows for inter-arrival times to have any distribution as long as the inter-arrival times are i.i.d. with a finite mean. The Sparre-Anderson model is also defined as
$$ \Phi (t)=\phi + ct-\sum \limits_{i=1}^{N(t)}{X}_i,\mathrm{for}\hskip0.32em t\ge 0, $$
but
$ N{(t)}_{t\ge 0} $
follows the Poisson process with arbitrary distribution functions for i.i.d. claim amounts
$ {X}_i $
(Thorin, Reference Thorin1974).
The i.i.d. requirement is a common assumption underlying statistical analyses (Gill and Bao, Reference Gill and Bao2024). Simply, a dataset of random variables is assumed to be i.i.d. if each random variable has the same probability distribution as the others but is produced without being conditional on any other variable in the dataset. In mathematical terms, the assumption of i.i.d. relies on random variables
$ X $
and
$ Y $
in a sample being independent, defined as
and the sample of
$ n $
random variables,
$ {X}_i,{X}_2,\dots, {X}_n $
, being identically distributed with each variable
$ {X}_i $
having the same mean
$ \mu $
and variance
$ {\sigma}^2 $
, defined as
Random variables that are identically distributed do not necessarily all have the same probability. Furthermore, the i.i.d. assumption is only met when
$ n $
random variables are both independent and identically distributed—violation of either independence or being identically distributed within the sample means that the variables are not i.i.d. (Gill and Bao, Reference Gill and Bao2024).
This study examines whether the identically distributed requirement for i.i.d. is satisfied for extortion payments, including common violations such as the underlying distribution of the sample shifting over time, known as “concept drift” (Webb et al., Reference Webb, Hyde, Cao, Nguyen and Petitjean2016). Random variables must share the same probability distribution, but that distribution can be of any type, such as normal distribution, exponential distribution, Poisson distribution, etc. More precisely, random variables in a sample are considered identically distributed if and only if the cumulative distribution function (CDF)—the probability that a random variable will take on a value equal to or less than said value in a sample—of each random variable is the same as that of other random variables in the sample. Additionally, as the CDF is the integral of the probability density function (PDF), which is the probability that a random variable will take on a value equal to said value in a sample, it follows that identically distributed random variables in a sample would have the same PDF (Gill and Bao, Reference Gill and Bao2024).
In this study, we specifically assess whether the i.i.d. assumption for sub-claim cryptocurrency extortion payments within claim amounts
$ {X}_i $
is satisfied for both Cramér-Lundberg and Sparre–Anderson models.
3. Data and methodology
3.1. Data
The data in this study are extortion payments paid in Bitcoin (BTC) cryptocurrency to ransomware threat actors using one of 145 ransomware strains or variants between 2011 and 2024. Cryptocurrency transaction data have been a source of data for research on ransomware threat actors and victims in previous studies (e.g., Wang et al., Reference Wang, Pang, Chen, Zhao, Huang, Chen and Han2021; Turner et al., Reference Turner, Ikram and Uhlmann2025). The dataset was created by the authors. Extortion payment amounts were collected from the public ledger for the Bitcoin blockchain using BTC addresses, as explained in the next subsection. The structure of our data does not reveal how many victims did not pay extortion amounts, as we are only able to conduct our analyses on observed transactions. Transactions of zero do not show up in the blockchain, but we do discuss payment ratios below.
Regarding the completeness of our aggregated dataset, one limitation is that the data included are a sample of convenience, and we acknowledge that there are many payments not present. Regarding the reliability of our data, each extortion amount was verified by transaction information on the BTC blockchain, and sometimes in the news as well. We excluded any extortion amounts that could not be confirmed by the blockchain public ledger; thus, the data included in this study are traceable and peer-reviewable. Replication and validation of this study’s findings can be conducted by requesting data from authors; independently recreating the dataset based on our Methodology; or using crowdsourced data, e.g. the ransomware project, or data from a blockchain analytics firm such as Chainalysis (Cable, Reference Cable2024).
3.2. Methodology
Bitcoin addresses were gathered by extracting BTC addresses from malware binary samples and ransom notes from confirmed ransomware attacks, following existing accepted practices. For example, one could acquire malware binary samples from publicly available repositories used by security researchers, incident responders, and/or forensic analysts (e.g. VirusShare), as well as publicly available scanning and verification tools for malware (e.g. VirusTotal). Malware binary samples are executable files that contain code for executing the malware on a device or system for an attack, and these samples are often used for reverse engineering for cybersecurity research (Ferguson et al., Reference Ferguson, Kaminsky, Larsen, Miras and Pearce2008).
The process for aggregating BTC addresses involved using a regular expression to extract BTC addresses from malware binary samples and text files associated with ransomware attacks. To protect against false positives, every BTC address was validated with the address validation mechanisms publicly available on the Rosetta Code website (RosettaCode, n.d.). After a BTC address was validated as a legitimate BTC address, soft attribution was performed via fuzzy hashing—a matching technique using a hash function to focus on common patterns or structures and allow for small variations or differences—of the ransom note or the malware itself, e.g. with TLSH (Trendmicro Locality Sensitive Hash). With this technique, each malware binary sample was assigned a ransomware strain or variant by matching the binary malware sample or the ransom note to previous samples from a known ransomware strain or variant. Many of these soft attributions have been validated with other data sources, such as Ransomlook, that classify BTC addresses into ransomware strains and variants (Dulaunoy et al., Reference Dulaunoy, FafnerKeyZee and Harper2024). When this soft attribution was not possible, the BTC address was put into a bucket known as “unaffiliated” until soft attribution may be possible in the future. All of these unaffiliated addresses (
$ N $
= 2022) were excluded from this study.
Once collated into a ransomware strain or variant, extortion payments to the BTC address were traced using the public BTC blockchain. We focused only on the incoming transactions (extortion payments from victims) and not on outgoing transactions (withdrawals by threat actors). Confirmed transactions were converted to US dollar (USD) amounts using the average price of BTC on the day of the transaction, which is a standard practice in any research assessing the value of cryptocurrency or other digital assets over time (Conti et al., Reference Conti, Gangwal and Ruj2018; Cable et al., Reference Cable, Gray and McCoy2024). Additionally, to explore if the demand value of extortion amounts was independent of fluctuating BTC price, we analysed a 3-year moving average of both BTC price and extortion payments to establish a lack of correlation (see Appendix A).
Herein, we conduct descriptive statistics of our data, as well as empirical analyses of specific ransomware strains/variants and incidents. We then determine whether the i.i.d. assumption is valid for extortion payments across and within our data using multiple tests, including a Kruskal–Wallis, Spearman’s Correlation, Cramer von Mises, and Kolmogorov–Smirnov. Violations of either independence or identical distribution for extortion payments violate the i.i.d. assumption, suggesting that ransom demands are generated by the different processes (tools, tactics, and procedures) associated with certain ransomware strains/variants or specific threat actors.
4. Results
4.1. Descriptive statistics
Table 1 displays summary statistics for our ransom paid dataset. The data include 121,762 extortion payments paid to threat actors, totalling more than a billion dollars (USD) in extortion payments ($1,275,242,024.72 USD). The largest extortion payment included in the study is $24,837,732.88. The median extortion payment across the entire timespan of our dataset was $251.90 (all USD herein), and more recently, over the last three years, the median extortion payment was $396.36. Overall, our data indicate that from 2011 through 2024, 92% of extortion payments were below the average of $10,473.23, and over the last three years, 93% of extortion payments were below the average of $186,277.40. Beyond these aggregate statistics, we also investigated the mean-to-median ratio of extortion payment amounts across all ransomware strains and variants in the dataset (see Appendix B). The mean-to-median ratio is an indicator of distributional skewness, as discussed in Florêncio and Herley (Reference Florêncio and Herley2011), which suggests a range of variability in ransom demand behaviours that are associated with ransomware strain or variant used by a threat actor.
Table 1. Summary statistics

As discussed, the frequency and severity of cybersecurity incidents, especially those involving large-scale breaches or high-impact attacks, often exhibit heavy-tailed distributions, deviating substantially from the expectations of normal or Poisson-like event models. These fat-tailed characteristics suggest that extreme events, though rare, occur with disproportionate impact and frequency. Figure 1 provides an overview of the frequency and severity of extortion payments in our dataset. In reference to the two tiers of cyber risk, while extortion payments stem from tier 1 risks, these payments also represent tier 2 risks, with the number of payments in a certain time frame stochastically fluctuating.

Figure 1. Characteristics of extortion payments by ransomware victims over time in our dataset. (a) Quarterly plot of the number of extortion payments by ransomware victims from 2011 to 2024. (b) Quarterly plot of the summed USD amount for extortion payments by ransomware victims from 2011 to 2024. Note particularly that the decrease in payments in (a) does not lead to reductions in (b): bigger payments are happening less often for equivalent profits.
Shown in Figure 1a, there is high variance in the frequency of extortion payments, with some quarters having thousands of extortion payments by victims, accumulating in a global level impact. Further complicating the issue is the varying adjusted USD amount for these payments as well, as seen in Figure 1b. To additionally visualize the differences between ransomware strain variants, Figure 2 provides a graphical representation of the log variances in extortion payments across all of the ransomware strains and variants in our data. Along with the extortion amount, the frequency of extortion payment also varies between ransomware strains/variants, as well as over time (see Appendix C).

Figure 2. Variance of extortion payments according to ransomware strain or variant. Boxplot of log variance of extortion payments for each of the ransomware strains and variants in the data, including lower and upper quartiles and outliers. Variance was logged for numerical stability and to maintain positive values.
4.2. Anecdotal case examples
With known differences in extortion payments between ransomware strains/variants, we provide case examples of unique distributions of extortion payments for three different ransomware strains or variants to qualitatively and empirically investigate the i.i.d. assumption. Some ransomware strains/variants used a fixed-price business model where a ransom demand amount was fixed regardless of the victim, and others used a negotiated-price business model where victims negotiated with ransomware actors on the amount of the extortion payment. Fixed-price extortion payments are uniformly distributed, as expected for such a business model, and the negotiated extortion payments are characterized by a variety of distributions from normal distribution to power-law distribution. Our first case example is the WannaCry ransomware attack.
4.2.1. WannaCry ransomware
The 2017 WannaCry ransomware attack was wide-ranging, affecting more than 300,000 computers across 150 countries (see Figure 3) (Akbanov et al., Reference Akbanov, Vassilakis and Logothetis2019; BBC, 2017). The initial ransom demand for victims was $300 USD to be paid in Bitcoin (Figure 3a), but doubled to $600 USD during the ongoing campaign (Gibbs, Reference Gibbs2017). While the threat actor behind the attack was using a fixed-price model, the price doubling makes clear that the distribution of extortion amounts can change over time, even within one ransomware attack, highlighting the ‘concept-drift’ breaking element of the i.i.d. assumption.

Figure 3. Ransom demand for WannaCry 2.0 ransomware and distributions of extortion payments. (a) Example WannaCry extortion demand note requesting a USD extortion payment be transferred in BTC. (b) WannaCry extortion payment frequency on a weekly basis, with extortion payments continuing after the initial primary campaign. (c) Kernel density estimation (KDE) of the bimodal distribution of fixed-price WannaCry extortion payments.
The initial attack on May 12, 2017, was massively curtailed within hours by a researcher who discovered a kill-switch, and though there were attempts to update the mechanisms of attack to continue spreading, the initial campaign was ultimately ended in about a week’s time. This is shown in Figure 3b with the extreme spike in extortion payment frequency in a short time frame in 2017 followed by a much lower frequency, though still continuing into 2023. This can be attributed to numerous WannaCry ransomware variants and old software. WannaCry is a ransomware family, with the 2017 attack involving the WannaCry 2.0 strain out of three strains currently in existence (Yang, Reference Yang2017), and those strains have now led to thousands of variants that continue to be used by threat actors. Additionally, WannaCry ransomware relies upon the EternalBlue exploit for older Microsoft systems that are still vulnerable if they have not been effectively patched (Goodin, Reference Goodin2017).
While there has been debate as to attribution, the United States and United Kingdom formally announced that North Korea’s Reconnaissance General Bureau was responsible for the ransomware attack (DOJ, 2018). North Korea may have benefited geopolitically by targeting states deemed adversarial to their interests, but the attacks were predominantly financially motivated, considering the increase in ransom demand amount during the attack, and knowledge from the intelligence community that North Korea uses such attacks to fund military objectives (CISA, 2023). The extortion amounts in USD had a bimodal distribution (see Figure 3c for estimation of PDF with kernel smoothing), and while other ransomware strains/variants in the dataset also have bimodal distributions, each is distinctly different. This speaks to the twinned issues of ransomware as an economic national security threat, which is distinct from national security threats that use ransomware as a cover story; it benefits everyone to sort the financially motivated crime from the espionage.
4.2.2. Deadbolt ransomware
We next examined Deadbolt ransomware attacks. Compared to WannaCry, one may expect that more recent ransomware attacks would involve negotiations for ransom demands due to technological leverage, but there are still examples, such as Deadbolt ransomware attacks, of a fixed-price approach to reject this notion (Figure 4). Deadbolt ransomware attacks operate with a flexible flat fee demand of between 0.03 BTC and 0.05 BTC (see Figure 4a) without any mechanism for negotiation if desired (no email address, chat handle, or domain listed in the ransom demand notes) (Chainalysis, 2023; Ellzey, Reference Ellzey2022; Muncaster, Reference Muncaster2022). There is an additional pattern of low-high extortion payments due to how the decryption key is sent to victims who pay ransoms. Once a victim pays the ransom, the decryption key is sent automatically via the blockchain as a low-value Bitcoin transaction to the ransom address with the decryption key written in the OP_RETURN field of the transaction (Chainalysis, 2023).

Figure 4. Value of Deadbolt and Ryuk extortion payments over time. (a) Value of Deadbolt extortion payments from January 2022 through September 2022. (b) Value of Ryuk extortion payments from August 2018 through January 2022.
It is also important to note that other variants of Deadbolt don’t necessarily follow the same behaviour. Deadbolt ransomware specifically targets network-attached storage (NAS) devices, and while there is a fixed-price model for end customers, another variant targets NAS vendors for much higher amounts that vary, e.g. $192,000 USD to $959,000 USD (Muncaster, Reference Muncaster2022). Though attribution remains unclear, intelligence agencies have advised that Deadbolt ransomware is also routinely used by North Korean actors, state-sponsored or otherwise. Compared to the WannaCry 2.0’s extortion payment bimodal distribution (PDF in Figure 5a, CDF in Figure 5d), Deadbolt has a similar bimodal distribution for extortion payments but is still distinctly different based on extortion payment amount and frequency over time (PDF in Figure 5b, CDF in Figure 5e). One possible explanation for the difference is that Deadbolt has also been observed to be used by other actors outside of North Korea (CISA, 2023).

Figure 5. (top) Probability density function (PDF) graphs for (a) WannaCry, (b) Deadbolt, and (c) Ryuk extortion payments demonstrating the probability density at specific amounts across the range of payment values, respectively. (bottom) Cumulative distribution function (CDF) graphs for (d) WannaCry, (e) Deadbolt, and (f) Ryuk extortion payments demonstrating the probability that a random extortion payment is less than or equal to an extortion amount across the range of payment values, respectively.
4.2.3. Ryuk ransomware
Lastly, we examined Ryuk ransomware for our third case example. Ryuk ransomware first emerged in February 2018, and by 2021, threat actors had made more than $150 million USD (Cimpanu, Reference Cimpanu2021). The ransomware, which is actually a variant of Hermes ransomware, has been attributed to the Wizard Spider group of Russian affiliation (Micro, Reference Trend2024). Ryuk ransomware attacks followed a negotiated-price business model and specifically targeted high-profile victims for maximum profit, including healthcare facilities, schools, government facilities, etc. (CISA, 2020) (Micro, Reference Trend2024). In contrast to the fixed-price models of the WannaCry and Deadbolt strains/variants in our data, Ryuk’s negotiated-price model is observably different in frequency of and variance in extortion payment amounts, as shown in Figure 4b, as well as in the PDF and CDF, as shown in Figure 5c and Figure 5f, respectively (Gray et al., Reference Gray, Cable, Brown, Cuiujuclu and McCoy2023).
Alongside the negotiated-price business model, the Wizard Spider group also had distinct patterns in their methods of attack, using specific malware like Trickbot and Emotet to compromise systems and other commercial products to gain access to ultimately execute Ryuk (CISA, 2020) (Micro, Reference Trend2024). Additionally, the group is also associated with other notable ransomware strains and variants, including Conti ransomware. We analysed the PDF and CDF of Conti, as well as LockBit (Russian-affiliated with double extortion business model) and Cuba (Russian-affiliated with negotiated-price model for high-profile targets) (see Appendix D). While comparing frequencies, ransomware business models, PDFs, and CDFs is not quantitatively causal for rejecting our null hypothesis, it does anecdotally indicate that there is high variability in ransomware attacks and ransom payments, which does not support the use of monolithic insurance coverage from cyber insurance providers.
4.3. Testing for i.i.d. assumptions across ransomware strains/variants
We first conducted the Kruskal–Wallis (KW) test (also known as the one-way ANOVA on ranks test) to determine if multiple samples are produced from the same distribution for non-parametric data (Kruskal and Wallis, Reference Kruskal and Wallis1952). The null hypothesis for the KW test for our study is that a randomly observed extortion amount from any ransomware strain/variant is neither more nor less likely to be larger in value than a randomly observed extortion amount from any other ransomware strain/variant in our dataset. Our analysis of variance of extortion payment against ransomware strain/variant resulted in a Kruskal H-value of 21732.919363219826 (p-value = 0.00000) (DATAtab, n.d.).
Since the KW test is insensitive to outliers, the H-value calculation generally relies only on extortion amounts within each ransomware strain or variant having identically distributed values with finite means and variances (Kruskal and Wallis, Reference Kruskal and Wallis1952). Under the strict assumption of identical distribution, our high H-value and significant p-value would suggest that we could reject the KW null and assume that at least one randomly observed extortion amount from any ransomware strain/variant is more likely to be larger in value than a randomly observed extortion amount from any other ransomware strain/variant. However, if we accept that the distributions in our dataset are not each identically distributed, as we have shown with our case examples, our interpretation of the result is that though we may reject the null, there is evidence that least one ransomware strain/variant is stochastically larger than at least one other ransomware strain/variant at the level of
$ p\le 0.05 $
significance.
We next conducted an analysis of Spearman’s rank correlation coefficient, a non-parametric measure of statistical dependence between the rankings of two variables, to check for correlations between extortion payments and time (Spearman, Reference Spearman1904). The coefficient assesses the strength and direction of the relationship between two variables based on a monotonic function, which is less restrictive than that of Pearson correlation coefficient based on a linear relationship (Kowalski, Reference Kowalski1972). The calculation uses the assumption that the two variables represent paired observations, and these observations are independent. For our analysis, we calculated Spearman’s coefficient to determine if time has an impact on extortion amounts, and our results of a low Spearman’s coefficient of 0.311502709075461 (p-value = 0.00000) indicates statistically dissimilar ranks between extortion payment amounts and time. This result suggests that the assumption of i.i.d. may not hold, as extortion payment amounts appear to be dependent on the ransomware strain/variant. While they may be independent of time, their dependence on strain/variant introduces a potential source of non-independence in the data.
To further investigate the relationship between extortion payment amounts in our data, we conducted various tests to ascertain distance metrics—degree of similarity or difference between two data points—within our dataset. The Kolmogorov–Smirnov test (KS test) takes two forms: 1) a one-sample KS test that compares whether an empirical sample comes from a theoretical probability distribution, and 2) a two-sample KS test that compares whether two independent empirical samples come from the same theoretical probability distribution (Massey, Reference Massey1951). Both tests involve comparing the CDFs to determine whether samples come from the same underlying distribution. We used the two-sample KS test to evaluate extortion payments of different amounts matched in their timestamp distributions (Figure 6). Additionally, the use of the two-sample KS test also mirrors a widely accepted approach for pricing distribution fitting (Parodi, Reference Parodi2023), adding relevance to the discussion of pricing and solvency.

Figure 6. Pairwise Kolmogorov–Smirnov test (KS test) heat maps for distance with p-values. (a) Heat map of pairwise KS test between all ransomware strains/variants in our dataset. (b) Heat map of corresponding p-values for the results of the KS test in (a).
With inspiration from Liao et al. (Reference Liao, Zhao, Doupé and Ahn2016), we used a time-series analysis of the data, expanding across multiple ransomware strains/variants, to examine how likely it is that extortion payments for two ransomware strains/variants are drawn from the same probability distribution. In particular, we compared different strains/variants across the entire dataset, as well as compared variants of the same strain against each other (Figure 6a). We additionally assessed the p-value of these two-sample KS tests (Figure 6b) for assessment of the significance of the results across all ransomware strains/variants. While the KS test is a regularly used goodness-of-fit test for non-parametric data, it does have some weaknesses. In particular, since the KS test compares the underlying CDFs for heavy-tailed distributions, the tails of the CDFs are more precise than the middle of the function, leading to the overall KS test losing sensitivity in the tails. Thus, we decided to examine distribution distance with higher sensitivity metrics than the KS test, including Wasserstein distance and the Cramer von Mises metric (see Supplementary Appendix E).
Wasserstein distance determines the minimum cost for transforming one probability distribution into another with considerations for spatial arrangements of the data and overall shapes of the distributions (Patriarca et al., Reference Patriarca, Heinsalu, Marzola, Chakraborti and Kaski2016). This method is more sensitive to shifts in the central data than outliers and more robust to differences in the tails of distributions. The Cramér von Mises metric assesses the goodness of fit of a CDF compared to the empirical distribution function, i.e. the distribution function of the empirical measure of the sample, by evaluating the sum of squares of differences (Cramér, Reference Cramér1928). The metric modifies the parameters of a normal distribution to account for positive and negative deviations. The results from the Wasserstein distance Supplementary Figure S4a) and the Cramer von Mises metric (Supplementary Figure S3a) further support that ransomware payments do not have similar probability distributions across families.
We performed additional testing for distribution fit and autocorrelation within our dataset, though it should be noted that further investigation along these lines of inquiry is beyond the scope of the present study. The Anderson–Darling test (AD test) is a statistical method for evaluating the goodness-of-fit between empirical data and a theoretical distribution, including those with heavy-tailed characteristics (Anderson and Darling, Reference Anderson and Darling1952). Unlike the KS test, the AD test applies greater weight to discrepancies in the distribution’s tails, thereby enhancing its sensitivity to extreme values often encountered in finance and risk modelling. This property makes it particularly well-suited for assessing distributions where tail behaviour—often indicative of rare but high-impact events—is of analytical importance. We conducted the AD test using well-established distributions (normal, exponential, logistic, Weibull min, and Gumbel (Extreme Value Type I) distributions) at different levels of significance. Our results indicate that ransom payments across ransomware strains/variants in the dataset do not fit these chosen distributions, though a limited number of observations within strain/variant showed alignment with these distributions (see Appendix E).
Lastly, we performed two tests for autocorrelation between observations, given the time span of our dataset and considerations for temporal activity in claim payments for solvency. Again, it should be noted that further time-series analyses of the data are beyond the scope of the present study. Understanding the non-normal distribution of our data, we conducted the Durbin-Watson test (DW test) (Durbin and Watson, Reference Durbin and Watson1971) and the Ljung–Box test (LB test) (Ljung and Box, Reference Ljung and Box1978) for serial correlation. Neither test requires normality, though they do assume independence under their respective null hypotheses for no autocorrelation. Our results indicated that ransom payments within ransomware strains/variants in the dataset have positive serial correlation, though a few ransomware strains/variants demonstrated no autocorrelation (see Supplementary Appendix E). Taken with the other results from our analyses, this suggests that even within ransomware strain/variant groupings, the assumption of independence is predominantly violated across strains/variants in our dataset.
5. Discussion
The findings of this study raise critical concerns about the applicability of classical solvency models—particularly those grounded in the assumption of independent and identically distributed loss events i.e. Classic Ruin models—in the context of ransomware incidents. Empirical analyses of extortion payment data reveal significant departures from i.i.d. assumptions, calling into question the reliability of classical ruin theory when used to assess financial resilience against solvency in the face of ransomware extortion. These violations are not merely statistical artifacts; they reflect deeper structural features of the ransomware threat landscape, including temporal clustering of attacks, heavy-tailed payment distributions, and potential for coordinated threat actor behaviour. As such, our results suggest that existing frameworks may systematically underestimate the risk of insolvency for cyber insurers providing coverage for extortion payments, necessitating new models and updated regulatory standards.
Overall, the results of the present study suggest an underlying tension between observed ransomware extortion payment data and the core assumptions underpinning classical ruin theory. The distribution of ransom payments violates key conditions necessary for solvency assessment within this framework—most notably, the assumptions of independence and identical distribution in both severity and inter-arrival times, as well as the requirement of a finite mean. The violation of ruin theory conditions affects not only cyber insurance firms offering extortion payment coverage based on these specific model assumptions, but also self-insured organizations, third-party negotiators, and risk modellers operating in adjacent spaces. Many stakeholders may assume that ransom payments and associated losses are closely related, even interchangeable from a modelling perspective, but such assumptions have not been empirically validated and should not be assumed. If the distribution of ransomware payments lacks a finite mean, as heavy-tailed behaviour would suggest, then the perceived insurability of these events may be an artefact of insufficient data rather than a reflection of actual risk structure. As ransomware threats continue to evolve, any risk assessment framework that overlooks these distributional properties is likely to underestimate the true exposure facing firms and insurers alike.
We do not argue that ransomware extortion payments are difficult to insure solely due to their statistical variance or volatility. Rather, our findings suggest that insuring ransoms as a sub-coverage introduces significant solvency concerns, rooted not in randomness but in the underlying heterogeneity of causal processes that generate ransomware claims. Modern ransomware operations employ a range of coercive tactics, including data exfiltration, system encryption, and operational disruption, often within a single campaign. These tactics represent distinct mechanisms of leverage, reflecting varied threat actor objectives and capabilities. To assume that ransom demands emerge from an i.i.d. process is to overlook the strategic heterogeneity among threat actors and the resulting complexity in how extortion is executed and monetized.
Once we acknowledge that extortion amounts are neither identically distributed nor reliably independent, a deeper causal inquiry becomes necessary. This variation is not incidental; it is shaped by both predator and prey dynamics. Specifically, it is important to consider whether this variability is primarily driven by characteristics of the threat actors (i.e. predator effect) or by the attributes and behaviours of the targeted organizations (i.e. prey effect). On the threat actor side, technological factors such as the preferred exploits of initial access brokers, the quality of lateral movement tools, or the ability to scale across networks can all influence the magnitude of potential losses. Additionally, targeting preferences, whether based on geography, sector, or organization size, are not static, but evolve in response to shifting incentives and observed payer behaviour. For instance, there is evidence that threat actors adjust their targeting to focus on jurisdictions or industries with a higher historical likelihood of ransom payment. These behavioural adaptations reflect learning processes that further disrupt any assumption of statistical independence or distributional uniformity.
Victim-side characteristics introduce further heterogeneity in loss outcomes. Factors such as industry sector, revenue size, and negotiation practices may all contribute to the heterogeneity observed in extortion payment outcomes. Additionally, organizational factors such as sectoral regulation, technological maturity, incident response capabilities, and the economic cost of downtime contribute to the severity and composition of ransomware-related losses. For example, while privacy breaches may be of paramount concern in healthcare or finance, downtime may be far more consequential for energy utilities or logistics firms. Losses are thus a complex mixture of paid and unpaid ransoms, operational and reputational impacts, and idiosyncratic, systematic, and systemic risks—echoing frameworks described by Awiszus et al. (Reference Awiszus, Knispel, Penner, Svindland, Voß and Weber2023).
The complexity deepens with the involvement of state-sponsored or state-sanctioned ransomware operations, where the motivations of threat actors may shift away from purely financial gain toward geopolitical disruption. This possibility introduces a significant source of model risk, particularly for insurers and regulators relying on attribution-based exclusions to delineate the boundaries of insurable events. As several cyber war exclusions hinge explicitly on attributing intent or affiliation, models premised on rational economic actors may fail under scenarios where the primary objective is systemic damage rather than ransom extraction. In such contexts, the underlying assumptions of risk modelling and not just the statistical distributions require fundamental re-examination.
Attribution is a difficult task, and association with a nation-state actor can be murky, though there are some organized groups who are well known to be proxies for adversarial nation states (Egloff, Reference Egloff2017; Egloff and Egloff, Reference Egloff and Egloff2022; Nershi and Grossman, Reference Nershi and Grossmann.d.). Studying victimization statistics has the potential to identify what is ‘targeted’ and what is “opportunistic.” For instance, the WannaCry ransomware attacks in 2017 are widely believed by multiple countries to be attributable to North Korea’s Reconnaissance General Bureau (Horsley, Reference Horsley2018; Bendiek and Schulze, Reference Bendiek and Schulze2021; Sumortier et al., Reference Dumortier, Papakonstantinou and De Hert2020). North Korea, one of the major global threat countries, is also known to have a modus operandi of using such types of cybercrimes for monetary gain, specifically in cryptocurrency. In WannaCry, the ransomware attacks were conducted for financial gain, though of significance was the damage and disruption to critical sectors and services globally. However, it is important to note that these attacks were opportunistic, taking advantage of technological vulnerability, rather than specifically targeting certain victims.
Ransomware threat actors emanating from Russia behave differently, wherein threat actors are not state-sponsored but behave as indirect proxies for state interests. These threat actors follow a similar tactic of primarily pursuing financial gain alongside causing significant levels of damage and disruption. However, Russian-based actors tend to be more politically strategic in the choice of target and what type of disruption and damage they cause, which in turn increases their chances of safe harbour and Russia’s indifference towards extradition efforts. Many of these ransomware incidents contribute to differences in ransom demand distributions and victim payment distributions. Furthermore, many ransomware threat actors operate as Ransomware-as-a-Service (RaaS), providing their ransomware variant to be used by other threat actors for a cut of the profits in return. From a cyber insurance perspective, insurers must be cautious as to how categorization and potential attribution of ransomware attacks are used for coverage determinations. We are thus careful to call out potential, directly state-sponsored threat actors or campaigns in this study, as well as note that attribution to a ransomware strain herein may represent actions by the respective organized group and actions by other threat actors using said strain (or respective variants).
This discussion of attribution is not dialectical—there are real implications for the payout of insurance claims (attribution can impact policy coverage when involving nation state threat actors. For example, see Lloyd’s Market Associations cyber war clauses and requirements regarding state-backed cyber attack exclusions; (Lloyd’s Market Association, 2025)). Millions of dollars (USD) can and will be denied from insurance payouts if a claim is found to be generated by a nation-state threat actor. Moreover, a threat actor seeking to maximize damage rather than achieve financial gain also supports our findings that ransom payments are not i.i.d., as these ransom payments are more side effects than natural variance from an underlying generating function. Ransomware threat actors operate with distinct monetization strategies and business models, which influence not only how access is priced but also how victims are selected and engaged. These strategic differences suggest that ransomware attacks may reflect both predator effects (e.g. the chosen strain or operational variant) and prey effects, including sector, geography, and the target’s willingness to pay extortion demands. For example, a ransomware variant tailored to German-speaking users or designed to exploit infrastructure in US-based firms will inevitably shape the demographic and economic profile of affected victims, and these prey-driven dynamics may result in observed ransom amounts that are more reflective of victim attributes than attacker intent (Rose-Ackerman and Palifka, Reference Rose-Ackerman and Palifka2016; Ridinger, Reference Ridinger2018; Cimpanu, Reference Cimpanu2019).
Supporting this view, a pioneering study by Cybersecurity Centre Belgium found that the frequency of ransomware attacks correlates more strongly with national GDP than with population size, suggesting that financial capacity and not sheer exposure is a key determinant of targeting (Centre for Cybersecurity Belgium, 2024). This introduces a potentially endogenous relationship between victim demographics and ransom payment valuations, complicating assumptions of randomness or uniformity often used in risk models. Understanding this interaction between attacker behaviour and victim characteristics is crucial for refining both threat intelligence and actuarial assumptions in ransomware risk assessment. Additionally, this cultural variability applies both to perpetrators and victims in different but important ways (Banuri and Eckel, Reference Banuri and Eckel2012). These dynamics would affect victims’ Willingness To Pay (WTP) extortion demands, as well as the amount that victims can or do ultimately pay (Everett, Reference Everett2016; Hernandez-Castro et al., Reference Hernandez-Castro, Cartwright and Cartwright2020). Our study has only gently explored this as a causal factor in why extortion amounts are not identically distributed and future research will hopefully expand on these themes.
While our findings challenge the assumption that extortion payments are identically distributed and thus readily insurable, we do not extend this argument to ransomware losses in general. The distinction between extortion payments and losses is both practical and conceptual: ransom demands may correlate with observed losses, but correlation alone is insufficient to establish equivalence or causality. It is entirely possible for a ransom-related loss to correlate with both threat actor behaviour and victim characteristics, yet emerge from distinct underlying distributions. In fact, we suggest that ransomware losses may align more closely with organizational factors such as revenue, sector, and geography than with the statistical properties of ransom demands and extortion payments. The misconception of equivalency persists in both public discourse and industry practice and warrants careful correction in both actuarial modelling and policy development.
Ultimately, the continued practice of insuring ransomware extortion payments without addressing the underlying statistical irregularities raises important questions about the rationale guiding both insurer behaviour and broader market practices. Regulatory bodies tasked with overseeing financial risk in the cyber domain should treat these findings as a signal for urgent re-evaluation.
5.1. Implications for regulatory standards and policy
Ransomware presents a complex and escalating risk management challenge for both governments and the global economy. In the absence of reliable data on the frequency and cost of attacks, organizations struggle to justify sustained investment in cybersecurity defence measures, perpetuating a cycle of vulnerability and loss. This uncertainty contributes to a growing economic burden, with ransomware-related damages now estimated to exceed $ 1 billion (USD) annually (Chainalysis, 2024). From a regulatory and policy standpoint, there are two critical paths forward: 1) the promotion of open, transparent risk models that enable organizations to quantify cyber risk and allocate defensive resources accordingly; and 2) coordinated efforts to de-monetize ransomware through legislative, regulatory, and insurance-based interventions. However, as our findings demonstrate, the statistical properties of extortion payments, specifically their non-i.i.d. nature, potentially undermine the reliability of insurance as a stand-alone solution. Any regulatory strategy aimed at managing ransomware risk must therefore account for the limitations of classic Ruin Theory in modelling and mitigating this uniquely dynamic threat.
Ransomware operates as a globally correlated risk, and the continued payment of ransoms may be artificially inflating cyber insurance premiums by introducing a direct path to insolvency. The inability to achieve effective risk pooling in the face of systemic cyber threats has been well documented in industry analyses (Lee, Reference Lee2023), and it offers empirical validation for predictions made by Böhme and Kataria (Böhme and Kataria, Reference Böhme and Kataria2006). Our findings suggest that when insurers repeatedly cover ransom payments, they expose themselves not just to increased claim frequency but to a compounding solvency threat. In effect, insurers that routinely finance extortion payments take on an accumulating tail risk, one that brings them incrementally closer to ruin. It is essential to emphasize that ransomware has already proven capable of bankrupting organizations across sectors (Bilton, Reference Bilton2025)—cyber insurance providers are not exempt. These firms face at least two existential exposures, including the threat of becoming targets themselves and the financial consequences of underwriting a class of risk that defies underlying model assumptions. Without intervention through regulatory oversight, revised underwriting practices, or market-wide norms, this dynamic may undermine the long-term viability of ransomware sub-coverages altogether.
The distinction between pricing adequacy and solvency assurance is central to the regulatory oversight of insurance markets, particularly when confronting dynamic risks such as ransomware. As defined by the Actuarial Standards Board of America, capital represents “the funds intended to assure payment of obligations from insurance contracts, over and above those funds backing the liabilities” (Actuarial Standards Board, 1997). This highlights a crucial point: while accurate pricing is necessary, it is not sufficient to guarantee solvency, particularly in the presence of non-i.i.d. processes that undermine certain models. Ruin Theory, rather than pricing alone, provides a formal framework for assessing long-term solvency. From this perspective, the insurability of ransom payments depends on whether their statistical properties satisfy the assumptions underpinning that theory, specifically independence and identical distribution. In the absence of such conditions, it is incumbent upon regulators to reevaluate solvency requirements for insurers engaged in ransomware payment coverage, potentially mandating more conservative capital buffers.
This regulatory imperative is echoed in established guidance on general insurance pricing, which states that “regulation may prescribe a minimum amount of capital that an insurer has to hold for the risks it underwrites” (Parodi, Reference Parodi2023). Moreover, the guidance reinforces that “specifically for pricing, the firm will need to be able to demonstrate that all pricing decisions and assumptions are documented, monitored, and linked to the firm’s capital model” (Parodi, Reference Parodi2023). In the context of pricing and ransomware coverage, such documentation should explicitly account for the tail-risk properties of ransom payments and their divergence from standard distributional assumptions. Without this linkage, pricing decisions may appear actuarially sound while still exposing firms to unacceptable solvency risk.
Addressing these challenges will require more than technical adjustments to existing models; it demands a reconsideration of the data and assumptions that underpin cyber risk assessment. Heavy-tailed phenomena, by their nature, require substantially more data to produce reliable statistical inferences. However, the opacity of ransomware incidents, under-reporting, and inconsistent disclosure practices continue to limit the availability of comprehensive datasets. Regulators, insurers, and cybersecurity firms must collaborate to develop data-sharing mechanisms that respect privacy and legal constraints while enabling rigorous empirical analysis. Without such efforts, the modelling of ransomware risk will remain constrained by structural blind spots, leading to potentially flawed pricing, inadequate reserves, and misplaced confidence in the insurability of extreme events. In this context, regulatory guidance could play a catalytic role in setting minimum standards for disclosure, particularly where systemic cyber risk threatens broader financial stability.
Another critical regulatory implication of our findings is the need for enhanced transparency and oversight of solvency models used by insurers offering ransomware extortion payment coverage. Given the demonstrated violation of i.i.d. assumptions in ransom payment distributions, regulators should require firms to rigorously document the assumptions, limitations, and data sources underlying their capital models. This documentation would not only support more accurate solvency assessments but also create the conditions for improved regulatory monitoring and cross-firm comparison. The benefits of such oversight include greater confidence in insurers’ financial resilience, reduced systemic exposure, and the ability to identify market-level modelling blind spots. However, there are potential drawbacks to this approach. Increased documentation requirements may impose administrative burdens on firms, especially smaller insurers, or incentivize overly conservative capital holdings that reduce market competitiveness. There is also a risk that regulatory standardization could stifle innovation in modelling methodologies or create incentives for firms to align with regulatory expectations rather than empirical realities. Nevertheless, on balance, transparent solvency model governance remains a necessary step toward ensuring the insurability of cyber risks does not rest on unexamined or outdated assumptions.
Lastly, and perhaps most controversially, is the implementation of a ban on payment of ransoms to ransomware threat actors. Policy discussions in both the United States (Logue and Shniderman, Reference Logue and Shniderman2022) and the United Kingdom (Office, Reference Home2025) have considered the prospect of banning ransom payments to cybercriminals—a move that reshapes the role of insurers in managing ransomware risk. In particular, the UK government recently announced its intention to ban ransomware payments for public sector bodies and critical infrastructure operators (Office, Reference Home2025). While outright prohibition remains politically and logistically complex, our findings suggest that a more targeted intervention may offer a pragmatic alternative. Specifically, regulators might consider restricting cyber insurers from directly covering ransom payments, thereby shifting the burden of that decision, and its associated costs, to the affected organization’s leadership.
Under such a framework, insurers would continue to support recovery through coverage for forensic investigation, business interruption, and remediation, but ransom payment itself would become a board-level decision, not an insured event. Companies would still be obligated to comply with sanctions regimes, such as those enforced by OFAC in the United States or OFSI in the United Kingdom, and this compromise preserves organizational autonomy while reducing the perverse incentives associated with risk externalization. Importantly, it also reinforces the need for insurers offering ransomware extortion payment coverage to demonstrate their own solvency, particularly if ransom payments remain part of their offerings. As ransomware continues to evolve, regulatory innovation must strike a balance between reducing harm, discouraging payment, and preserving institutional resilience, but the fundamental message is clear: insuring ransom payments without recognizing their structural risk may not just be unsustainable, it may be part of the problem.
6. Limitations and future work
As with any empirical analysis of cyber risk, this study is subject to several limitations. First, while our dataset is reflective of real-world ransomware payment events, it under-represents the full spectrum of incidents, particularly those that go unreported due to reputational, legal, or regulatory concerns. As with many studies in this space, our dataset captures only a subset of the overall landscape, and we are aware that larger institutions may have access to more comprehensive or proprietary data. Additionally, the opacity of ransomware payments continues to constrain the field’s ability to make definitive claims about distributional properties across all sectors and geographies. That said, we have documented our approach and provided sufficient methodological detail to allow others to reproduce and extend this analysis on alternative datasets. We view this openness as a necessary step toward improving the empirical foundations of ransomware research.
Second, while both frequency and severity are fundamental dimensions of cyber risk, this study has focused exclusively on the severity of ransom payments. A rigorous treatment of frequency dynamics would require a separate analytical framework, alongside additional assumptions to handle common data quality issues, such as right-censoring, under-reporting, and time lags in disclosure. These complexities place such an analysis beyond the scope of the present study but offer fertile ground for future research into the temporal evolution and recurrence risk of ransomware attacks. Third, in terms of statistical methodology, we note that the Anderson-Darling test used in our analysis is not inherently limited to a narrow set of heavy-tailed distributions. Other distributions could be tested through bootstrapping or Monte Carlo simulations to generate appropriate critical values. We also acknowledge that under certain conditions, central limit theorems can still apply even with non-identically distributed data as covered by the Lyapunov condition (Taleb, Reference Taleb2022). Additionally, the Cramer condition may hold if sub-population distributions have finite moments. However, this places the burden of proof on showing that all active ransomware threat actors produce ransom demands with finite moments, rather than across a single aggregate distribution—a demanding requirement in both theory and practice. This raises the possibility that while some threat actors’ ransom demand behaviour may meet insurability thresholds, others may not, particularly those driven by non-financial or geopolitical motives.
Fourth, though our work may demonstrate significant violations of i.i.d. assumptions in ransom payments, we do not claim to have exhaustively characterized the full set of causal mechanisms driving this heterogeneity. This study focuses primarily on ransom payments as a sub-coverage within cyber insurance products, rather than on total ransomware losses. Although we emphasize the conceptual and statistical distinction between ransoms and losses, more research is needed to clarify the relationship between these two variables, especially under different regulatory and threat landscapes. Additionally, although this study has focused on variations in ransom demands as signals of threat actor (predator) behaviour, we stress the need for future research to more fully examine victim-side (prey) dynamics. Organizational revenue, sectoral regulation, geographic location, and technological preparedness may all influence both the likelihood and severity of ransomware losses. Furthermore, we also do not model the effect of ransom payment bans or other policy interventions—an important next step for evaluating the behavioural responses of both attackers and defenders under constrained insurance markets. Future work should explore more granular correlates of ransom variation, including organizational maturity, enforcement regimes, and cybersecurity hygiene investments. Exploring how these factors interact with threat actor strategies will be essential to improving pricing and solvency models, informing regulatory design, and developing a more complete understanding of the conditions under which ransomware risks may or may not be considered insurable.
Further interdisciplinary collaboration between researchers, cybersecurity threat intelligence experts, economists, insurers, and regulators will be essential for addressing the challenges identified here. In particular, there is an urgent need for the development of open, standardized datasets and transparency mechanisms that enable the robust evaluation of solvency models under adversarial risk. As ransomware continues to evolve in tactics and scope, policy and risk modelling frameworks must evolve in parallel—not only to reflect the empirical realities of the threat landscape but also to ensure that insurance remains a tool for resilience rather than a vector for systemic vulnerability.
7. Conclusion
While the use of ruin theory for solvency calculations is suitable for a single distribution risk framework, such determinations do not account for the non-i.i.d. multi-distribution of ransom payments as shown in our findings. Compounding this problem is the industry’s prevailing emphasis on pricing over solvency in underwriting ransomware coverage. Actuarial pricing approaches are fundamentally ill-equipped to capture the deliberate, strategic, and intentional nature of ransomware events and ransom payments. Cyber insurance firms often prioritize competitive premium structures and market share when offering ransom payment coverage, with far less scrutiny given to whether their solvency models can withstand the extreme tail risks posed by ransom demand payments and/or coordinated ransomware campaigns.
Beyond identifying this methodological shortfall, our work also highlights a pressing regulatory vacuum. The pricing-centric approach of cyber insurance firms is mirrored—and in many ways enabled—by regulatory frameworks that continue to endorse ruin theory-based solvency standards for cyber risks without examining assumptions robustly. These models, grounded in assumptions of statistical independence and loss regularity, offer little resilience in the face of systemic cyber threats that violate such premises. At present, insurance regulators lack coherent solvency standards tailored to the realities of ransomware or even general cyber risk. This absence permits cyber insurance firms to underestimate their exposure, compromising not only firm-level stability but also the broader economic resilience of the insurance ecosystem. As a corrective, our work calls for the establishment of minimum solvency requirements for cyber insurance firms and a regular re-examination of the assumptions underpinning the insuring of cyber risks.
Cyber risks also include the potential for criminal enterprises or nation-state adversaries to intentionally induce simultaneous, overwhelming attritional losses via coordinated ransomware attacks, in a strategic bid to severely disrupt or destroy the cyber insurance industry. Extortion is an intentional act, and it is erroneous to assume that all the threat actors are solely financially motivated. Insurance regulators should factor not just catastrophic cybersecurity incidents but also intentionally designed market failures into any solvency metrics and requirements.Footnote 2 Ransomware represents a paradigm shift in cyber risk—one that breaks the assumptions of traditional actuarial logic. Without urgent regulatory reform and a rethinking of risk modelling frameworks, the cyber insurance industry may find itself structurally unprepared for the threats it purports to underwrite.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/dap.2025.10040.
Data availability statement
The data that support the findings of this study are available on request from the corresponding author, D.R. The data are not publicly available due to the sensitive nature of the data.
Author contribution
Conceptualization: E.L.; Data curation: E.L.; Formal analysis: D.R., E.L.; Investigation: D.R., E.L.; Methodology: D.R., E.L.; Resources: D.R., E.L.; Software: D.R., E.L.; Validation: D.R., E.L.; Visualization: D.R., E.L.; Writing - original draft: D.R., E.L.; Writing - review & editing: D.R.
Funding statement
The authors would like to thank the DNS Federation for their financial support of this project while author D.R. was previously at American University in Washington, DC. Author E.L.’s work was supported by a grant from the Lighthill Risk Network, which also supports the economic portfolio modelling of ransomware at Oasis Loss Modelling Framework. Author E.L. would also like to thank the members of the MSR-SIG at FIRST.org for their continued pressure on the ransomware ecosystem.
Competing interests
The authors report no competing interests.
Ethical standard
The research meets all ethical guidelines, including adherence to the legal requirements of the study countries.



Comments
No Comments have been published for this article.