Guaranteed Minimum Withdrawal Benefit (GMWB)

Updated 6 July 2026

GMWB is a variable annuity rider that guarantees periodic premium withdrawals by decrementing a benefit account while allowing exposure to market gains.
The product typically uses a two-account structure—an investment account and a guarantee account—where withdrawals reduce both and penalties apply for excess amounts.
Valuation involves solving an optimal stochastic control problem that incorporates dynamic policyholder behavior, surrender options, and sensitivity to market and model risk.

Searching arXiv for recent and foundational papers on GMWB to ground the article in the literature. {} Guaranteed Minimum Withdrawal Benefit (GMWB) is a variable-annuity rider that guarantees the return of premiums in the form of periodic withdrawals while allowing policyholders to participate fully in market gains. In the formulations studied across the literature, the policyholder pays an initial premium into a market-linked account, withdrawals reduce a separate guarantee balance or benefit base, and the contract typically pays any remaining account value at maturity; if the account is depleted earlier, the guarantee continues to support withdrawals until the guaranteed amount is exhausted. Under dynamic policyholder behavior, GMWB valuation is an optimal stochastic control problem, and once surrender or lapse is admitted it becomes a combined control-and-stopping problem (Hyndman et al., 2013, Luo et al., 2015).

1. Contract architecture and payoff structure

A recurring formulation uses two state accounts. The wealth account $W(t)$ or account value $A_t$ is the invested balance linked to a risky asset or fund, while the guarantee account $A(t)$ , benefit account $B_t$ , or benefit base records the remaining withdrawal entitlement. In discrete-time models with withdrawal dates $0=t_0<t_1<\dots<t_N=T$ , the initial conditions are typically $W(0)=A(0)$ or $A_0=B_0=P$ , where $P$ is the premium. Withdrawals reduce both the investment account and the guarantee balance, usually dollar-for-dollar in the classical GMWB specification (Luo et al., 2015, Luo et al., 2014).

The contractual withdrawal amount is commonly written as

$G_n=W(0)\frac{t_n-t_{n-1}}{T}, \qquad g=\frac{1}{T},$

or, in continuous-withdrawal formulations, $G=gP$ with $A_t$ 0. In the canonical penalty structure, the cashflow at a withdrawal date is

$A_t$ 1

where $A_t$ 2 is the penalty rate on the excess above the contractual amount. After withdrawal,

$A_t$ 3

At maturity, a standard terminal payoff is

$A_t$ 4

or, equivalently in alternative notation, the larger of the residual investment account and the remaining guarantee net of penalty (Shevchenko et al., 2016, Luo et al., 2014).

Between withdrawal dates, the investment account evolves under the risky asset net of fees. In one common risk-neutral Black–Scholes specification,

$A_t$ 5

and

$A_t$ 6

where $A_t$ 7 is the continuous fee. In continuous-withdrawal models, the account dynamics may be written as

$A_t$ 8

with non-negativity enforced through a positive-part representation (Hyndman et al., 2013, Luo et al., 2014).

Later product designs enrich this architecture. Ratchet contracts replace the return-of-premium-type benefit base by a base that can step up to the post-fee account value, with guaranteed withdrawal amount

$A_t$ 9

while some taxed or hybrid designs track an additional tax base $A(t)$ 0 or a cash fund that accumulates unwithdrawn guaranteed amounts at a contractual cash rate $A(t)$ 1 (Alonso-Garcia et al., 10 Jul 2025, Molent, 2019).

2. Valuation perspectives and contract decompositions

A central analytical distinction is between the policyholder’s total contract value and the insurer’s residual guarantee value. In the continuous-time decomposition developed for static-withdrawal GMWBs, the policyholder value at inception is

$A(t)$ 2

The fair fee $A(t)$ 3 is defined by the no-arbitrage condition

$A(t)$ 4

Under the standing assumption $A(t)$ 5, $A(t)$ 6 is continuous and strictly decreasing in $A(t)$ 7, which yields existence and uniqueness of the fair fee (Hyndman et al., 2013).

From the insurer’s viewpoint, the rider is the embedded guarantee net of future fee income. If $A(t)$ 8 denotes the trigger time at which the account first hits zero, one representative insurer value process is

$A(t)$ 9

with $B_t$ 0. The core identity is

$B_t$ 1

or, in discrete-time binomial form, $B_t$ 2. This decomposition separates the already-funded account component from the residual guarantee liability and is extended in the literature to include lapses, mortality, and death benefits (Hyndman et al., 2014, Hyndman et al., 2013).

When surrender is permitted, the no-lapse rider value $B_t$ 3 and the value of the surrender option $B_t$ 4 satisfy

$B_t$ 5

This isolates the economic value of early termination flexibility. In the lapse model of the continuous-time decomposition, the policyholder chooses a stopping time $B_t$ 6 to maximize

$B_t$ 7

where $B_t$ 8 is a deterministic surrender-charge schedule (Hyndman et al., 2013).

Analytical work based on hitting-time identities shows that policyholder-side and insurer-side fair-charge formulations can coincide under a consistent no-arbitrage model. In particular, under the assumptions of the hitting-time approach and with $B_t$ 9, the fair charge obtained from the policyholder’s viewpoint and the fair charge obtained from the insurer’s viewpoint are the same (Feng et al., 2013).

3. Optimal control, surrender, and policyholder behavior

Under dynamic withdrawals, GMWB valuation is a finite-horizon stochastic control problem. The Bellman recursion takes the generic form

$0=t_0<t_1<\dots<t_N=T$ 0

with terminal condition $0=t_0<t_1<\dots<t_N=T$ 1. In discrete withdrawal-date formulations, the jump condition is

$0=t_0<t_1<\dots<t_N=T$ 2

Once surrender is admitted, the recursion becomes a combined control-and-stopping problem by taking the maximum of continuation value and surrender value (Luo et al., 2015, Langrené et al., 26 May 2026).

The surrender feature is economically important because it allows the policyholder to terminate before maturity. One numerical setup uses

$0=t_0<t_1<\dots<t_N=T$ 3

as surrender payoff. Reported results indicate that the extra value added by the surrender option can be very significant. At $0=t_0<t_1<\dots<t_N=T$ 4 and $0=t_0<t_1<\dots<t_N=T$ 5, the fees for GMWB-S and GMWB are almost identical, whereas at $0=t_0<t_1<\dots<t_N=T$ 6 the fee for GMWB-S can be much larger, and when the penalty drops from $0=t_0<t_1<\dots<t_N=T$ 7 to $0=t_0<t_1<\dots<t_N=T$ 8, the fee for GMWB-S can be more than twice the fee of plain GMWB in some cases (Luo et al., 2015).

A recurrent simplification is the bang-bang strategy, in which the holder can only choose among no withdrawal, withdrawal at the contractual amount, and full surrender. For GMWB, this simplification is delicate. A convexity-based analysis shows that the related GMWB contract is not convexity preserving, and hence does not satisfy the bang-bang principle other than in certain degenerate cases such as $0=t_0<t_1<\dots<t_N=T$ 9 or $W(0)=A(0)$ 0. Computationally, this means that one cannot generally reduce the optimization to a finite, state-independent control set, and a search over the control continuum is typically required (Azimzadeh et al., 2015).

At the same time, numerical evidence shows that bang-bang can still be a useful approximation. In the surrender study, bang-bang fees are typically only about $W(0)=A(0)$ 1 lower at most than the optimal-with-surrender case, and the static-with-surrender contract is less than $W(0)=A(0)$ 2 smaller than bang-bang, indicating that the “no-withdrawal” option adds little value in that setting (Luo et al., 2015). This suggests a distinction between structural non-equivalence and numerical near-equivalence: the bang-bang principle does not generally hold for GMWB, but restricted-action policies may nonetheless approximate optimal values well in some parameter regimes.

Behavioral assumptions also create model risk. Under the benchmark approach with the minimal market model (MMM), a policyholder may optimally delay withdrawals because the model recognizes long-term growth and leverage effects, whereas an insurer reserving under Black–Scholes may expect earlier or smaller withdrawals. In the reported asymmetric case of a BSM insurer and MMM policyholder, the reserve account ends with a deficit of about 1 million units (Sun et al., 2019).

4. Market models and valuation methodologies

The GMWB literature spans a wide range of market models. Classical formulations use Black–Scholes dynamics with deterministic rates; extensions introduce Vasicek or Hull–White short-rate models, Heston stochastic volatility, Heston–Hull–White joint stochastic volatility and interest rate, jump-diffusions of Merton or Kou type, and the benchmark approach with the minimal market model (Shevchenko et al., 2016, Goudenège et al., 2016, Goudenège et al., 2019, Lu et al., 2023, Sun et al., 2019).

Under stochastic rates, a particularly influential device is the bond-price numeraire. In the Vasicek setting, the bond price

$W(0)=A(0)$ 3

permits a change of measure under which the discounted continuation value over one withdrawal interval becomes a two-dimensional Gaussian expectation in $W(0)=A(0)$ 4. This transforms the backward step into repeated evaluation of two-dimensional integrals and is central to the Gauss–Hermite quadrature on cubic spline interpolation methodology (Shevchenko et al., 2016).

Several numerical engines have been proposed.

The GHQC method computes the expectation in the backward step by high-order Gauss–Hermite quadrature applied to cubic-spline interpolation. In the one-factor setting it replaces finite-difference PDE time stepping between withdrawal dates; in the stochastic-rate Vasicek setting it is extended to two-dimensional integrals after the numeraire change. Reported comparisons show very close agreement with finite-difference and Monte Carlo benchmarks and substantial speed gains (Luo et al., 2014, Shevchenko et al., 2016).

The binomial framework provides a discrete-time pricing and hedging model with explicit perfect hedging strategies funded using only periodic fee income. In the no-lapse case, if the initial capital is set equal to the rider value $W(0)=A(0)$ 5, the replicating portfolio satisfies $W(0)=A(0)$ 6 pathwise; with lapses, the hedge remains valid and additional insurer consumption arises if the policyholder behaves suboptimally (Hyndman et al., 2014).

For higher-dimensional models, the literature uses hybrid tree-finite difference, Hybrid Monte Carlo, ADI finite difference, and Standard Monte Carlo. In the comparison of Heston and Black–Scholes–Hull–White models, the hybrid PDE method is identified as offering the best overall balance of speed, stability, and accuracy for pricing and Greeks, whereas Monte Carlo remains useful for confidence intervals and scenario-based risk analysis (Goudenège et al., 2016).

Machine-learning and simulation-based surrogates have also become prominent. In Heston–Hull–White, Gaussian Process Regression (GPR) is trained on prices and Greeks produced by a Hybrid Tree-PDE solver and can be used to compute prices, Greeks, and the no-arbitrage fee with very low computational cost (Goudenège et al., 2019). In general endogenous-control settings, Deep Least Squares Monte Carlo modifies classical LSMC through control randomization and regression on state-control pairs; polynomial LSMC can be very accurate but requires manual feature engineering, while neural network LSMC is slightly less accurate, requires more training time, and appears more stable when interest rates become stochastic (Langrené et al., 26 May 2026).

Continuous-withdrawal GMWBs with jump-diffusion and stochastic interest rate lead to a three-dimensional HJB quasi-variational inequality with cross derivatives. A recent numerical approach combines a semi-Lagrangian discretization with Green’s-function/Fourier convolution to produce an $W(0)=A(0)$ 7-monotone Fourier method and proves convergence to the viscosity solution as $W(0)=A(0)$ 8 (Lu et al., 2023).

Alongside these numerical methods, the hitting-time approach provides analytical and semi-analytical pricing formulas. By exploiting an identity in distribution between the hitting time of Yor’s process and the hitting time of an affine-drift diffusion, one obtains analytical solutions for the fair charge from both the policyholder’s and insurer’s viewpoints (Feng et al., 2013).

5. Product extensions: mortality, death benefits, taxation, ratchets, and low rates

Mortality and death benefits materially alter the structure of the problem. In GMWDB models, the state expands to $W(0)=A(0)$ 9, where $A_0=B_0=P$ 0 indicates whether the policyholder is alive, has just died, or died earlier. Three death-benefit designs are studied: $A_0=B_0=P$ 1 Under optimal withdrawals, DB0 adds only modestly to fair fees, but DB1 and DB2 can become problematic for long maturities if the extra cost is charged continuously as a proportion of account value: in some cases, there is no fair fee solution at all. The proposed alternative is to keep the standard continuous GMWB fee and charge the extra death-benefit cost upfront or in fixed periodic installments (Luo et al., 2014).

Mortality also interacts with hedging. In the binomial mortality extension, biometric risk is assumed independent of financial risk and diversified in large pools. The aggregate number of deaths and survivors satisfies strong-law limits, and the aggregate hedging portfolio converges to the aggregate rider value on a per-policy or per-premium basis (Hyndman et al., 2014).

Taxation introduces another layer of nonlinearity because post-tax value differs for the policyholder and the insurer. In the Moenig–Bauer style taxed GMWB with Black–Scholes–Hull–White rates, the state variables are $A_0=B_0=P$ 2, where $A_0=B_0=P$ 3 is the tax base. Withdrawals are taxed under a last-in first-out convention, outside investments used to replicate the policy are taxed at rate $A_0=B_0=P$ 4, and valuation becomes subjective rather than unique. The policyholder value satisfies a nonlinear implicit equation, while the insurer value remains a standard risk-neutral expectation. Numerically, taxation reduces optimal withdrawals and lowers the insurer’s break-even fee, while the policyholder’s subjective value can exceed the insurer’s value at the insurer’s break-even fee (Molent, 2019).

Ratchet and cash-fund designs alter the meaning of the guarantee base. In the ratchet GMWB with taxation and an internal cash fund, the guaranteed withdrawal amount is a fixed proportion of a benefit base that can step up to the post-fee account value, and any shortfall between the guaranteed amount and the chosen withdrawal can be transferred to a cash fund earning a contractual cash rate $A_0=B_0=P$ 5. The cash fund’s principal is not taxed as income, while its interest earnings are taxed; the paper characterizes the resulting mechanism as a tax-shielding effect. Reported numerical results show that the ratchet tends to discourage early surrender, the cash fund discourages active withdrawals, and sufficiently high tax rates can eliminate the market entirely without ratchets or cash fund, whereas ratchet plus cash fund can restore viability (Alonso-Garcia et al., 10 Jul 2025).

Low and negative interest-rate environments affect both discounting and the equity-account drift under $A_0=B_0=P$ 6. In a general GMWB with step-up, bonus, surrender, mortality, and death benefits under a correlated Hull–White short rate, low or negative rates profoundly affect optimal withdrawal behavior and can increase fair values significantly. In the reported 2021 low/negative-rate scenario, a standard GMWB can be priced at par only with very large management fees and penalties; without step-up, dynamic GMWB can be priced at par around $A_0=B_0=P$ 7 with $A_0=B_0=P$ 8, whereas in the 2022 positive-rate scenario the same contract can be priced at par with $A_0=B_0=P$ 9 for $P$ 0 (Fontana et al., 2022).

6. Quantitative sensitivities, hedging, and recurring research issues

A dominant empirical finding is that GMWB values and fair fees are highly sensitive to volatility, interest-rate dynamics, withdrawal flexibility, and surrender penalties. In the surrender study, quarterly-withdrawal contracts with $P$ 1, $P$ 2, and $P$ 3 show near equality between GMWB and GMWB-S fees at $P$ 4, but much larger fees for GMWB-S at $P$ 5; the paper emphasizes that ignoring surrender can substantially underprice the contract in volatile or low-penalty environments (Luo et al., 2015).

Stochastic rates have a pronounced impact even when the effect on contract price is smaller than the effect on fair fee. In the Vasicek study, with positive asset-rate correlation $P$ 6, the static fair fee is about 143 basis points under stochastic rates versus 95.8 basis points under deterministic rates, while the dynamic fair fee is about 188 bp versus 136 bp. With negative correlation $P$ 7, the static difference is negligible but the dynamic fair fee remains about 161 bp versus 136 bp, still roughly 19\% higher (Shevchenko et al., 2016).

Jointly modeling jumps and stochastic rates changes the decomposition of cost drivers. For continuous-withdrawal GMWBs, four models are compared: GBM-C, GBM-V, JD-C, and JD-V. Reported results show that stochastic interest rates increase prices and fair fees relative to constant rates, while jumps reduce fair fees substantially compared with diffusion-only models; under $P$ 8, the relative fee change versus GBM-C ranges from about 0.8\% for GBM-V to 25.1\% for JD-C, whereas the corresponding price change is around 0.1\% to 2.7\% (Lu et al., 2023).

Hedging and reserve adequacy are likewise model-sensitive. In the binomial setting, perfect hedging strategies funded only by periodic fee income can replicate the rider exactly under the model assumptions, and mortality risk diversifies in large pools (Hyndman et al., 2014). By contrast, when pricing, reserving, and policyholder optimization are based on different market models, substantial hedge deficits can appear, as in the BSM-insurer/MMM-policyholder case (Sun et al., 2019).

Fee design itself can be used to manage sensitivity. In a stochastic-volatility-with-jumps framework, replacing a fixed rider fee by a VIX-linked fee makes fee income more positively aligned with volatility-sensitive liability. The reported fair base fees decline as the VIX-link multiplier $P$ 9 increases, from approximately 2.4650\% at $G_n=W(0)\frac{t_n-t_{n-1}}{T}, \qquad g=\frac{1}{T},$ 0 to 1.0300\% at $G_n=W(0)\frac{t_n-t_{n-1}}{T}, \qquad g=\frac{1}{T},$ 1, and the insurer’s liability becomes less sensitive to temporary volatility changes (Kouritzin et al., 2017).

Several recurrent issues therefore define current research on GMWB. One is the gap between structural control theory and practical approximation: bang-bang control is not generally valid for GMWB, yet restricted strategies can be close numerically. Another is state-space growth: taxation, mortality, stochastic rates, stochastic volatility, jumps, and ratchets expand the control problem rapidly and motivate quadrature, tree, Fourier, surrogate, and deep-LSMC methods. A plausible implication is that GMWB valuation is best understood not as a single closed-form pricing problem, but as a family of embedded-option problems whose tractability and risk profile depend sharply on contract design, behavioral specification, and market model.