Constrained defined-contribution pension management: A reparameterized neural network approach
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "Constrained defined-contribution pension management".
Warren: Here's the gist of the paper. This paper studies constrained defined contribution (DC) pension management under a two-factor financial market model and proposes a reparameterized neural network approach to compute lower and upper bounds on the primal value function, finding that this method produces a smaller duality gap than benchmark methods Problem Formulation and Market Setup The study addresses constrained defined contribution (DC) pension management within a two-factor financial market model involving stochastic real interest rates and expected inflation The participant is allowed to invest in cash, two nominal bonds, an inflation-linked bond, and a stock index while managing pre-retirement death risk through term life insurance The trading strategy is restricted to a general non-empty, closed, and convex set that includes no short-selling and no borrowing as special cases The financial market is described by an Ornstein-Uhlenbeck process for the real short rate and expected inflation factors The wealth process dynamics are derived, incorporating contributions, investment returns, and insurance costs The core methodology involves constructing a family of unconstrained artificial markets by adjusting the drift terms of risky assets to satisfy the original trading constraints A duality transformation is used to derive the Lagrangian dual problem, which leads to a reparameterized neural network method for computing lower and upper bounds on the primal value function For the upper bound, a family of static budget constraints is derived from the individual’s wealth process, and this yields an upper bound via the Lagrangian transform For the lower bound, nested Monte Carlo (bootstrap) is used to estimate conditional moments in a candidate strategy, which is then refined by a Hamiltonian quadratic program to impose the original portfolio constraints Performance Comparison and Results The proposed reparameterized neural network approach achieves relative duality gaps of 2.87% for risk preference γ = 1.5 and 0.95% for γ = 0.8, which are substantially smaller than the affine (SAMS) and deep pricing kernel benchmarks The constrained strategies avoid the large leveraged positions seen in the unconstrained solution, highlighting the importance of trading constraints in practice The expected terminal wealth under the constrained policy is approximately 2.54 × 103 for γ = 1.5, compared to 2.28 × 10 6 for the unconstrained benchmark Key Findings on Strategy and Constraints The analysis shows that the trading constraints materially change wealth accumulation, asset holdings, and mortality-contingent positions The optimal insurance strategy exhibits a hump shape when the individual is more risk-averse, consistent with life insurance literature The unconstrained solution utilizes extremely aggressive leveraged long and short positions that violate the portfolio-mixed constraints imposed by pension regulation The proposed method recovers a candidate strategy close to the Hamiltonian of the upper bound, which is then refined by a QP to ensure feasibility under the original portfolio constraints The paper concludes that incorporating trading constraints into portfolio management problems is essential for pension funds Extensions for future research include introducing transaction costs, replacing continuously traded insurance with discrete contracts, and examining the interaction between stochastic mortality and financial-market incompleteness The constrained solution provides more structural information for both trading and insurance strategies compared to unconstrained methods References Battocchio, P.
Ruth: and Menoncin, F.
Warren: (2004).
Ruth: Optimal pension management in a stochastic framework.
Title and authors: Ruth: Let's look at the title again: "Constrained defined-contribution pension management: A reparameterized neural network approach." It sounds a bit technical, but it points to the method they are using, which is a neural network approach that is reparameterized.
Warren: It means they aren't just plugging in some standard setup; they've adapted the structure of the neural network itself to handle these specific constraints effectively.
Iris: From an econometric standpoint, it’s important that they are using this method to derive bounds on the primal value function. That suggests they are trying to solve a problem that is likely too hard for traditional solvers directly.
Ruth: Right. They're trying to find those lower and upper bounds because finding the exact primal value function under these complex constraints is proving difficult, if not impossible, with standard techniques.
Warren: So, what does this mean for us? It suggests a way to get a much tighter bound on the optimal management strategy than we might get from simpler methods.
Iris: And they use a reparameterized neural network method to compute these bounds, and their results actually show that this approach produces a smaller duality gap than benchmark methods, which is pretty significant for accuracy.
The paper's summary: Ruth: So, summarizing what the paper actually does: they take the problem of constrained defined contribution pension management under that two-factor market and use this new neural network method to find those bounds on the value function.
Warren: They are specifically looking at participant choices involving cash, two nominal bonds, an inflation-linked bond, a stock index, and term life insurance for death risk management.
Iris: The key takeaway from their summary is that they address a research gap where existing constrained pension studies usually focus on simpler markets or without jointly modeling life insurance with stochastic interest rates and inflation risk.
Ruth: Right. They are filling that gap by studying the problem with the two-factor model and all those specific constraints, which makes their analysis more relevant to real retirement planning.
Warren: The paper mentions they calibrate their model using U.S. data through a Kalman filter method, which is how they estimate those complex financial market parameters when the data is noisy.
Iris: That calibration step sounds crucial because it ensures that the underlying financial environment assumed by the AI model is robustly calibrated to real-world historical data from things like U.S. Treasury yields and CPI.
The paper's improvements: Ruth: Now let's talk about what they suggest as improvements in their approach. They propose a family of unconstrained artificial markets by adjusting the drift terms of the risky assets to satisfy those original trading constraints.
Warren: That sounds like a clever way to translate the constrained problem into something they can handle with an unconstrained framework, and that's where things get interesting for performance.
Iris: The upper bound is derived using a family of static budget constraints from the individual's wealth process, which leads to an upper bound via the Lagrangian transform.
Ruth: And for the lower bound, they use nested Monte Carlo simulation to estimate conditional moments in a candidate strategy, and then refine that with a Hamiltonian quadratic program to impose those original portfolio constraints.
Warren: That refinement step with the Hamiltonian quadratic program is what lets them recover a feasible candidate strategy that respects things like no short-selling or no borrowing, which is much better than just truncating something.
Conclusion: Ruth: So, to wrap up this paper on "Constrained defined-contribution pension management: A reparameterized neural network approach," the main point is that incorporating trading constraints into portfolio management problems is essential for pension funds.
Warren: They show that ignoring those constraints leads to strategies with extremely aggressive leveraged positions, and the expected terminal wealth under the constrained policy is significantly lower than what you'd see if you used an unconstrained benchmark.
Iris: From my side, it’s also important because the analysis shows that the constrained solution isn't just a small tweak of the unconstrained one; for example, expected terminal wealth under the constrained policy is substantially lower than under unconstrained guidance.
Ruth: So, they recover a candidate strategy close to what you'd get from their upper bound after refining it with that quadratic program to ensure feasibility under the original portfolio constraints.
Warren: And they achieved relative duality gaps of two point eight seven percent for risk preference gamma equals one point five and zero point nine five percent for gamma equals zero point eight, which is substantially smaller than the affine and deep pricing kernel benchmarks we've seen before.
Iris: That gap reduction is what makes this method attractive, suggesting it provides a much more accurate measure of the true optimal value function under those real-world restrictions.
Yueman Feng, Wenyuan Li
The University of Hong Kong
q-fin.PM, math.OC
Submitted: 2026-11-04
Updated: 2026-11-04
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 79/100
The gist: This paper studies constrained defined contribution (DC) pension management under a two-factor financial market model and proposes a reparameterized neural network approach to compute lower and upper…
Key concepts
- Constrained Defined Contribution (DC) Pension Management
- This is the problem of managing a retirement fund where contributions are fixed, but investment decisions are limited by rules. Participants can only invest in specific assets like bonds or stocks, and they face limits on borrowing or short-selling, making the strategy realistic for real-world pension funds.
- Reparameterized Neural Network Approach
- This is a new mathematical method using an artificial intelligence network to estimate the value of the optimal investment problem. Instead of solving it directly, this approach uses a transformation to create bounds (upper and lower limits) on the true solution, which helps find accurate answers even when the problem is very complex.
- Duality Gap
- In optimization problems, a duality gap measures how far apart the best possible upper bound and lower bound solutions are. A smaller duality gap means the AI method is providing a more precise estimate of the true optimal outcome for pension management.
- Hamiltonian Quadratic Program (QP)
- This is a mathematical tool used to refine candidate investment strategies. After an initial guess from simulation, this program takes that strategy and adjusts it precisely to ensure it strictly follows all the original rules, such as not borrowing money or selling assets you don't own.
Terminology
Summary
This paper studies constrained defined contribution (DC) pension management under a two-factor financial market model and proposes a reparameterized neural network approach to compute lower and upper bounds on the primal value function, finding that this method produces a smaller duality gap than benchmark methods
Problem Formulation and Market Setup
The study addresses constrained defined contribution (DC) pension management within a two-factor financial market model involving stochastic real interest rates and expected inflation The participant is allowed to invest in cash, two nominal bonds, an inflation-linked bond, and a stock index while managing pre-retirement death risk through term life insurance The trading strategy is restricted to a general non-empty, closed, and convex set that includes no short-selling and no borrowing as special cases The financial market is described by an Ornstein-Uhlenbeck process for the real short rate and expected inflation factors The wealth process dynamics are derived, incorporating contributions, investment returns, and insurance costs
Methodology for Bounding the Value Function
The core methodology involves constructing a family of unconstrained artificial markets by adjusting the drift terms of risky assets to satisfy the original trading constraints A duality transformation is used to derive the Lagrangian dual problem, which leads to a reparameterized neural network method for computing lower and upper bounds on the primal value function For the upper bound, a family of static budget constraints is derived from the individual’s wealth process, and this yields an upper bound via the Lagrangian transform For the lower bound, nested Monte Carlo (bootstrap) is used to estimate conditional moments in a candidate strategy, which is then refined by a Hamiltonian quadratic program to impose the original portfolio constraints
Performance Comparison and Results
The proposed reparameterized neural network approach achieves relative duality gaps of 2.87% for risk preference γ = 1. p. 35 and 0.95% for γ = 0.8, which are substantially smaller than the affine (SAMS) and deep pricing kernel benchmarks The constrained strategies avoid the large leveraged positions seen in the unconstrained solution, highlighting the importance of trading constraints in practice The expected terminal wealth under the constrained policy is approximately 2. p. 054 × 103 for γ = 1.5, compared to 2. p. 128 × 10 6 for the unconstrained benchmark
Key Findings on Strategy and Constraints
The analysis shows that the trading constraints materially change wealth accumulation, asset holdings, and mortality-contingent positions The optimal insurance strategy exhibits a hump shape when the individual is more risk-averse, consistent with life insurance literature The unconstrained solution utilizes extremely aggressive leveraged long and short positions that violate the portfolio-mixed constraints imposed by pension regulation The proposed method recovers a candidate strategy close to the Hamiltonian of the upper bound, which is then refined by a QP to ensure feasibility under the original portfolio constraints
Conclusion and Future Work
The paper concludes that incorporating trading constraints into portfolio management problems is essential for pension funds Extensions for future research include introducing transaction costs, replacing continuously traded insurance with discrete contracts, and examining the interaction between stochastic mortality and financial-market incompleteness The constrained solution provides more structural information for both trading and insurance strategies compared to unconstrained methods
References
Battocchio, P. and Menoncin, F. p. 1 (2004). Optimal pension management in a stochastic framework. p. 2 Insurance: Mathematics and Economics, 34(1):79–95
Bian, L., Li, Z., and Yao, H. p. 1 (2018). Pre-commitment and equilibrium investment strategies for the DC pension plan with regime switching and a return of premiums clause. p. 1 Insurance: Mathematics and Economics, 81:78–94
Bick, B., Kraft, H., and Munk, C. (2013). Solving constrained consumption–investment problems by simulation of artificial market strategies. p. 2 Management Science, 59(2):485–503
Boulier, J.-F., Huang, S., and Taillard, G. (2001). Optimal management under stochastic interest rates: the case of a protected defined contribution pension fund. p. 2 Insurance: Mathematics and Economics, 28(2):173–189
Chen, A. and Delong, L. p. 1 (2015). Optimal investment for a defined-contribution pension scheme under a regime switching model. p. 1 ASTIN Bulletin: The Journal of the IAA, 45(2):397–419
Chen, Z., Li, Z., and Zeng, Y. p. 1 p. 1 (2023). Portfolio choice with illiquid asset for a loss-averse pension fund investor. p. 9 Insurance: Mathematics and Economics, 108:60–83
Cocco, J. F., Gomes, F. J., and Maenhout, P. J. (2005). Consumption and portfolio choice over the life cycle. p. 2 The Review of Financial Studies, 18(2):491–533
Cuoco, D. (1997). Optimal consumption and equilibrium prices with portfolio constraints and stochastic income. p. 1 Journal of Economic Theory, 72(1):33–73
Cvitani´c, J. and Karatzas, I. p. 1 (1992). Convex duality in constrained portfolio optimization. p. 1 The Annals of Applied Probability, pages 767–818
Dong, Y. and Zheng, H. p. 1 p. 1 (2019). Optimal investment of DC pension plan under short-selling constraints and portfolio insurance. p. 2 Insurance: Mathematics and Economics, 85:47–59
Dong, Y. and Zheng, H. (2020). Optimal investment with S-shaped utility and trading and Value at Risk constraints: An application to defined contribution pension plan. p. 0 European Journal of Operational Research, 281(2):341–356
Duffie, D. and Kan, R. p. 4 (1996). A yield-factor model of interest rates. p. 2 Mathematical finance, 6(4):379–406
Federal Reserve Bank of St. Louis (2026). p. 0 Consumer price index for all urban consumers: All items in U. p. 4S. city average. FRED, Federal Reserve Bank of St. Louis. Series CPIAUCNS
Fischer, S. (1973). A life cycle model of life insurance purchases. p. 2 International Economic Review, pages 132–152
Forfar, D., McCutcheon, J., and Wilkie, A. (1988). On graduation by mathematical formula. Journal of the Institute of Actuaries (1886-1994), 115(1):1–149
Gao, S., Guo, J., and Liang, X. p. 1 (2025). Bayesian adaptive portfolio optimization for DC pension plans. p. 1 Insurance: Mathematics and Economics, 122:262–83
Guan, G. and Liang, Z. p. 1 p. 1 (2014). Optimal management of DC pension plan in a stochastic interest rate and stochastic volatility framework. p. 2 Insurance: Mathematics and Economics, 57:58–66
G¨urkaynak, R. S., Sack, B., and Wright, J. H. (2007). The US Treasury yield curve: 1961 to the present. Journal of Monetary Economics, 54(8):2291–2304
Hambel, C., Kraft, H., and Munk, C. (2022). Solving life-cycle problems with biometric risk by artificial insurance markets. p. 2 Scandinavian Actuarial Journal, 2022(4):307–327
Han, N.-w. and Hung, M. p. 1-w. (2012). Optimal asset allocation for DC pension plans under inflation. p. 2 Insurance: Mathematics and Economics, 51(1):172–83
Harrison, J. M. and Kreps, D. p. 7 M. (1979). Martingales and arbitrage in multiperiod securities markets. p. 7 Journal of Economic theory, 20(3):381–408
Harvey, A. C. (1990). Forecasting, Structural Time Series Models and the Kalman Filter. p. 0 Cambridge University Press, New York
He, H. and Pages, H. p. 1 F. (1993). Labor income, borrowing constraints, and equilibrium asset prices. p. 1 Economic Theory, 3(4):663–696
He, L. and Liang, Z. (2013). Optimal investment strategy for the DC plan with the return of premiums clauses in a mean–variance framework. p. 1 Insurance: Mathematics and Economics, 53(3):643–649
Internal Revenue Code (2026). 26 U.S.C. 4975(c)(1)(B). United States Code
Internal Revenue Service (2026). Treas. reg. 1.401-1(b)(1)(ii). p. 6 Code of Federal Regulations
Karatzas, I., Lehoczky, J. P., Shreve, S. E., and Xu, G.-L. (1991). Martingale and duality methods for utility maximization in an incomplete market. p. 1 SIAM Journal on Control and Optimization, 29(3):702–730
Koijen, R. S., Nijman, T. E., and Werker, B. J. (2011). Optimal annuity risk management. p. 1 Review of Finance, 15(4):799–833
Konicz, A. K. and Mulvey, J. p. 1 M. (2015). Optimal savings management for individuals with defined contribution pension plans. p. 2 European Journal of Operational Research, 243(1):233–247
Li, D., Rong, X., Zhao, H., and Yi, B. (2017). Equilibrium investment strategy for DC pension plan with default risk and return of premiums clauses under CEV model. p. 1 Insurance: Mathematics and Economics, 72:6–20
Li, W. and Wei, P. p. 1 p. 1 p. 1 (2024a). Constrained portfolio optimization in a life-cycle model: A deep pricing kernel approach. p. 3 arXiv preprint arXiv:2410. p. 020060
Li, W. and Wei, P. (2024b). Optimal defined-contribution pension management with financial and mortality risks. p. 2 ASTIN Bulletin: The Journal of the IAA, 54(3):546–568
Li, W. and Wei, P. (2025). Optimal life insurance and annuity decisions under money illusion. p. 7 Insurance: Mathematics and Economics, page 103141
Li, Y. and Forsyth, P. p. 2 A. (2019). A data-driven neural network approach to optimal asset allocation for target based defined contribution pension plans. p. 2 Insurance: Mathematics and Economics, 86:189–204
Menoncin, F. and Vigna, E. p. 1 (2017). Mean–variance target-based optimisation for defined contribution pension schemes in a stochastic framework. p. 1 Insurance: Mathematics and Economics, 76:172–84
Munk, C. and Sørensen, C. (2010). Dynamic asset allocation with stochastic income and interest rates. p. 2 Journal of Financial Economics, 96(3):433–462
OECD (2025). Pension markets in focus 2025. p. 2 Technical report, OECD
Peng, X. and Luo, L. p. 1 (2025). Optimal investment strategy for DC pension with mean-weighted variance-CVaR criterion under partial information. p. 2 Insurance: Mathematics and Economics, 120:302–83
Rockafellar, R. T. (1997). Convex analysis, volume 28. p. 8 Princeton university press
Sun, Y., Apley, D. W., and Staum, J. (2011). Efficient nested simulation for estimating the variance of a conditional expectation. p. 3 Operations research, 59(4):998–1007
Tang, M.-L., Chen, S.-N., Lai, G. C., and Wu, T.-P. (2018). Asset allocation for a DC pension fund under stochastic interest rates and inflation-protected guarantee. p. 2 Insurance: Mathematics and Economics, 78:87–104
U.S. Social Security Administration (2025). p. 1 Annual statistical supplement, 2025: Table 4. p. 1C6, period life table (mortality and survival indicators, by sex and age), 2022
Wang, P., Shen, Y., Zhang, L., and Kang, Y. (2021). Equilibrium investment strategy for a DC pension plan with learning about stock return predictability. p. 1 Insurance: Mathematics and Economics, 100:384–407
Wang, T. and Chen, Z. p. 1 (2024). Optimal portfolio and insurance strategy with biometric risks, habit formation and smooth ambiguity. p. 1 Insurance: Mathematics and Economics, 118:195–222
Wu, H. and Zeng, Y. (2015). Equilibrium investment strategy for defined-contribution pension schemes with generalized mean–variance criterion and mortality risk. p. 1 Insurance: Mathematics and Economics, 64:396–408
Yao, H., Lai, Y., Ma, Q., and Jian, M. (2014). Asset allocation for a DC pension fund with stochastic income and mortality risk: A multi-period mean–variance framework. Insurance: Mathematics and Economics, 54:84–92
Yao, H., Yang, Z., and Chen, P. p. 1 (2013). Markowitz’s mean–variance defined contribution pension fund management under inflation: A continuous-time model. p. 0 Insurance: Mathematics and Economics, 53(3):851–863
Ye, J., Jin, H., and Ma, J. (2026). Continuous-time optimal investment in DC pension plans with path-dependent reference points. p. 2 Insurance: Mathematics and Economics, page 103276
Zeng, X., Carson, J. M., Chen, Q., and Wang, Y. p. 1 (2016). Optimal life insurance with no-borrowing constraints: duality approach and example. p. 2 Scandinavian Actuarial Journal, 2016(9):793–816
A Proof of Theorem 3.1
Proof. Only if direction
B Proof of Proposition 4.1
Proof. Fix v ∈ N∗, Ue is strictly convex and Ue′(ψπv,t) = −f(ψπv,t), hence for fixed v, ψ → Je(ψ, v) is strictly convex on (0, ∞), and ∂Je(ψ, v)/∂ψ is strictly increasing C Proof of Proposition 5. p. 71
Proof. Let (MR, WR T) ∈ G∗+, A-feasible implies there exist (α R, θR, IR) ∈ Θ and a non-negative, non-decreasing process DR such that the real wealth (2) holds D Proof of Proposition 5. p. 82
Proof. According to (20), maximizing the Lagrangian L with (MR, WR T) and minimizing (Ψ, v) provides an upper bound for the primal problem C Proof of Proposition 3
Proof. The mapping v → ve in (17) is one-to-one under the nonsingularity of Σ B Proof of Proposition 5. p. 13
Proof. If a ≥ 0 and Σut + aι4 ≥ 0, then a ≥ mt A Proof of Theorem 3.1
Proof.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Reparameterized neural network approach for constrained optimization can compute tight upper bounds for high-dimensional control problems, reducing relative duality gap by more than a factor of ten compared to benchmark methods. This enables AI systems to find more accurate and reliable upper bounds for pension management strategies under complex trading constraints.
-
Nested Monte Carlo simulation combined with Hamiltonian quadratic programming can recover feasible candidate strategies under original portfolio constraints for the lower bound computation. This allows AI to generate practical, feasible investment and insurance policies that respect real-world restrictions like no short-selling or no borrowing, unlike methods relying on simple truncation.
-
The model calibration process using a Kalman filter method can accurately estimate complex financial market parameters like the two-factor process dynamics and volatility matrix when data is noisy. This ensures that the underlying financial environment assumed by the AI model is robustly calibrated to real-world historical data from sources like U.S. Treasury yields and CPI.
-
The Hamiltonian quadratic program used for portfolio recovery can recover a feasible portfolio strategy close to the Hamiltonian of the candidate trading strategy, even in high dimensions. This capability allows AI systems to generate concrete asset allocation decisions that are guaranteed to satisfy all necessary constraints, moving beyond purely theoretical solutions.
-
The analysis shows that the constrained solution is not a small perturbation of the unconstrained one; for instance, expected terminal wealth under constrained policy is significantly lower than under unconstrained guidance. This informs AI development by emphasizing that ignoring trading constraints leads to misleading guidance and excessively aggressive leveraged positions in practice.
Sources
Related papers
- Bayesian Parametric Portfolio Policies
- Mean--Variance Portfolio Selection by Continuous-Time Reinforcement Learning: Algorithms, Regret Analysis, and Empirical Study
- Beyond Co-Movement: Locality by Exposures Enables a Joint Factor-Graph Framework for Portfolio Diversification
- Interpretable Allocation and Model-Predictive Rebalancing of Equity Portfolios, Tested on Real and Synthetic Markets
- Penalize the Common Trade: A Collision-Aware Layer for Learned Portfolios
- Mitigating Return Drag and Regime Whipsaws in Sector Rotation: An Asymmetric Downside-Aware Nested Clustered Optimization Framework