{"paper_id":"2410.05524","title":"Deep Learning Methods for S Shaped Utility Maximisation with a Random Reference Point","authors":["Ashley Davey","Harry Zheng"],"paper_date":"2024-10-07","paper_type":"theoretical","paper_url":"https://arxiv.org/pdf/2410.05524v1","page_url":"https://thequant.space/flowcharts/deep-learning-methods-for-s-shaped-utility-maximisation-with/","scores":{"empirical_rigor":3,"math_complexity":9.5,"quadrant":"Lab Rats"},"proposes":{"status":"available","statement":"The paper proposes several deep learning based numerical methods, including Physics Informed Neural Networks and deep BSDE/SMP solvers, to solve the S‑shaped utility maximisation problem with a random benchmark in both complete and incomplete markets.","quote":"We develop several numerical methods for solving the problem using deep learning and duality methods. We use deep learning methods to solve the associated Hamilton-Jacobi-Bellman equation for both the primal and dual problems, and the adjoint equation arising from the stochastic maximum principle.","summary":"We consider the portfolio optimisation problem where the terminal function is an S-shaped utility applied at the difference between the wealth and a random benchmark process. We develop several numerical methods for solving the problem using deep learning and duality methods. We use deep learning…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The investment horizon is a fixed interval [0,T] with a risk‑free asset following dS0_t = αS0_t dt and a risky asset following dS_t = μS_t dt + σS_t dW_t.","type":"model","stated":"explicit","quote":"PROBLEM FORMULATION Assume (Ω,F,{Ft},P) is a filtered probability space with {Ft} the natural filtration gen-erated by two standard Brownian motions W,W R with correlation ρ ∈[−1,1], augmented with all P-null sets. Assume the investment period is [0,T] and the risk-free asset price S0 and the risky asset price S satisfy the following stochastic differential equations (SDEs), for t ∈[0,T], dS0 t = αS0 t dt, dSt =","support":"full"},{"assumption":"The two Brownian motions driving the assets have correlation ρ ∈ [−1,1].","type":"model","stated":"explicit","quote":"PROBLEM FORMULATION Assume (Ω,F,{Ft},P) is a filtered probability space with {Ft} the natural filtration gen-erated by two standard Brownian motions W,W R with correlation ρ ∈[−1,1], augmented with all P-null sets.","support":"full"},{"assumption":"The benchmark is a non‑negative random variable (or process) that may be non‑replicable in incomplete markets.","type":"other","stated":"explicit","quote":"INTRODUCTION In this paper we apply the deep learning method to solve utility maximisation problem with a benchmark that is a non-negative random variable.","support":"full"},{"assumption":"For the SMP formulation a complete market is assumed so that the random reference R is replicable.","type":"market","stated":"explicit","quote":"To apply the SMP as it is given in the literature, for example [21], we assume a complete market here, so RT is replicable.","support":"full"},{"assumption":"The control variable π (or Π) is admissible, i.e., the resulting wealth process remains non‑negative where required.","type":"model","stated":"explicit","quote":"We therefore require Π to be in the admissible control set, defined by, for (t,x) ∈[0,T]×[0,∞), A (t,x) = \b Π ∈H 2 t (R) Xt,x,Π s ≥0 a.s. for all s ∈[t,T] , where Xt,x,Π satisfies (4.3) with Xt,x,Π t = x.","support":"full"},{"assumption":"The utility function is S‑shaped: concave on (0,∞) and convex on (−∞,0).","type":"model","stated":"explicit","quote":"The agent wishes to maximise an S-shaped utility, a function which is concave on (0,∞) and convex on (∞,0), applied at the difference between their terminal wealth and the benchmark.","support":"full"}]},"datasets":{"status":"available","items":[{"name":"Synthetic grid data for PINN training","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"We choose a grid size X ..= (Xmin,Xmax)d ∈(0,∞)d, then uniformly sample three batches: the collocation data within the time horizon, given by (t j,Xj) ∈[0,T)×X for j = 1,...,Mc, the boundary data (T,Xb j ) ∈{T} × X for j = 1,...,Mb and the zero state data (ts j,Xs j) ∈ S SHAPED UTILITY MAXIMISATION WITH A RANDOM REFERENCE POINT 9 [0,T]×∂(Xmin,Xmax)d for j = 1,...,Ms. Using these sample data we define the loss","support":"full"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"synthetic"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R3","candidates":{"front_matter":"n/a","evidence_card":"n/a","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"none","findings":[{"finding":"Numerical experiments for power and log utilities demonstrate that the deep learning algorithms achieve high accuracy compared to reference solutions.","kind":"performance","quote":"We give some numerical results for power and log utilities to show the accuracy of the suggested algorithms. Results show the PINN-based algorithms that reduce dimensionality of the problem, solving the scaled problem with z = x r, then rescaling.","support":"full"},{"finding":"The PINN‑based methods reduce dimensionality and successfully solve the scaled problem, with results matching those of the dual and primal formulations.","kind":"performance","quote":"Results show the PINN-based algorithms that reduce dimensionality of the problem, solving the scaled problem with z = x r, then rescaling. (x, r) Primal General Primal Non-Conc General Dual General SMP Solution (0.5, 1.0) -0.0188 -0.0577 -0.0050 -0.0450 -0.0149 (1.0, 1.0) 0.3793 0.3340 0.3866 0.3735 0.3872 (5.0, 1.0) 2.0546 2.0631 2.0787 2.0678 2.0653 (1.0, 0.5) 0.7335 0.7429 0.7320 0.7361 0.7353 (1.0, 5.0) -0.6758","support":"full"},{"finding":"In incomplete markets, the primal and dual value approximations agree, suggesting that the concavification principle still holds despite lack of analytical solutions.","kind":"theoretical","quote":"However, the concave primal, and dual PINN methods appear to agree, suggesting the concavification principle still occurs. As the primal and dual val-ues coincide, up to numerical errors, we conclude that the payoff XT = −∂y ˜U(YT,RT) has been replicated by the primal algorithm, where YT is the optimal payoff for the dual problem.","support":"full"},{"finding":"The SMP method computes solutions in about one minute and produces consistent average results over multiple runs.","kind":"performance","quote":"The SMP method takes about 1 minute. We run the algorithms 10 times and present the average results.","support":"full"},{"finding":"Overall, the deep neural network approach is shown to be viable and robust for solving nonlinear HJB equations and adjoint BSDEs in this non‑concave utility setting.","kind":"theoretical","quote":"These numerical results show robustness of our methods outside of problem classes that are well understood. FIGURE 2. We present numerical results for different solution methods (HJB equations for primal, concavified primal, dual value functions, SMP for adjoint BSDE, reduced and full problems) and show that the DNN method is viable and robust in solving nonlinear PDEs and BSDEs.","support":"full"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"available","supports":[],"contradicts":[],"qualifies":[],"compared_with":33,"citations":{"cited_by_in_corpus":1,"by_independent_authors":1,"how_cited":{"used_as_baseline":0,"uses_method_or_data":1,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[{"paper":{"paper_id":"2510.21156","title":"Portfolio selection with exogenous and endogenous transaction costs under a two-factor stochastic volatility model","paper_date":"2025-10-24","authors":["Dong Yan","Ke Zhou","Zirun Wang","Xin-Jiang He"],"empirical_rigor":5,"page_url":"https://thequant.space/flowcharts/portfolio-selection-with-exogenous-and-endogenous-transaction-costs-under-a-two/"},"how":"uses_method_or_data","quote":null}],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":true,"code_url":"https://github.com/Ashley-Davey/ML-for-S-Utility","note":null},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":2,"extracted_at":"2026-11-14T19:54:42Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}