{"paper_id":"2601.03175","title":"Breaking the Dimensional Barrier: Dynamic Portfolio Choice with Parameter Uncertainty via Pontryagin Projection","authors":["Jeonggyu Huh","Hyeng Keun Koo"],"paper_date":"2026-01-06","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2601.03175v1","page_url":"https://thequant.space/flowcharts/breaking-the-dimensional-barrier-dynamic-portfolio-choice-w/","scores":{"empirical_rigor":4,"math_complexity":9,"quadrant":"Lab Rats"},"proposes":{"status":"available","statement":"The paper proposes a simulation‑only two‑stage solver called PG‑DPO, consisting of Stage 1 (DPO) that performs BPTT‑based stochastic gradient ascent on an Euler simulator, and Stage 2 (Pontryagin projection) that aggregates costate blocks across the latent parameter distribution and enforces the q‑a","quote":"We propose a simulation-only two-stage solver.Stage 1 (DPO) performs BPTT-based stochastic gradient ascent through an Euler simulator while sampling $θ$ only inside the simulator. Stage 2 (Pontryagin projection) aggregates costate blocks across $θ\\sim q$ and enforces the $q$-aggregated stationarity condition within the deployable class; the resulting correction can be amortized via interactive distillation.","summary":"We study continuous-time portfolio choice in diffusion markets with parameter θ∈ Θ and uncertainty law q(dθ). Nature draws latent θ~ q at time 0; the investor cannot observe it and must deploy a single θ-blind feedback policy maximizing an ex-ante CRRA objective averaged over diffusion noise and θ.…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Nature draws a latent parameter θ from a fixed distribution q at time 0 and keeps it constant over the investment horizon.","type":"model","stated":"explicit","quote":"Nature draws a latent parameter $θ\\sim q$ at time $0$ and keeps it fixed; the investor never observes $θ$ and must commit to a single $θ$-blind policy maximizing an ex-ante objective, treating $q$ as a decision-time input.","support":"full"},{"assumption":"The investor never observes θ and must commit to a single θ‑blind Markov feedback policy evaluated under an ex‑ante objective.","type":"model","stated":"explicit","quote":"Nature draws a latent parameter $θ\\sim q$ at time $0$ and keeps it fixed; the investor never observes $θ$ and must commit to a single $θ$-blind policy maximizing an ex-ante objective, treating $q$ as a decision-time input. The deployed controller is θ-blind: it depends only on observable states (t, Xt, Yt) and never conditions on the realized θ.","support":"full"},{"assumption":"Asset prices follow a continuous‑time diffusion with constant risk‑free rate r and stochastic excess returns driven by Brownian motions.","type":"market","stated":"explicit","quote":"Let d risky assets satisfy, for s ∈[t, T], dSs Ss = r 1 ds + θ ds + Σ1/2dWs, St ∈(0, ∞)d, (24) where Σ ∈Rd×d is symmetric positive definite and θ ∈Rd is a latent constant excess-return vector.","support":"full"},{"assumption":"Utility is CRRA with risk‑aversion parameter γ>1.","type":"model","stated":"explicit","quote":"To obtain a transparent closed-form target, we restrict to portfolio fractions constant over [t, T]: πs ≡π ∈Rd, s ∈[t, T], (28) and take CRRA utility U(x) = x1−γ/(1 −γ) with γ > 1.","support":"full"},{"assumption":"The decision‑time input law q (or qt) is treated as an ex‑ante fixed description and is not updated during the investment horizon.","type":"statistical","stated":"explicit","quote":"Fix a decision time t ∈[0, T) and let an external routine provide an Fobs t -measurable input law qt for the uncertain premium parameter/state. We treat qt as a decision-time input (not a belief process controlled/updated within a run): it is held fixed over the remaining horizon [t, T].","support":"full"},{"assumption":"The policy class is deployable and Y‑blind, i.e., it depends only on observable wealth and time‑to‑go and not on latent factor realizations.","type":"execution","stated":"explicit","quote":"Throughout this section, all reported policies are deployable and Y -blind: the control is a function of observ-able wealth and time-to-go only, and never takes as input (i) a point estimate of Y0, (ii) the realized initial premium Y0, or (iii) the factor path (Yt) (including the PPO baseline). At de-ployment, the investor is assumed to have access only to an uncertainty description q0(dy) for 26 Y0","support":"full"}]},"datasets":{"status":"available","items":[{"name":"High‑dimensional Gaussian drift‑uncertainty benchmark","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":"d risky assets (d∈{5,10,50,100})","quote":"Section 2.3.1 considers static drift uncertainty and provides the benchmark used in the high-dimensional scaling/geometry experiments of Section 4. Let d risky assets satisfy, for s ∈[t, T], dSs Ss = r 1 ds + θ ds + Σ1/2dWs, St ∈(0, ∞)d, (24) where Σ ∈Rd×d is symmetric positive definite and θ ∈Rd is a latent constant excess-return vector.","support":"full"},{"name":"OU premium factor benchmark with Gaussian decision‑time uncertainty","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":"d risky assets with factor Y_t","quote":"Sec-tion 2.3.2 considers a mean-reverting (OU) premium factor with Gaussian decision-time un-certainty for the current factor level; OU propagation and time-averaging induce a horizon-dependent effective Gaussian premium law, yielding a tractable reference used in the hedging-demand recovery study of Section 5. Let Yt ∈Rm be a mean-reverting premium factor and Rt ∈Rd the risky excess returns: dYt = κY (¯y −Yt) dt +","support":"full"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"synthetic"},"transaction_costs":{"status":"available","modeled":"no","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R1","candidates":{"front_matter":"no","evidence_card":"no","evidence_card_quoted":false},"meaning":"a trading or execution result is reported without costs"},"results":{"status":"available","evaluation_type":"unclear","findings":[{"finding":"The Pontryagin projection step (Stage 2) substantially reduces decision‑time RMSE compared to the warm‑up DPO policy across all dimensions and uncertainty levels.","kind":"performance","quote":"Across all d and s0, the post-hoc Pontryagin projection (PG–DPO) substantially reduces decision-time RMSE relative to the deployable Stage 1 warm-up (DPO) policy (Table 2).","support":"full"},{"finding":"In a high‑dimensional setting (d=100, aligned P0, s0=10⁻³), DPO RMSE is 3.54×10⁻³ while PG‑DPO achieves 1.56×10⁻⁴.","kind":"performance","quote":"For example, under aligned P0 with s0 = 10−3 and d = 100, DPO attains 3.54 × 10−3 whereas PG–DPO achieves 1.56 × 10−4.","support":"full"},{"finding":"Interactive distillation further improves the deployable policy but the projected PG‑DPO rule remains the most accurate.","kind":"performance","quote":"Amortization, robustness, and the PPO baseline. Interactive distillation improves the deployable Stage 1 policy relative to basic Stage 1 (DPO), while the most accurate object remains the post-hoc projected policy (PG–DPO) (Table 2).","support":"full"},{"finding":"The PPO baseline remains far from the analytic OU reference under the same Y‑blind restriction and can be unstable.","kind":"negative","quote":"Finally, the PPO baseline remains far from the analytic OU reference under the same Y -blind deployability restriction, and can be unstable in this terminal-only setting (Table 2).","support":"full"},{"finding":"Projection also recovers the intertemporal hedging demand, with hedging‑component RMSE dropping to ~1.5×10⁻⁴ in the same setting.","kind":"performance","quote":"The component-wise diagnostics indicate that the remaining discrepancy is largely driven by the hedging channel: in the same setting, the hedging RMSE is 1.55 × 10−4 (basic) and 1.42 × 10−4 (distill.) (Table 3), while the myopic RMSE is an order of magnitude smaller (Appendix Table 4).","support":"full"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":18,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"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":"2026-11-14T19:44:31Z"},"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/"}