{"id":"1bb0c2c2-9763-48ab-ad5d-519d2c44833c","arxiv_id":"1908.01602","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A single-objective deep learning algorithm for high-dimensional optimal stopping problems computes both approximate optimal exercise strategies and option prices, demonstrated on Bermudan max-call options in up to 5000 dimensions.","lead":"This paper introduces a deep learning algorithm that prices American and Bermudan options by learning an approximate optimal exercise strategy directly, and reports numerical results in dimensions up to 5000. It matters because it addresses the curse of dimensionality in optimal stopping, a core difficulty in quantitative finance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The d=5000 benchmark in §4.4.1.2 uses the algorithm's own M=6000 output as the reference value, so the reported relative errors measure training convergence, not accuracy; no independent upper bound supports the high-dimensional effectiveness claim.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption being an unproved expressiveness and trainability assumption for the neural-network architecture and the absence of convergence guarantees. My stress-test identifies a more specific, concrete manifestation of that concern: the signature 5000-dimensional numerical result is validated only against the algorithm's own output at a later training step, making the reported relative error a measure of training convergence rather than accuracy. This is a genuine load-bearing weakness for the paper's headline claim of high-dimensional effectiveness, because the example that best demonstrates 'many underlyings' lacks any independent reference value or dual upper bound. I therefore do not change the reader's conditional verdict: the paper remains a valuable numerical contribution whose central high-dimensional accuracy claim should be endorsed only after an independent validation, such as a dual upper bound or a comparison against an independent implementation. The reader's broader assumption about expressiveness/trainability is related but not identical; my concern is narrower and more directly tied to the specific experimental evidence, hence 'partial' agreement.","tokens_in":39177,"tokens_out":5690,"duration_ms":64314,"concrete_test":"For the d=5000 max-call example of §4.4.1.2, compute an upper-bound confidence interval using the Andersen–Broadie dual method, taking the trained neural-network stopping rule as the exercise policy. If the resulting 95% confidence interval is tight and contains 165.430 within, say, 1%, the reported value is credible; if the upper bound lies several percent above 165.430, the reported relative errors are optimistic and the high-dimensional effectiveness claim is not established. As an alternative or complement, run the independent deep-optimal-stopping algorithm from [9] with the same discretisation and compare its price estimate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that the algorithm is highly effective in high dimensions rests primarily on the d=5000 max-call example of §4.4.1.2. There, the reference value 165.430 is not an independent benchmark: the text states that 'the exact value of the price (120) has been replaced by the value 165.430, which corresponds to a realisation of P with M=6000' (Table 8). Consequently, the relative approximation errors in Table 8 and Figure 1 compare the algorithm only to a later run of itself; they demonstrate that training eventually stabilizes near one particular realisation, not that this realisation is close to the true optimal stopping value. Since (48) guarantees only that the reported price is a lower bound, and no dual upper bound is computed for d=5000, the numerical evidence cannot rule out a substantial low bias. The low-dimensional comparisons and the one-dimensional-representation examples provide useful validation, but the headline high-dimensional accuracy claim—and especially the 5000-dimensional result—remains unconfirmed by any independent reference. A secondary symptom of the same issue is visible in Table 8: the estimate at M=750 (103.764) is far worse than at M=500 (156.038), showing that training can temporarily stall in poor local optima, which further weakens the reliability of a single self-referential baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a deep-learning algorithm for high-dimensional optimal stopping problems arising in American and Bermudan option pricing. The method represents randomized stopping times through neural-network stopping-time factors, maximizes the expected discounted payoff over network parameters by stochastic gradient ascent (Adam), and then converts the trained randomized rule into a true stopping time whose payoff is evaluated by Monte Carlo. The authors prove a factorisation lemma for discrete stopping times (Lemma 2.2), observe that the final price estimator is low-biased (inequality (48)), and report numerical experiments for Brownian-motion stopping, geometric-average options, Bermudan max-call options up to 5000 dimensions, basket options, and a path-dependent derivative. Most lower-dimensional results are compared with binomial-tree or literature reference values.","tokens_in":39452,"tokens_out":4712,"duration_ms":45312,"significance":"If the numerical claims hold, this is a useful contribution: it offers a single-objective, non-recursive training formulation, avoids the temporal recursion of earlier deep optimal stopping algorithms, and shows empirically stable pricing in hundreds of dimensions. The theoretical lemmas are clearly proved, and the low-bias caveat is stated explicitly. The lower-dimensional comparisons against Longstaff-Schwartz, Andersen-Broadie, Broadie-Cao, and other literature benchmarks are genuine strengths. However, the headline high-dimensional claim rests on the d=5000 max-call experiment whose reference value is generated by the algorithm itself; without an independent upper bound or benchmark, the accuracy in high dimension is not established.","major_comments":[{"comment":"The reference value 165.430 is not independent: the text states that the exact value of the price (120) has been replaced by a realisation of P with M=6000. Consequently, the relative errors in Table 8 and Figure 1 compare the algorithm to another run of itself and measure training stability, not proximity to the true optimal stopping value. Moreover, the 95% confidence interval for that reference is [165.378, 165.483], so the reported errors below 10^-3 are within the Monte Carlo noise of the reference. The row M=750 (103.764) also shows that training can stall in poor local optima, so a single self-referential baseline cannot certify reliability. This is load-bearing because the abstract and introduction claim effectiveness in 5000 dimensions.","section":"§4.4.1.2, Table 8"},{"comment":"Since the paper only establishes a lower bound and no dual upper bound or other independent bound is provided for d=5000, the numerical evidence cannot rule out a substantial low bias. The authors should either compute a dual upper bound (for example, an Andersen-Broadie type bound) for the max-call example or clearly restrict the high-dimensional effectiveness claim to the validated lower-dimensional cases. This is fixable, but as it stands the central high-dimensional claim is not fully supported.","section":"§2.7, Eq. (48)"},{"comment":"The approximation power of the network class is an assumption rather than a theorem: no expressiveness or trainability result is given for the two-hidden-layer architecture of width d with the recursive factorisation (60). This is acceptable for a numerical paper only if independent benchmarks carry the validation. Since the main high-dimensional benchmark is self-referential, the high-dimensional claim remains conjectural. A convergence or approximation result, even for a simplified setting, or a broader set of independently benchmarked high-dimensional examples would strengthen the paper considerably.","section":"§2.4 and §3.2, Eq. (60)"}],"minor_comments":[{"comment":"The sentence 'the runtime in seconds need or calculating the realisation of P' contains a typo; it should read 'needed for calculating'. In addition, J0 appears as '2 20' and should be typeset as 2^20.","section":"§4.4.1.2, text after Table 8"},{"comment":"The phrase 'under suitable hyptheses' contains a typo; it should read 'under suitable hypotheses'.","section":"§2.6"},{"comment":"The table lists no reference values for d=10 to 500. The caption should state explicitly that these rows are unbenchmarked, so that readers do not mistake them for validated accuracy claims.","section":"§4.4.1.1, Table 7"},{"comment":"The formula 'Itˆ o's formula' has a formatting artifact in the proof; it should be 'Itô's formula'. The same artifact appears in the text before equation (81).","section":"§4.1, Proposition 4.3 proof"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper fits the scope of the journal and the authors' citation practice appears fair. The main concern is the self-referential benchmark in §4.4.1.2, which is decisive for the high-dimensional claim. I would not recommend rejection because the flaw is local and repairable: replacing the d=5000 reference with an independent method, or removing that claim, would resolve it. The lower-dimensional results are credible and useful."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a genuinely useful reformulation of the deep optimal stopping approach in Becker–Cheridito–Jentzen. Instead of solving a backward recursion, they optimize a single objective over randomized stopping times, then convert the relaxed factors into a true stopping time via (44). The factorisation lemma (Lemma 2.2) is correct and clean, and the low-bias caveat is stated plainly in (48). I came away with a good opinion of the algorithmic idea and the exposition.\n\nWhat is new and good: the single-objective construction is a real simplification, not a repackaging. The numerical results for standard Bermudan max-call benchmarks in 2, 3, and 5 dimensions match the Andersen–Broadie and Broadie–Cao confidence intervals well. The geometric-average examples are validated against exact one-dimensional representations from Proposition 4.3; those are genuine checks, since the distributional symmetry is mathematically exact. Runtime scaling up to a few hundred dimensions looks plausible.\n\nThe soft spot, and it is a real one: the d=5000 max-call benchmark is not independent. Table 8 replaces the exact price (120) by 165.430, which is itself a realisation of the same algorithm at M=6000. So the relative errors in Figure 1 and Table 8 measure training convergence, not accuracy against any external truth. Inequality (48) guarantees only a lower bound, and no dual upper bound is computed for d=5000, so the evidence cannot rule out a substantial low bias. The non-monotonic training path (M=750 at 103.764 after M=500 at 156.038) reinforces that a single run can stall in a poor local optimum. The authors disclose the self-reference, so this is not a hidden flaw, but the advertised claim of “highly effective in 5000 dimensions” is stronger than the evidence supports.\n\nA second, smaller concern is the unproven expressiveness assumption: two hidden layers of width d, plus the recursive factorisation in (60), may not represent optimal stopping times for all payoff structures. That is typical for this literature and is mitigated by the breadth of examples, so I treat it as minor.\n\nWho is this for? Researchers in quantitative finance and machine learning who want a practical, scalable algorithm for high-dimensional early-exercise problems. It deserves a serious referee, not a desk reject. My recommendation: send it to peer review, and ask the authors to either (a) compute dual upper bounds for the high-dimensional cases, (b) release the code, or (c) at minimum re-label the d=5000 comparison as a training consistency check rather than an independent accuracy result.","headline":"A solid and honest deep-optimal-stopping paper whose low-dimensional validation is convincing, but whose 5000-dimensional headline number is self-referential and should not be read as an independent accuracy claim.","tokens_in":39954,"tokens_out":1861,"would_cite":true,"duration_ms":21553,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G40","65C05","91G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep learning algorithm prices high-dimensional American and Bermudan options by training neural networks to represent stopping times directly.","keywords":["optimal stopping","American options","Bermudan options","deep learning","curse of dimensionality","neural networks","stopping time factorisation","stochastic gradient ascent"],"falsifier":"Run the same algorithm on a 100-asset payoff whose optimal exercise region is fragmented into many disjoint zones, for example a basket payoff with several separate in-the-money bands, and compare the low-biased estimate with a dual upper bound; if the gap exceeds the Monte Carlo error, the chosen network architecture is too small to represent the optimal stopping time.","tokens_in":38981,"feed_emoji":"📈","tokens_out":4292,"duration_ms":45853,"temperature":0.7,"pith_summary":"The paper proposes a deep learning algorithm that turns a high-dimensional optimal stopping problem, such as pricing an American or Bermudan option on many underlying assets, into a single neural network training problem. The central move is to represent a discrete stopping time through measurable factors that decide, at each time step, whether to stop, and to relax those binary factors into smooth network outputs. Training then maximizes the expected discounted payoff by stochastic gradient ascent, producing both an approximate optimal exercise strategy and a low-biased estimate of the option price. The numerical results suggest the method can handle problems with hundreds or thousands of assets, including a Bermudan max-call option in 5000 dimensions with a reported price of about 165.430 and a runtime of about 2899 seconds. If the method holds up beyond the tested examples, it would offer a practical route around the curse of dimensionality for early-exercise derivatives.","feed_headline":"Neural nets price 5,000-asset options in under an hour","feed_subtitle":"Training a network to represent the exercise decision prices Bermudan options on 5,000 underlyings without grids.","key_machinery":"The load-bearing identity is the factorisation lemma for stopping times (Lemma 2.2), which expresses any discrete stopping time as a sum of measurable functions of the observed path, together with the recursive construction of approximate stopping-time factors $U_{n,\\theta}$ that always sum to one. The paper pairs this with a fixed feedforward architecture: one network per time step, with two hidden layers of width equal to the problem dimension, logistic output activation, batch normalisation, and the Adam optimiser. This machinery converts optimal stopping into an unconstrained parameter search and yields the explicit exercise rule that stops at the first time the current stopping factor outweighs all later factors.","core_discovery":"The paper's central claim is that the supremum over all stopping times in a discrete-time optimal stopping problem can be approximated by a supremum over neural network parameters. This is made possible by a factorisation lemma: any stopping time adapted to a Markov process can be written as a sum over time indices of indicator functions of the observed path, and these indicators can be relaxed into neural-network outputs that define a randomised stopping time. The resulting objective function, an expectation of the payoff weighted by these randomised stopping factors, is maximized by stochastic gradient ascent; a final threshold rule converts the trained network into a genuine stopping time. The paper reports that this single-training-pass procedure matches reference values across many benchmark problems, including American puts, geometric average options, max-call options, basket options under local volatility, and a path-dependent derivative recast as a 100-dimensional problem.","pith_inferences":["Beyond the paper: the factorisation construction could be reused for related discrete decision problems, such as optimal stopping with multiple exercise rights or swing options, by replacing the single stopping indicator with a sequence of exercise decisions.","Beyond the paper: pairing this low-biased primal estimate with a dual upper bound method would turn the reported point estimates into proper confidence intervals, which is a natural next step the paper does not pursue.","Beyond the paper: a testable extension is to apply the same architecture to payoffs with fragmented or discontinuous exercise regions, such as barrier-like payoffs, where the smooth network representation may need more layers or width to stay accurate."],"forward_implications":["Bermudan and American options on hundreds or thousands of underlying assets can be priced without any grid or mesh in the asset dimension, because the networks scale with the dimension of the state rather than the size of a state space.","The method outputs an approximate exercise strategy, not only a price, so it can be used to guide early-exercise decisions in practice.","The same formulation applies to any simulable Markov process, including path-dependent derivatives once they are embedded in a higher-dimensional Markovian state.","Because the final Monte Carlo estimate uses a genuine stopping time, the reported price is a low-biased lower bound; reference values are needed to judge how close it is to the true price.","The 5000-dimensional max-call result suggests that many high-dimensional exercise problems have enough low-dimensional structure for a moderately sized network to capture."],"supporting_citations":[{"why":"Provides the earlier recursive deep optimal stopping approach that this paper contrasts and replaces with a single training pass over one objective.","marker":"[9]"},{"why":"Supplies the classic least-squares regression baseline and the American put reference value used for relative error comparisons.","marker":"[73]"},{"why":"Supplies the binomial reference values and confidence intervals used to validate the Bermudan max-call benchmark.","marker":"[3]"},{"why":"Supplies the correlated max-call example with quoted prices that are compared in the numerical section.","marker":"[5]"},{"why":"Supplies the path-dependent derivative example and the lower bound used to assess the 100-dimensional pricing results.","marker":"[88]"},{"why":"Defines the Adam optimiser used in every numerical experiment.","marker":"[58]"},{"why":"Supplies a neural-network pricing baseline and the American strangle spread basket example.","marker":"[64]"},{"why":"Provides the binomial-tree reference values used to compute relative approximation errors for several examples.","marker":"[85]"}],"fun_headline_variants":["Deep learning cracks 5000-dim optimal stopping","Neural nets beat curse of dimensionality for 5000-asset options","5000 underlyings, one neural net: optimal stopping solved","Neural approach prices 5000-dimension Bermudan options"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a feedforward network with two hidden layers, each as wide as the number of underlying assets, can represent the optimal stopping rule closely enough for the tested payoff structures and dynamics; this is supported by numerical examples but not by a convergence proof.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning cracks 5000-dim optimal stopping","Neural nets beat curse of dimensionality for 5000-asset options","5000 underlyings, one neural net: optimal stopping solved","Neural approach prices 5000-dimension Bermudan options"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001203,"raw_usage":{"total_tokens":4948,"prompt_tokens":929,"completion_tokens":4019,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":3947}},"tokens_in":545,"tokens_out":4019,"duration_ms":29056,"temperature":1.0,"reasoning_tokens":3947,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:06:43.417502+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same algorithm on a 100-asset payoff whose optimal exercise region is fragmented into many disjoint zones, for example a basket payoff with several separate in-the-money bands, and compare the low-biased estimate with a dual upper bound; if the gap exceeds the Monte Carlo error, the chosen network architecture is too small to represent the optimal stopping time.","supporting_citations":[{"cited_title":"Deep optimal stopping","cited_arxiv_id":"1804.05394","evidence_quote":"Provides the earlier recursive deep optimal stopping approach that this paper contrasts and replaces with a single training pass over one objective."},{"cited_title":"A., and Schwartz, E","cited_arxiv_id":null,"evidence_quote":"Supplies the classic least-squares regression baseline and the American put reference value used for relative error comparisons."},{"cited_title":"N., and Van Roy, B","cited_arxiv_id":null,"evidence_quote":"Supplies the path-dependent derivative example and the lower bound used to assess the 100-dimensional pricing results."},{"cited_title":"Adam: a method for stochastic optimization","cited_arxiv_id":null,"evidence_quote":"Defines the Adam optimiser used in every numerical experiment."},{"cited_title":"Pricing of high-dimensional American options by neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies a neural-network pricing baseline and the American strangle spread basket example."},{"cited_title":"Javascript options and implied volatility calculator","cited_arxiv_id":null,"evidence_quote":"Provides the binomial-tree reference values used to compute relative approximation errors for several examples."}],"review_version":1}