{"id":"720b7a32-1d6d-4038-a0fd-cfa502c154a1","arxiv_id":"2506.06653","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Risk attributions for ML financial models can be produced by applying the Shapley value to a risk measure of the model output, yielding a complete and symmetric allocation of risk to features.","lead":"This paper extends the Shapley value framework from explaining model predictions to explaining risk, by treating a risk measure of the model's output as the thing to allocate across features. It proposes 'BShap (RAM)' and shows on portfolios, factor models, and option pricing that the resulting risk attributions are plausible and satisfy basic fairness axioms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.3 is false as stated: the Symmetry (RAM) density condition does not force vR-symmetry, so the Shapley allocation need not treat symmetric features equally.","rationale":"The reader's conditional verdict correctly notes the arbitrary baseline x' as a weakness. That concern is real: attributions are defined relative to x', and the paper gives no principled rule for choosing it, so the meaning of a 'fair' allocation is partly convention. However, the more load-bearing defect is in the paper's main theoretical result. Theorem 3.3 is the axiomatic justification for calling BShap(RAM) fair, and its Symmetry claim is false as stated: the distributional condition in Definition 7 is insufficient to imply the characteristic-function symmetry required by the Shapley axioms. The counterexample above is simple and exact. The construction itself remains salvageable by strengthening the symmetry axiom or the proof; the baseline issue also needs a sensitivity analysis or selection rule. Because the central idea can be repaired with local corrections, a conditional verdict remains appropriate, but the condition must now explicitly include fixing Theorem 3.3 and Definition 7, not only the empirical baseline analysis.","tokens_in":12057,"tokens_out":15624,"duration_ms":161700,"concrete_test":"Compute Eq. (2) for m=2, f(x1,x2)=x1+2x2, X1,X2 iid N(0,1), x'=(0,0), rho=STD. The stated Symmetry (RAM) hypothesis holds, but vR({1})=1 and vR({2})=2, yielding BSR_1=(sqrt(5)-1)/2 and BSR_2=(sqrt(5)+1)/2, so the attributions are unequal. Also re-check the proof of Theorem 3.3: it needs vR(S∪{i})=vR(S∪{j}) for every S, which the density-of-f(X) condition does not imply.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central fairness guarantee is Theorem 3.3, which says BShap(RAM) preserves Symmetry (RAM). Definition 7 states Symmetry (RAM) as: if the density g of f(X) is invariant under swapping coordinates xi and xj, and x'_i = x'_j, then Ai = Aj. The proof in Appendix A asserts this implies Symmetry (SH) for the characteristic function vR(S) = rho(f(X_S; x'_{M\\S})). That implication is false. Shapley symmetry requires vR(S∪{i}) = vR(S∪{j}) for every coalition S avoiding i,j, a condition that depends on the functional form of f, not merely on the law of f(X). Concretely, take m=2, f(x1,x2)=x1+2x2, X1,X2 iid N(0,1), x'=(0,0), and rho=STD. The density of f(X) is N(0,5); after swapping arguments, f(X*) = X2+2X1 is also N(0,5), so the stated Symmetry (RAM) hypothesis is satisfied. Yet vR({1}) = STD(X1) = 1 while vR({2}) = STD(2X2) = 2, so vR is not symmetric and the Shapley attributions differ (BSR_1 = (sqrt(5)-1)/2, BSR_2 = (sqrt(5)+1)/2). Thus Theorem 3.3 is not correct as written; the theorem can be repaired by strengthening Symmetry (RAM) to require vR-symmetry (or f-symmetry plus exchangeability), but the current proof does not establish the central fairness claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes risk attribution methods (RAMs) for financial models by defining a characteristic function vR(S)=rho(f(X_S; x'_M\\S)) and applying the Shapley value. It claims that BShap(RAM) preserves completeness, dummy, and symmetry axioms, proves a sub-additivity bound for additive portfolios, and illustrates the method on portfolio CVaR, nonlinear factor models, and option pricing.","tokens_in":12410,"tokens_out":12335,"duration_ms":115387,"significance":"If the symmetry theorem were correct, the paper would offer a general and axiomatically motivated method for allocating model risk to features, with closed-form results for linear portfolios and a clear connection to the diversification principle. The paper is useful in framing the RAM problem and in showing that standard BAM axioms do not transfer directly to risk measures. The analytical examples and Shapley computations are carefully worked out. However, the central fairness theorem requires a substantive repair before the paper's main claim can be accepted, and the empirical sections are illustrative rather than statistically validated.","major_comments":[{"comment":"Theorem 3.3 is false as stated. The proof in Appendix A asserts that Symmetry (RAM) implies Symmetry (SH) for vR(S)=rho(f(X_S; x'_M\\S)), but the density condition in Definition 7 does not force vR(S∪i)=vR(S∪j) for every coalition S. Counterexample: let m=2, f(x1,x2)=x1+2x2, X1,X2 iid N(0,1), x'=(0,0), and rho=STD. The law of f(X) is N(0,5); after swapping inputs, f(X*) = X2+2X1 is also N(0,5), so the stated symmetry hypothesis is satisfied under the natural reading that the law of f(X) is invariant under coordinate swapping (it is also satisfied if g is read as the density of X, because X1 and X2 are exchangeable). Yet vR({1})=STD(X1)=1 while vR({2})=STD(2X2)=2, so vR is not symmetric and the Shapley attributions differ: BSR_1=(√5−1)/2 and BSR_2=(√5+1)/2. The theorem can be repaired by requiring f(x)=f(x*) for all x together with exchangeability of Xi and Xj, or by directly assuming vR-symmetry; as written, the central fairness guarantee is not established. This also affects Example 4.4, which relies on Symmetry (RAM) to conclude that X1 and X2 contribute equally to f.","section":"§3.2, Theorem 3.3 and Appendix A"},{"comment":"The choice of baseline constant x' is a load-bearing modeling input. In the portfolio example x'=0 is justified as a risk-free asset, but in the option example x'=(ST, sigma_T, r_T) is chosen as 'current market conditions' without a principled criterion. Since vR(S) replaces features by x'_M\\S, every attribution depends on x'; different baselines will generally produce different risk allocations. The paper does not analyze this sensitivity or provide a general rule for selecting the baseline, which limits the claim that the resulting allocation is 'fair' rather than merely axiomatic relative to an arbitrary reference point.","section":"§3.1 and §5.3"},{"comment":"The empirical sections present point estimates without error bars, repeated training runs, or any comparison to alternative risk allocation methods such as Euler's decomposition (which the paper discusses in §4.1) or Aumann–Shapley values. As a result, the statement that risk 'can be well allocated' is supported only by qualitative agreement with domain knowledge. If the empirical material is meant to demonstrate practical utility, it needs robust validation; otherwise it should be explicitly labeled as a toy illustration.","section":"§5 (Figures 1–4)"}],"minor_comments":[{"comment":"The notation g(x) for the density of f(X) is dimensionally inconsistent; rewrite the symmetry condition in terms of the law of f(X) or the density of X.","section":"Definition 7"},{"comment":"The phrase 'only for i = j' is ambiguous; specify whether it means the monotonicity axioms hold only when the two characteristic functions in the comparison involve the same player.","section":"Theorem 2.1"},{"comment":"'Symmetryic Monotonicity' should be 'Symmetric Monotonicity'.","section":"Section 4.1"},{"comment":"The text says log prices, log volatility, and log interest rates are used as inputs, but the baseline is given as x'=(ST, sigma_T, r_T); the baseline should be the logs of these values for consistency.","section":"Section 5.3"},{"comment":"Both BAM and RAM attributions appear to use the same symbol A_i; the claimed distinction is not visible in the text.","section":"Remark 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for Quantitative Finance, but the central theorem requires repair. I recommend major revision. The self-citations are frequent but not disqualifying. The paper would benefit from an explicit discussion of how the symmetry condition should be strengthened and from a sensitivity analysis over baselines. If the symmetry theorem cannot be repaired, the main fairness claim would not hold, and a rewrite focusing only on the definitional framework might be more appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chen applies Baseline Shapley to risk measures by defining v_R(S) = rho(f(X_S; x'_{M\\S})). That's a natural extension, and the discussion of which BAM axioms fail for RAMs (Linearity, Symmetric Monotonicity) is useful. The analytical examples are neat, and the sub-additivity proposition is correct as far as it goes.\n\nBut the central theorem is not. Theorem 3.3 claims BShap(RAM) preserves Symmetry (RAM). As stated, the axiom only requires the density of f(X) to be invariant under swapping coordinates. That doesn't imply the characteristic function v_R is symmetric. The stress-test counterexample is right: with f(x1,x2)=x1+2x2, X1,X2 iid standard normal, and x'=0, f(X) and f(X*) have the same distribution, yet v_R({1})=1 and v_R({2})=2. So the Shapley values differ. The proof in Appendix A just asserts the SH symmetry condition holds; it doesn't. The fix is to either require f itself symmetric in i,j plus exchangeability of X, or state Symmetry (RAM) directly in terms of v_R. Without that, the fairness guarantee isn't established.\n\nOther issues are more moderate. The baseline x' is ad hoc; for risk, choosing it as 'no randomness' is often not principled, and no sensitivity analysis is given. The empirical sections are illustrative: no error bars, no repeated training runs, and no comparison to Euler's decomposition—which the paper itself mentions and even shows satisfies the axioms for the two-asset case. The self-citations are not a concern.\n\nOn balance, the core idea is sound but incrementally novel; the Shapley machinery is well known. The paper would be worth a serious referee: the topic is practically important, and the flaw is repairable. A referee should demand a corrected symmetry axiom, a principled baseline discussion, and at least one real benchmark comparison.","headline":"Extends Shapley to risk measures, but the central symmetry theorem is false as stated; the idea is worth referee time if fixed.","tokens_in":12934,"tokens_out":3817,"would_cite":false,"duration_ms":37751,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A12","91B32","91G10","91G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Risk can be fairly attributed to features by running the Shapley value on a risk-measure characteristic function.","keywords":["risk attribution","Shapley value","explainable machine learning","risk measures","conditional value-at-risk","baseline attribution","portfolio risk allocation","nonlinear factor models"],"falsifier":"Compute BShap (RAM) for a fixed model and risk measure under two different baseline choices, such as $x'=(0,0)$ and $x'=(r,r)$ for a two-asset portfolio; if the attributed risks change materially, the method's allocation is convention-dependent rather than determined by the model and data.","tokens_in":11852,"feed_emoji":"📉","tokens_out":5604,"duration_ms":46025,"temperature":0.7,"pith_summary":"This paper tries to establish that risk, as measured by a risk measure such as standard deviation or conditional value-at-risk, can be fairly allocated across the input features of a general predictive model. The proposed method replaces the usual baseline explanation of a prediction with a Shapley value computed on characteristic functions that evaluate the risk of the model output when only a subset of features is random. If the paper is right, risk managers can decompose the risk of complex models, including neural networks, into per-feature contributions that respect completeness, dummy, and symmetry axioms. The authors demonstrate the allocation on optimized portfolios, nonlinear factor models, and an option-pricing example.","feed_headline":"The Shapley value now splits risk, not just predictions","feed_subtitle":"A game-theory tool that explains predictions also allocates portfolio, factor, and option risk fairly.","key_machinery":"The carrying object is the Shapley value formula, applied not to a model prediction but to a characteristic function built from a risk measure: $v_R(S) = \\rho(f(X_S; x'_{M\\setminus S}))$ for population data, and its sample analogue for realized data. Absent features are pinned to a baseline constant $x'$, so the characteristic function asks what the risk would be if only the features in $S$ were random. This single substitution converts the classical cooperative-game allocation into a risk allocation, and all axiomatic results follow from Shapley's original axioms applied to $v_R$. For one special case, Euler's decomposition gives an alternative closed-form split for homogeneous functions, but it is not linear in the characteristic-function sense and does not generalize; the Shapley route is proposed for general nonlinear models.","core_discovery":"The central claim is that defining $v_R(S) = \\rho(f(X_S; x'_{M\\setminus S}))$ and substituting it into the Shapley value formula allocates the risk $\\rho(f(X))$ among features in a way that is fair in the axiomatic sense. For this BShap (RAM) method, Theorem 3.3 shows that Completeness (RAM), Dummy (RAM), and Symmetry (RAM) hold. The paper also shows that Linearity cannot hold for RAMs together with completeness and dummy, and that symmetric monotonicity across different models is incompatible with the other axioms. Empirically, the method attributes the largest option risk to stock price, followed by volatility, and gives positive risk attributions to risky factors in a neural-network factor model, treating idiosyncratic error as an extra feature.","pith_inferences":["The method's output inherits the baseline-dependence of baseline Shapley: choosing different $x'$ can change every attribution, and the paper exercises ad hoc choices without a criterion; a principled baseline-selection rule would be needed for regulatory use.","Although the paper demonstrates CVaR and standard deviation, the construction applies to any risk measure $\\rho$, including expected shortfall, VaR, and drawdown-based measures; whether the axioms keep their force for non-coherent measures is a testable extension.","The Shapley value requires evaluating $2^m$ risk computations, so for high-dimensional factor models an approximation or sampling scheme would be needed before the method scales to production risk systems.","The symmetry axiom uses symmetry of the joint distribution of $f(X)$, which is stronger than symmetry of the model function alone; this suggests the method will behave differently from prediction explanations exactly when feature distributions are skewed or correlated."],"forward_implications":["For a linear portfolio with a sub-additive risk measure, BShap (RAM) assigns each asset an attribution no larger than its standalone risk, so diversification lowers attributed risk and the attributions sum to portfolio risk.","In nonlinear factor models, the method ranks factors by their risk contribution and can isolate idiosyncratic risk by treating the residual as an additional feature.","In option pricing, the method allocates the bulk of risk to the log stock price, then volatility, then the interest rate, suggesting which sources to hedge.","The axiomatic guarantees give a formal sense in which the allocation is fair, extending baseline Shapley explanations from predictions to risk.","Because Linearity is impossible for risk attributions, users should not expect additive feature decompositions across different models or risk measures."],"supporting_citations":[{"why":"Supplies the value formula and its characteristic-function axioms that the RAM construction relies on.","marker":"Shapley et al. (1953)"},{"why":"Defines the baseline Shapley (BShap) method and its axioms, which the paper extends from predictions to risk.","marker":"Sundararajan and Najmi (2020)"},{"why":"Supplies the sub-additivity property for coherent risk measures used to prove diversification under BShap (RAM).","marker":"Artzner et al. (1999)"},{"why":"Provides CVaR as the risk measure used in the portfolio and factor-model experiments.","marker":"Rockafellar and Uryasev (2002)"},{"why":"Underlies the option-pricing formula whose risk is decomposed in the empirical example.","marker":"Black and Scholes (1973)"},{"why":"Prior use of Shapley value for risk attribution in specific financial settings that the paper generalizes.","marker":"Tarashev et al. (2016)"},{"why":"Gives the analytical variance-allocations result for Shapley value that the portfolio theory section builds on.","marker":"Colini-Baldeschi et al. (2018)"},{"why":"Supplies the four fundamental axioms for baseline attribution methods that motivate the RAM axioms.","marker":"Lundstrom and Razaviyayn (2023a)"}],"fun_headline_variants":["Shapley value now slices risk, not just predictions","Fair risk splits via Shapley axioms","From value to risk: Shapley's new role","Risk attribution axiomatized for finance","Shapley goes beyond mean to risk sharing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The allocation is defined relative to a chosen baseline feature state $x'$, and the paper selects that baseline without a principled rule; if no natural no-randomness baseline exists, the claimed fairness of the split is not well defined.","fun_headline_variants_meta":{"raw":{"variants":["Shapley value now slices risk, not just predictions","Fair risk splits via Shapley axioms","From value to risk: Shapley's new role","Risk attribution axiomatized for finance","Shapley goes beyond mean to risk sharing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1442,"prompt_tokens":793,"completion_tokens":649,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":409,"completion_tokens_details":{"reasoning_tokens":577}},"tokens_in":409,"tokens_out":649,"duration_ms":6265,"temperature":1.0,"reasoning_tokens":577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:52:36.686159+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute BShap (RAM) for a fixed model and risk measure under two different baseline choices, such as $x'=(0,0)$ and $x'=(r,r)$ for a two-asset portfolio; if the attributed risks change materially, the method's allocation is convention-dependent rather than determined by the model and data.","supporting_citations":[{"cited_title":"The many shapley values for model explanation","cited_arxiv_id":null,"evidence_quote":"Defines the baseline Shapley (BShap) method and its axioms, which the paper extends from predictions to risk."},{"cited_title":"Coherent measures of risk","cited_arxiv_id":null,"evidence_quote":"Supplies the sub-additivity property for coherent risk measures used to prove diversification under BShap (RAM)."},{"cited_title":"The pricing of options and corporate liabilities","cited_arxiv_id":null,"evidence_quote":"Underlies the option-pricing formula whose risk is decomposed in the empirical example."},{"cited_title":"Risk attribution using the shapley value: Methodology and policy applications","cited_arxiv_id":null,"evidence_quote":"Prior use of Shapley value for risk attribution in specific financial settings that the paper generalizes."},{"cited_title":"Variance allocation and shapley value","cited_arxiv_id":null,"evidence_quote":"Gives the analytical variance-allocations result for Shapley value that the portfolio theory section builds on."}],"review_version":1}