{"id":"70272093-c625-4d1a-94f0-2fd70d6e6ec5","arxiv_id":"1908.02164","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A stochastic-control framework yields optimal statistical arbitrage portfolios for multiple co-integrated stocks, with stability guarantees for risk-averse investors and backtests showing high sensitivity to parameter estimation.","lead":"This paper solves optimal portfolio choice for multiple stocks whose prices are co-integrated with eigenportfolio factors, with and without a market-neutrality constraint. The authors derive stability conditions for the optimal strategies and backtest them on S&P 500 data from 2000 to 2021, finding that performance is sensitive to parameter estimation and stronger in high-volatility periods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 2.8's rank condition (2.42) is algebraically impossible: (I-P)δ has rank d-m, so the constrained Riccati observability condition can never hold for m>0.","rationale":"The reader identified the rank condition (2.42) as the weakest assumption, noting the lack of justification for real data. Close inspection shows the condition is not merely unjustified but algebraically impossible: (I−P)δ necessarily has rank d−m, so it cannot equal d when m≥1. This makes Proposition 2.8 vacuous and removes the theoretical support for the constrained portfolio's boundedness and stable growth, which is a central part of the paper's claimed contribution. The unconstrained results may still be salvageable, but the paper as written contains a fundamental flaw in the constrained stability analysis. This warrants rejection or major revision rather than a conditional acceptance conditioned on a plausible assumption.","tokens_in":40633,"tokens_out":11650,"duration_ms":116415,"concrete_test":"Analytic check: for any d>m, full-column-rank β∈R^{d×m}, and invertible diagonal δ, compute rank(δ−β(β^Tβ)^{-1}β^Tδ) symbolically; it equals d−m < d. Numerical confirmation: set d=15, m=6, take β as a random 15×6 matrix and δ=diag(1,2,...,15); compute the rank of the observability matrix of (E_c,A_c) using the definitions in Proposition 2.8. The rank will be 9, not 15, confirming that (E_c,A_c) is unobservable and Proposition 2.8's conditions are impossible to satisfy.","verdict_should_be":"REJECT","load_bearing_attack":"The constrained stability theorem rests on Proposition 2.8, which requires rank(δ−β(β^Tβ)^{-1}β^Tδ)=d and rank(β^Tβ)=m. Let P=β(β^Tβ)^{-1}β^T. Since P is the orthogonal projection onto Ran(β) and has rank m, I−P has rank d−m. Because δ is diagonal with strictly positive entries, it is invertible, so rank((I−P)δ)=rank(I−P)=d−m. Thus the first equation in (2.42) holds only if m=0, contradicting the setup m≥1. Equivalently, from Proposition 2.7, −Pc=E_c^T E_c has null space spanned by δ^{-1}β, whose dimension is m, so E_c has rank d−m. Hence the observability matrix [E_c; E_c A_c; ...; E_c A_c^{d−1}] has rank at most d−m<d, and the pair (E_c,A_c) is never observable. Therefore Wonham's theorem cannot be invoked, and the claimed boundedness and stable growth for the market-neutral constrained portfolio have no proof. This is not a data-dependent issue: the hypotheses are empty for every nontrivial factor model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops optimal statistical arbitrage strategies for multiple co-integrated stocks whose spreads are driven by untradeable eigenportfolio factors. The authors derive HJB equations for both an unconstrained portfolio and a market-neutral constrained portfolio, reduce them via an exponential ansatz to a matrix Riccati ODE plus two linear ODEs, and claim long-term stability and absence of arbitrage for risk aversion gamma<0. They then present sliding-window backtests on S&P 500 constituents from 2000 to 2021, reporting that the strategies are sensitive to parameter estimation and more profitable in high-volatility periods.","tokens_in":40947,"tokens_out":14268,"duration_ms":149796,"significance":"If the theoretical claims were valid, the paper would provide a multivariate extension of Avellaneda-Lee co-integration trading with explicit optimal controls and sufficient conditions for long-term stable growth in both unconstrained and market-neutral settings. The unconstrained HJB reduction and the use of Wonham's theorem for the unconstrained Riccati equation are standard and mostly correct, and the empirical study is transparent about data cleaning, eigenportfolio construction, survivorship adjustment, and hyperparameter sensitivity. The paper also honestly acknowledges transaction costs, liquidity, and practical implementation issues. However, the constrained market-neutral stability theorem rests on an impossible rank condition, so the paper's central theoretical contribution for the constrained portfolio is not established as written.","major_comments":[{"comment":"The first full-rank condition in (2.42) is impossible for any m >= 1. Since beta(beta^T beta)^{-1} beta^T is the orthogonal projection onto the column space of beta, it has rank m, so I - beta(beta^T beta)^{-1} beta^T has rank d - m. Because delta is diagonal and invertible (delta_i > 0 by the OU stationarity assumption), rank((I - beta(beta^T beta)^{-1} beta^T) delta) = rank(I - beta(beta^T beta)^{-1} beta^T) = d - m. Hence rank(delta - beta(beta^T beta)^{-1} beta^T delta) = d requires m = 0, contradicting the factor model with m >= 1. Consequently Proposition 2.8 is vacuous, and the boundedness, observability, and convergence conclusions for the constrained Riccati equation are unproved for every nontrivial factor model. The necessary condition stated in Remark 2.7 is also inaccurate: non-commutation of the projection with delta does not change the rank of (I - P)delta.","section":"§2.4, Proposition 2.8 and Eq. (2.42)"},{"comment":"The conclusion that R_c(t) has all positive eigenvalues because it is 'the summations and produces of positive semi-definite matrices' is not a valid inference. Products of positive semi-definite matrices are not generally symmetric positive semi-definite, and even when each term has nonnegative eigenvalues, the eigenvalues of a sum of non-normal matrices cannot be inferred from the eigenvalues of the summands. A separate argument is needed. Since this proposition underpins the finite steady state of b(t) and hence the constrained long-term growth claim, the constrained stability analysis is not established even if the issues with Proposition 2.8 were resolved.","section":"§2.4, Proposition 2.9"},{"comment":"A similar gap appears in the unconstrained analysis. The proof that R_u(t) has all positive eigenvalues relies on the assertion that the terms are sums and products of positive semi-definite matrices and on Lemma 2.1. This is not sufficient: -C(t) Q_u is a product of two positive semi-definite matrices and is not symmetric, and Lemma 2.1(a) concerns the field of values rather than eigenvalue positivity of a sum. Since Proposition 2.5 uses the finite steady state of b(t) to obtain the long-term growth rate, the unconstrained long-term stability claim also needs a corrected proof.","section":"§2.3, Proposition 2.4"}],"minor_comments":[{"comment":"The myopic control for the constrained portfolio is stated as pi_m^* = (1/(1-gamma))(Sigma_1^{-1} - Sigma_c)(mu + delta z). Substituting C = b = 0 into Eq. (2.36) gives the same expression with mu - delta z, matching the unconstrained myopic control in Eq. (2.21). The plus sign appears to be a typo; please correct it and verify which version was used in the backtests.","section":"§2.4, Eq. (2.37)"},{"comment":"In the wealth dynamics display, the risk-free contribution is written with an extra factor W_t after dividing by W_t; it should be r(1 - pi^T 1) dt, not r(1 - pi^T 1) W_t dt.","section":"§2.4, Eq. (2.29)"},{"comment":"The proof contains notational slips: after defining A_c, the text refers to Gamma(A_u^T, E_u^T) and to (A_u^T, E_u^T) instead of A_c in the controllability/observability discussion. These should be corrected.","section":"§2.4, Proof of Proposition 2.8"}],"recommendation":"major_revision","confidential_remarks":"The vacuous rank condition in Proposition 2.8 is a serious mathematical error that should not pass peer review. The paper's unconstrained theory and empirical framework have value, so I do not recommend outright rejection, but the authors must either provide a valid stability proof for the constrained market-neutral case or substantially revise their claims about constrained long-term stable growth. The backtests alone do not establish the theoretical assertions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the unconstrained half is a competent extension of Avellaneda–Lee to multi-stock stochastic control, but the constrained half has a load-bearing gap. Proposition 2.8's rank condition can never hold when there is at least one factor, so the paper's proof of boundedness and stable growth for the market-neutral portfolio is unsupported. This isn't a minor technicality; it's the theorem that the constrained stability claims rest on.\n\nWhat the paper does well: it takes the eigenportfolio factor model seriously, treats factors as non-tradeable, derives the HJB equations for power utility, reduces them to a matrix Riccati system, and gives a clean Wonham-based stability argument for the unconstrained case. That derivation is standard but carefully done, and the unconstrained stability result appears correct. The empirical section is also honest: it flags survivorship bias, shows sensitivity to estimation windows, and explicitly says transaction costs are not included. The backtest setup is reproducible in spirit and the high-volatility profitability claim is at least plausible.\n\nWhere it falls down: the constrained stability analysis. In Proposition 2.8 the paper requires rank(δ − β(βᵀβ)⁻¹βᵀδ) = d. Since δ is diagonal and invertible and β is d×m full rank, that matrix equals (I − P)δ where P is the orthogonal projection onto the column space of β. Its rank is d−m, not d. So for m ≥ 1 the hypothesis is empty. The paper's Remark 2.7 says a necessary condition is δ not proportional to identity, which is true but far from sufficient. As a result the observability condition needed for Wonham's theorem is never satisfied, and the claimed boundedness for the constrained Riccati equation has no proof. Proposition 2.9 has a separate gap: it asserts that sums and products of positive semi-definite matrices have positive eigenvalues, which is not generally true. Minor issues include a sign typo in equation (2.37), where the constrained myopic control should read µ − δz, and backtests that lack transaction costs and statistical significance tests—though the authors acknowledge the first omission.\n\nWho should read this: anyone working on stochastic control for statistical arbitrage, especially with multiple co-integrated assets and non-tradeable factors. The unconstrained part and the empirical framework are worth engaging with. But the market-neutral part needs real rework, likely a different argument than the rank condition or a reformulated constraint.\n\nRecommendation: send to peer review, but with the expectation of major revision. The unconstrained contribution is real; the constrained theorem must be fixed or substantially revised before the paper's central claims can be accepted.","headline":"The unconstrained HJB solution is a solid, useful extension of the Avellaneda–Lee model, but the market-neutral stability proof rests on a rank condition that is algebraically impossible for any nontrivial factor count, so the constrained claims currently lack proof.","tokens_in":41451,"tokens_out":2850,"would_cite":false,"duration_ms":33069,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P05","91B28","93E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"For risk-averse investors, optimal multi-stock arbitrage portfolios are bounded and stable, and backtests show they profit most in volatile markets.","keywords":["statistical arbitrage","co-integrated stocks","eigenportfolios","Hamilton-Jacobi-Bellman equation","matrix Riccati equation","market-neutral portfolio","Ornstein-Uhlenbeck process","stochastic optimal control"],"falsifier":"Estimate $\\beta$ and $\\delta$ on any sliding-window training set of S&P 500 returns and check whether $\\beta^\\top\\beta$ has full rank and whether $\\delta-\\beta(\\beta^\\top\\beta)^{-1}\\beta^\\top\\delta$ has rank $d$; if either fails, the constrained portfolio's boundedness theorem does not apply, and the Riccati ODE can be solved numerically to see whether $C(t)$ remains bounded or blows up.","tokens_in":40427,"feed_emoji":"📈","tokens_out":10592,"duration_ms":92567,"temperature":0.7,"pith_summary":"This paper claims that optimal statistical arbitrage among multiple co-integrated stocks, driven by eigenportfolios as untradeable factors, is mathematically well-behaved for risk-averse investors: with risk-aversion parameter $\\gamma<0$, the value functions of the stochastic control problems stay finite and optimal portfolios converge to steady states as the horizon grows. The authors solve the resulting Hamilton-Jacobi-Bellman equations through an exponential ansatz, reducing each problem to a matrix Riccati equation and two linear ordinary differential equations, and prove boundedness and stability for both an unconstrained portfolio and one forced to be market-neutral. In sliding-window backtests on S&P 500 constituents from 2000 through 2021, the strategies are reported to be sensitive to parameter estimation and to perform best when overall market volatility is high. If the stability result is right, it gives a theoretical reason why multi-asset mean-reversion strategies can grow steadily without admitting an arbitrage that would make the value function blow up.","feed_headline":"Multi-stock arbitrage portfolios stay stable for risk-averse traders","feed_subtitle":"Proved for unconstrained and market-neutral HJB solutions; backtests favor high-volatility periods.","key_machinery":"The load-bearing object is the exponential ansatz $g(t,z)=\\exp(a(t)+b(t)^\\top z+z^\\top C(t)z)$, which converts the nonlinear HJB PDE into a system of ODEs: a matrix Riccati equation for $C(t)$, a linear ODE for $b(t)$, and a scalar ODE for $a(t)$. Stability of the whole system is governed by the Riccati equation; the proof applies Wonham's existence-and-boundedness theorem for matrix Riccati equations, whose hypotheses for $\\gamma<0$ reduce to the coefficient matrix $Q$ being symmetric positive definite and $-P$ being symmetric positive (semi)definite. For the constrained portfolio the same structure appears with $\\Sigma_1^{-1}$ replaced by $\\Sigma_1^{-1}-\\Sigma_c$, and the observability half of Wonham's conditions is what forces the rank condition on $\\beta$ and $\\delta$. The optimal portfolio is then the myopic term plus a hedging term proportional to $\\Sigma_1^{-1}\\Sigma_2(2C(t)z+b(t))$.","core_discovery":"The paper's central claim is that for $\\gamma<0$, the HJB equations for both the unconstrained and the market-neutral constrained portfolio have global, bounded solutions, which the authors interpret as absence of arbitrage and stable long-term growth. Proposition 2.3 uses Wonham's theorem on matrix Riccati equations to establish that $C(t)$ is unique, negative semi-definite, and bounded on $(-\\infty,T]$, with a unique limit $\\bar C$; Remark 2.3 states that this stability rules out finite-time singularities that would correspond to a 'Nirvana' arbitrage. For the constrained case, Proposition 2.8 proves the same boundedness provided the rank conditions in (2.42) hold. The paper further shows that the long-term certainty-equivalent growth rate is a constant, and that the market-neutral constraint $\\pi^\\top\\beta=0$ makes the portfolio adapted to the spread processes and therefore immunised against factor moves. In backtests, optimal portfolios are reported to be more profitable and more volatile than myopic ones, while constrained portfolios reduce risk relative to unconstrained ones.","pith_inferences":["A natural next step, not taken in the paper, is to test the constrained theorem's rank condition on each sliding-window estimate; if it fails often, the boundedness guarantee for market-neutral portfolios would not apply to real data.","Because the stability proof relies mainly on $\\gamma<0$ and on sign properties of diffusion matrices, the same Riccati-based argument could be carried over to other state dynamics, such as Ornstein-Uhlenbeck processes with stochastic volatility, provided the analogous coefficient matrices keep their signs.","The backtests deliberately exclude transaction costs and liquidity constraints; adding a 5--10 basis point penalty, which the paper notes is feasible, would show whether the high-volatility outperformance survives realistic frictions.","The paper's sensitivity finding implies that choosing window lengths and factor counts is itself a risk factor; an adaptive hyperparameter rule, rather than a fixed grid, is the implicit practical recommendation."],"forward_implications":["If the stability theorem is correct, neither the unconstrained nor the market-neutral model can produce a finite-time blow-up in the value function for $\\gamma<0$, so the 'Nirvana' arbitrage is excluded within the model.","The steady-state limits $\\bar C$ and $\\bar b$ give explicit long-run optimal portfolios that can be computed by solving an algebraic Riccati equation, which is what the backtests use.","The market-neutral constraint lowers volatility, maximum drawdown, and profit relative to the unconstrained strategy, but still outperforms the S&P 500 ETF in high-volatility subperiods.","Backtest results vary substantially with training and testing window lengths, which the authors read as sensitivity to parameter estimation.","High-volatility periods, including 2000--2003, 2008, 2010--2012, and 2020, are where the strategies show their best relative performance against the ETF."],"supporting_citations":[{"why":"It supplies the theorem used to prove existence, boundedness, and uniqueness of the matrix Riccati solution, which is the backbone of the stability results.","marker":"Wonham (1968)"},{"why":"It defines the factor model with eigenportfolios and co-integrated spreads from which the paper's state dynamics are built.","marker":"Avellaneda and Lee (2010)"},{"why":"It implements the same model and reports sensitivity of Sharpe ratios to estimation windows, the pattern the backtests here extend.","marker":"Yeo and Papanicolaou (2017)"},{"why":"It provides the co-integration and unit-root testing procedure used to decide which stocks enter the portfolio.","marker":"Engle and Granger (1987)"},{"why":"It supplies the verification framework used to show the candidate optimal control is admissible and optimal.","marker":"Davis and Lleo (2014)"},{"why":"It links finite-time singularities in the value function to 'Nirvana' arbitrage, which is the interpretation behind the stability result.","marker":"Lee and Papanicolaou (2016)"},{"why":"It gives the asymptotic estimation theory for Ornstein-Uhlenbeck parameters used for $\\hat\\theta$ and $\\hat\\delta$.","marker":"Kutoyants (2004)"},{"why":"It provides the covariance shrinkage estimator used to estimate the factor diffusion matrix $\\Sigma_0$.","marker":"Ledoit and Wolf (2004)"}],"fun_headline_variants":["Stable HJB arbitrage for co-integrated stock baskets","Market-neutral arbitrage stays stable, thrives in volatility","Co-integrated arbitrage: optimal portfolios, stable long-term","Multi-stock arbitrage: proven stability, volatility profits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stability claim for the market-neutral portfolio rests on the rank condition in (2.42), which the paper assumes but never verifies on data; if real S&P 500 estimates violate it, the proof of boundedness and stable growth for the constrained strategy no longer holds.","fun_headline_variants_meta":{"raw":{"variants":["Stable HJB arbitrage for co-integrated stock baskets","Market-neutral arbitrage stays stable, thrives in volatility","Co-integrated arbitrage: optimal portfolios, stable long-term","Multi-stock arbitrage: proven stability, volatility profits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1329,"prompt_tokens":938,"completion_tokens":391,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":323}},"tokens_in":554,"tokens_out":391,"duration_ms":4437,"temperature":1.0,"reasoning_tokens":323,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:52:07.823515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate $\\beta$ and $\\delta$ on any sliding-window training set of S&P 500 returns and check whether $\\beta^\\top\\beta$ has full rank and whether $\\delta-\\beta(\\beta^\\top\\beta)^{-1}\\beta^\\top\\delta$ has rank $d$; if either fails, the constrained portfolio's boundedness theorem does not apply, and the Riccati ODE can be solved numerically to see whether $C(t)$ remains bounded or blows up.","supporting_citations":[{"cited_title":"Statistical arbitrage in the US equities market","cited_arxiv_id":null,"evidence_quote":"It defines the factor model with eigenportfolios and co-integrated spreads from which the paper's state dynamics are built."},{"cited_title":"Risk control of mean-reversion time in statistical arbitrage","cited_arxiv_id":null,"evidence_quote":"It implements the same model and reports sensitivity of Sharpe ratios to estimation windows, the pattern the backtests here extend."},{"cited_title":"Pairs trading of two assets with uncertainty in co-integration's level of mean reversion","cited_arxiv_id":null,"evidence_quote":"It links finite-time singularities in the value function to 'Nirvana' arbitrage, which is the interpretation behind the stability result."},{"cited_title":"Honey I shrunk the sample covariance matrix","cited_arxiv_id":null,"evidence_quote":"It provides the covariance shrinkage estimator used to estimate the factor diffusion matrix $\\Sigma_0$."}],"review_version":1}