{"id":"42874d2d-2ba4-470a-92a6-39e60c7f404e","arxiv_id":"1908.07102","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Short rates in the quasi-Gaussian HJM model with CEV type volatility explode in finite time with positive probability for exponents in (1/2,1] and almost surely for sufficiently large initial rates.","lead":"Short rates in a widely used interest-rate model (the quasi-Gaussian, or Cheyette, HJM model) are proved to explode to infinity in finite time with positive probability, and almost surely under stronger conditions. This matters because such explosions make zero-coupon bond prices collapse and Eurodollar futures infinite, restricting the maturities the model can price reliably.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1(b) invokes Proposition 1 from the set Γ, but the proof never shows that the stated initial condition (r0,0) reaches Γ.","rationale":"The central claim is Theorem 1(b), which is stated for the time-homogeneous system (20)-(21) starting at (r0,0). The proof constructs a Lyapunov function satisfying the hypotheses of Proposition 1 and then applies that proposition. The load-bearing step is therefore the activation of Proposition 1 from the actual initial state. As quoted in the paper, Proposition 1 requires the process to start in Γ, yet no argument shows that the process reaches Γ with positive probability from (r0,0). This is a genuine gap in the written proof, though likely repairable. The reader's identified concern, the comparison step near Eq. (19), is less central: it affects only the transfer from the time-homogeneous theorem to time-dependent λ, while Theorem 1 itself is already time-homogeneous. The reader's algebraic worries are also less severe than suggested. Equation (58) indeed contains a typo: x should be r^{δ2+1}/y rather than r^{δ1+1}/y; with that correction, the reduction in Eq. (57) is consistent. Equations (68) and (69) are equivalent after substituting the definitions of κ1, κ2, a, and b, so that alleged mismatch does not undermine the proof. The strongest remaining issue is the missing hitting step for Proposition 1, which supports the same conditional verdict rather than a rejection.","tokens_in":17991,"tokens_out":29250,"duration_ms":288871,"concrete_test":"Using the γ=1, β=0, σ=0.2, r0=0.1 case of Figure 1, maximize V(r0,0)-K2 over R≥ε, δ1,δ2>0 with (1+δ1)(1+δ2)=2, and C2,C3 satisfying the inequalities in Remark 1. If the maximum is nonpositive, the Lyapunov argument as written cannot be invoked from the theorem's initial condition; the authors must then supply a separate lemma proving P_{(r0,0)}(τ_Γ<∞)>0, for example via a support-theorem/control argument. If a parameter choice gives V(r0,0)>K2, the direct route may survive and the gap narrows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1(b) verifies hypotheses (A.1)-(A.3) of Proposition 1 and then concludes explosion with positive probability. But Proposition 1, as stated in the paper, gives that conclusion only when the process starts at some future time in Γ=[2R,∞)×[2R,∞). The theorem's initial condition is (r0,0), which is never in Γ and, for r0<R, lies inside D. No argument is given that the diffusion reaches Γ with positive probability from (r0,0). The Lyapunov inequality LV≥CV on D^c is inert for paths that stay in D or on the y=0 side of D^c. For the Figure 1 parameters (γ=1, β=0, σ=0.2, r0=0.1) and the natural choice R=1/δ2=1, V(r0,0)<K2, so the constants chosen in Remark 1 do not themselves put the initial point in the region where the explosion test applies. This gap is more load-bearing than the comparison in Eq. (19), because Theorem 1 is already time-homogeneous; the comparison only affects the extension to time-dependent λ.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the ε-CEV modified one-factor quasi-Gaussian HJM short-rate model (8)-(9). Its main results are for the time-homogeneous case λ(t)=λ0: Theorem 1 proves non-explosion for γ∈(0,1/2] and, for γ∈(1/2,1], finite-time explosion with positive probability under either of the parameter conditions (i) sup_R F(R;β,σ)>0 or (ii) sup_R (G(R)-(2β+1/2σ^2δ2(δ2+1)))≥0. Theorem 2 upgrades this to almost-sure explosion when β>0 and r0 satisfies (40). The proof uses a Lyapunov function V(r,y)=C1-C2/(1+y)^{δ1}-C3/(1+r)^{δ2} together with the Chow-Khasminskii explosion criteria. The paper closes with implications for zero-coupon bond prices and Eurodollar futures and a small numerical illustration.","tokens_in":18219,"tokens_out":25889,"duration_ms":251035,"significance":"The result, if fully established, is a useful rigorous addition to the interest-rate modeling literature: it identifies a wide and practically relevant parameter region in which the quasi-Gaussian HJM model with CEV-type volatility has explosive short-rate paths, and it quantifies the consequences for derivatives pricing. The construction of the Lyapunov function is explicit and the parameter conditions are checkable; the numerical section illustrates that the proved region is not vacuous. The main obstruction is that the proof of Theorem 1 as written does not connect the initial condition (r0,0) to the set Γ where the Khasminskii criterion applies, and the comparison reduction at Eq. (19) is asserted without proof. Both are localizable and appear repairable, so the manuscript merits a major revision rather than rejection.","major_comments":[{"comment":"The proof verifies hypotheses (A.1)-(A.3) of Proposition 1 and then concludes P(τ<∞)>0. Proposition 1, as quoted in the paper, gives this conclusion only when the process starts at a point of Γ=[2R,∞)×[2R,∞). The initial condition of Theorem 1 is (r0,0); this point never lies in Γ because y0=0, and for the parameter values used in Figure 1 (r0=0.1, R=1/δ2=1) it lies in D. The paper gives no argument that the diffusion reaches Γ with positive probability from (r0,0). The inequality LV≥CV is only used on D^c and says nothing about the crossing from D to Γ. Please add a reachability lemma, for example via the support theorem and an explicit control path from (r0,0) to Γ, or restrict the theorem to initial conditions in Γ.","section":"§3, proof of Theorem 1(b)"},{"comment":"The reduction from the time-inhomogeneous system (8)-(9) to the time-homogeneous system (20)-(21) is asserted in one sentence with a reference to Yamada [36], but no comparison proof is given. For the statement to hold one needs both λ'(t)+βλ(t)≥βλ(0) and the monotone dependence of y on r, and this should be shown explicitly. Since a declining forward curve can violate (19), the paper should either prove the comparison or clearly state that the rigorous explosion results are confined to λ(t)≡r0 and that Theorem 1 does not cover general forward curves.","section":"§3, Eq. (19)"}],"minor_comments":[{"comment":"The definition of x should be x = r^{δ2+1}/y, not x = r^{δ1+1}/y. With the printed definition the two terms in (57) do not collapse to a single function of x; with the corrected exponent the derivation is valid.","section":"Appendix, Eq. (58)"},{"comment":"For the record, Eq. (69) is consistent with Eq. (68): substituting a and b from (71) into (68) reproduces (69) exactly; the apparent discrepancy in the ordering of the terms inside the minimum is only a notational matter.","section":"Appendix, Eq. (69)"},{"comment":"The small-δ2 expansion of G(R0(δ2)) appears to have a sign error: from G(δ)=δ^{1+δ}(1+δ)^{-(1+δ)} one obtains G(δ)=δ+δ^2(logδ-1)+O(δ^3(logδ)^2), not δ2+δ2^2(logδ2+1). The leading-order conclusion σ_max=√2 is unaffected.","section":"§3.1, Eq. (38)"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript is within the journal's scope and the topic is appropriate. The main theorem is not proved as stated because of the missing reachability step; this is repairable with a lemma rather than a fundamental flaw. The authors' previous paper [30] is used appropriately as motivation and for the numerical comparison, so I do not see a self-citation problem. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dan, quick take on arXiv:1908.07102. The real contribution is a rigorous statement that the one-factor quasi-Gaussian HJM model with CEV-type volatility (γ∈(1/2,1]) can explode in finite time: positive probability for the baseline case, a.s. for sufficiently large initial short rate. That extends the earlier deterministic small-noise result to the stochastic system and to the CEV class, and it is a useful model-risk warning for swaption and Eurodollar pricing. The Lyapunov strategy from Chow-Khasminskii is appropriate, and the parameter regions in Figure 3 are informative. The citation pattern is unproblematic: self-citation appears only as motivation and for the numerical illustration.\n\nThe main problem is in the proof of Theorem 1(b). Proposition 1 only concludes explosion with positive probability if the process starts in Γ=[2R,∞)×[2R,∞). The initial condition is (r0,0), which is never in Γ; it lies in D when r0<R, and even when r0≥R, y0=0<2R. The paper never shows that the diffusion reaches Γ with positive probability. The Lyapunov inequality on D^c is inert for paths that stay in D. This is not a typo; as written, Theorem 1 does not follow from the cited criterion. It is likely repairable by a standard controllability/hitting argument, since the r-diffusion is nondegenerate and y is driven positive, but that argument must be added.\n\nThere are also algebra slips. Eq. (58) defines x=r^{δ1+1}/y, but the exponents only close if x=r^{δ2+1}/y (then 2γ=(1+δ1)(1+δ2) gives the desired cancellation). In Lemma 2's first region, the displayed exponent δ2(2γ-δ1-1) should be 2γ-δ1-1 (using δ2(δ1+1)=2γ-δ1-1). The reader's flag on Eq. (69) I think is a misread: after substituting a and b, (69) is exactly equivalent to (68). The comparison step for time-dependent λ is asserted with a citation, but it is not needed for Theorem 1 where λ is constant.\n\nBottom line: the central result is plausible and worth a serious referee, but the submission needs a revision that closes the Γ gap and fixes the typos. I'd send it to review rather than desk reject.","headline":"Worth a serious referee, but Theorem 1 has a missing hitting-Γ argument that must be repaired.","tokens_in":18768,"tokens_out":12963,"would_cite":false,"duration_ms":107078,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H10","60J60","91G30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The quasi-Gaussian HJM short rate explodes in finite time with positive probability for CEV exponents in $(1/2,1]$, and almost surely for large initial rates.","keywords":["HJM model","short rate explosion","CEV volatility","Lyapunov function","multidimensional diffusion","zero coupon bond","Eurodollar futures","finite-time explosion"],"falsifier":"Simulate the time-homogeneous SDE (20)-(21) with parameters satisfying condition (i) or (ii), for example $\\gamma=1$, $\\beta=0.05$, $\\sigma=0.2$, $r_0=0.1$, over a long horizon and estimate the probability of explosion from a large ensemble; the proof guarantees this probability is positive, so a carefully controlled run finding no explosion at all would contradict the claim. To test the comparison step directly, simulate the time-inhomogeneous model (8)-(9) with a forward curve that violates $\\lambda'(t)+\\beta\\lambda(t)\\ge\\beta\\lambda(0)$ and check whether explosions persist.","tokens_in":17739,"feed_emoji":"📈","tokens_out":10329,"duration_ms":83417,"temperature":0.7,"pith_summary":"The paper studies the one-factor quasi-Gaussian Heath-Jarrow-Morton interest-rate model with an $\\varepsilon$-CEV volatility specification, in which the short rate's volatility behaves like $\\sigma r^\\gamma$ for large $r$. It establishes that for elasticity parameter $\\gamma\\in(1/2,1]$, including the log-normal case $\\gamma=1$, the short rate explodes to infinity in finite time with positive probability whenever one of two explicit parameter conditions holds; for $\\beta>0$ and a sufficiently large initial short rate, the explosion occurs with probability one. The practical consequence is that a finite-time explosion makes zero-coupon bond prices collapse to zero and Eurodollar futures prices infinite at sufficiently long maturities, so the model's useful pricing horizon is bounded by the explosion time. The paper also proves that for $\\gamma\\in(0,1/2]$ the model is non-explosive.","feed_headline":"A popular bond model's short rate can explode in finite time","feed_subtitle":"CEV-type volatility makes the short rate blow up with positive probability, breaking long-dated bond and futures prices.","key_machinery":"The argument is carried by a Lyapunov function $V(r,y)=C_1-\\frac{C_2}{(1+y)^{\\delta_1}}-\\frac{C_3}{(1+r)^{\\delta_2}}$ with $(1+\\delta_1)(1+\\delta_2)=2\\gamma$, together with a comparison theorem that reduces the time-dependent forward-curve model to a time-homogeneous system when $\\lambda'(t)+\\beta\\lambda(t)\\ge\\beta\\lambda(0)$. Applying the infinitesimal generator of the diffusion to $V$ and invoking the Lyapunov-function explosion criterion of [12] turns the explosion problem into two algebraic inequalities in $\\beta,\\sigma,\\gamma$, giving the explicit conditions (i) and (ii). A second auxiliary function $V_0=e^{-r}+e^{-y}$ is used to verify the hitting-time condition needed for almost-sure explosion.","core_discovery":"The central discovery is that explosion is a genuine property of the quasi-Gaussian HJM model with a CEV-type volatility, not an artefact of a small-noise approximation. For the time-homogeneous model with constant initial forward rate $r_0$, Theorem 1 shows that the two-dimensional diffusion $(r_t,y_t)$ with $\\varepsilon$-CEV volatility explodes with positive probability for $\\gamma\\in(1/2,1]$ provided that either $\\sup_{R\\ge\\varepsilon}F(R;\\beta,\\sigma)>0$ or $\\sup_{R\\ge\\varepsilon}\\bigl(G(R)-(2\\beta+\\tfrac12\\sigma^2\\delta_2(\\delta_2+1))\\bigr)\\ge0$, where $F$ and $G$ are explicit functions of the model parameters and $\\delta_1,\\delta_2>0$ satisfy $(1+\\delta_1)(1+\\delta_2)=2\\gamma$. Under the extra assumptions $\\beta>0$ and a large enough initial short rate satisfying inequality (40), Theorem 2 upgrades this to almost-sure explosion. The log-normal case $\\gamma=1$ is included, and the authors note that this matches explosions observed numerically and previously derived in a deterministic small-noise limit.","pith_inferences":["The two conditions in Theorem 1 are sufficient, not necessary; the true explosion region in the $(\\beta,\\sigma,\\gamma)$ parameter space is likely larger, so explosions may also occur for parameter sets outside the proved region.","The comparison assumption $\\lambda'(t)+\\beta\\lambda(t)\\ge\\beta\\lambda(0)$ means the proved explosion applies most directly to flat or rising forward curves; for downward-sloping yield curves the mechanism could be delayed or suppressed, and this is testable by simulation.","The same Lyapunov ansatz may extend to multi-factor quasi-Gaussian models, where no necessary-and-sufficient explosion criterion is known, potentially yielding explicit sufficient conditions for explosion in higher-dimensional term-structure models.","The almost-sure explosion in Theorem 2 requires a large initial short rate; for small initial rates only positive-probability explosion is guaranteed, so low-rate calibrations may remain on the non-explosive side in practice."],"forward_implications":["Zero-coupon bonds with maturity beyond the explosion time have price zero with positive probability, because $P(T,T+\\delta)$ contains the factor $\\exp(-G(T,T+\\delta)x_T-\\tfrac12 G(T,T+\\delta)^2y_T)$ and $x_T,y_T$ explode.","Eurodollar futures prices, which involve $\\mathbb{E}^Q[P^{-1}(T,T+\\delta)]$, become infinite for sufficiently large $T$ when the Theorem 1 conditions hold, and LIBOR-linked caps, swaptions and CMS products inherit the singularity.","Explosion is confined to $\\gamma\\in(1/2,1]$; for $0<\\gamma\\le1/2$ the coefficients satisfy a linear-growth condition and the solution is non-explosive and square-integrable.","With mean reversion $\\beta>0$ and a sufficiently high initial short rate satisfying inequality (40), explosion occurs almost surely rather than merely with positive probability.","Capping the short-rate volatility at a finite level restores sub-linear growth and eliminates explosion, a practical remedy for pricing applications."],"supporting_citations":[{"why":"Supplies the Lyapunov-based sufficient conditions for explosion with positive probability and for almost-sure explosion that drive the proof of Theorems 1 and 2.","marker":"[12]"},{"why":"Provides the stochastic-stability framework and the auxiliary-function criterion used to verify the hitting-time condition (A.5) in Theorem 2.","marker":"[24]"},{"why":"Gives the standard existence, uniqueness and non-explosion theorem applied to the $\\gamma\\in(0,1/2]$ case and the comparison theorem used for positivity.","marker":"[22]"},{"why":"The comparison theorem invoked to transfer explosion from the time-homogeneous auxiliary system to the original time-inhomogeneous system and to prove $r_t>0$.","marker":"[36]"},{"why":"The paper's earlier deterministic small-noise analysis of the log-normal model, whose finite-time explosion result the present paper extends to the stochastic setting.","marker":"[30]"},{"why":"The original HJM framework whose log-normal explosion and zero bond collapse motivate the question studied here.","marker":"[16]"},{"why":"Provides the quasi-Gaussian model representation, the zero-coupon bond pricing formula (4), and the displaced log-normal substitution.","marker":"[3]"},{"why":"Introduces the $\\varepsilon$-cutoff modification of the CEV volatility that appears in equation (7) and handles the behavior near zero.","marker":"[1]"},{"why":"Supplies practical parameter ranges for the numerical study of the allowed $(\\beta,\\sigma)$ region and the Eurodollar futures pricing context.","marker":"[6]"}],"fun_headline_variants":["Explosion proven for quasi-Gaussian HJM short rate","CEV volatility makes HJM short rate explode","Short rate blow-up in bond model: rigorous proof","Finite-time explosion in popular yield curve model","HJM model's short rate can explode: new theorem"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The most fragile load-bearing premise is the comparison step near equation (19): the explosion proved for the time-homogeneous auxiliary model is transferred to the original time-inhomogeneous model only when the forward curve satisfies $\\lambda'(t)+\\beta\\lambda(t)\\ge\\beta\\lambda(0)$, an inequality that can fail for a declining forward curve.","fun_headline_variants_meta":{"raw":{"variants":["Explosion proven for quasi-Gaussian HJM short rate","CEV volatility makes HJM short rate explode","Short rate blow-up in bond model: rigorous proof","Finite-time explosion in popular yield curve model","HJM model's short rate can explode: new theorem"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1260,"prompt_tokens":895,"completion_tokens":365,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":289}},"tokens_in":511,"tokens_out":365,"duration_ms":3586,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:28:52.348335+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the time-homogeneous SDE (20)-(21) with parameters satisfying condition (i) or (ii), for example $\\gamma=1$, $\\beta=0.05$, $\\sigma=0.2$, $r_0=0.1$, over a long horizon and estimate the probability of explosion from a large ensemble; the proof guarantees this probability is positive, so a carefully controlled run finding no explosion at all would contradict the claim. To test the comparison step directly, simulate the time-inhomogeneous model (8)-(9) with a forward curve that violates $\\lambda'(t)+\\beta\\lambda(t)\\ge\\beta\\lambda(0)$ and check whether explosions persist.","supporting_citations":[{"cited_title":"Khasminskii","cited_arxiv_id":null,"evidence_quote":"Supplies the Lyapunov-based sufficient conditions for explosion with positive probability and for almost-sure explosion that drive the proof of Theorems 1 and 2."},{"cited_title":"(2012) Stochastic Stability of Differential Equations","cited_arxiv_id":null,"evidence_quote":"Provides the stochastic-stability framework and the auxiliary-function criterion used to verify the hitting-time condition (A.5) in Theorem 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the standard existence, uniqueness and non-explosion theorem applied to the $\\gamma\\in(0,1/2]$ case and the comparison theorem used for positivity."},{"cited_title":"(1973) On a comparison theorem for solutions of stochastic differential equations and its applications","cited_arxiv_id":null,"evidence_quote":"The comparison theorem invoked to transfer explosion from the time-homogeneous auxiliary system to the original time-inhomogeneous system and to prove $r_t>0$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The paper's earlier deterministic small-noise analysis of the log-normal model, whose finite-time explosion result the present paper extends to the stochastic setting."},{"cited_title":"Jarrow and A","cited_arxiv_id":null,"evidence_quote":"The original HJM framework whose log-normal explosion and zero bond collapse motivate the question studied here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the quasi-Gaussian model representation, the zero-coupon bond pricing formula (4), and the displaced log-normal substitution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the $\\varepsilon$-cutoff modification of the CEV volatility that appears in equation (7) and handles the behavior near zero."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies practical parameter ranges for the numerical study of the allowed $(\\beta,\\sigma)$ region and the Eurodollar futures pricing context."}],"review_version":1}