{"id":"01089abd-c9f6-42bd-87d7-97673d4120e7","arxiv_id":"1908.05089","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A symmetric Hawkes model is fit to S&P 500 tick data, and a heuristic diffusion counterpart is proposed that approximates its variance and skewness properties.","lead":"This paper fits a symmetric Hawkes model to millisecond stock price ticks and proposes a simpler diffusion model that mimics its behavior. It offers a fast, intraday volatility estimate from high-frequency trading data, with the caveat that the diffusion link is an approximation, not a proven limit.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Diffusion analogy's 'very close distributional property' is not established: using the paper's own Table 4/5 estimates, GE and T have only about 4–6 mid-price changes per minute, so the normal approximation in Section 3.1 is questionable.","rationale":"The reader's weakest assumption and my concern coincide: the diffusion analogy rests on an unquantified normal approximation of Poisson increments at the one-minute scale. I sharpen the point with the paper's own numbers. Under the MLEs in Tables 4 and 5, several stocks have only about 4–13 mid-price changes per minute, so replacing the Hawkes jump process by a Gaussian diffusion at that horizon is not merely lacking an error bound; it is likely to change the distribution materially. This matters because Proposition 4 (variance), Proposition 5 (skewness), and the simulated-likelihood estimation in Section 3.4 all use the diffusion model as if it were the Hawkes model. The paper is transparent that the derivation is heuristic, but the abstract and Section 3.1 still claim 'very close distributional property,' and that claim is not supported by the one-parameter, no-metric comparison in Figure 2. I do not object to the symmetric Hawkes MLE, the simulation study, or the TSRV/RRV comparisons; those are substantive and partly checkable. The load-bearing gap is the diffusion equivalence. A quantitative simulation-based distributional comparison would settle it. If the test fails, the paper would still contain a useful Hawkes volatility estimator, but the diffusion part would need to be repositioned as a tractable approximation with stated error bounds or restricted to high-intensity stocks. Since the reader already requires supporting evidence, the verdict remains CONDITIONAL and unchanged.","tokens_in":31484,"tokens_out":12364,"duration_ms":120247,"concrete_test":"Implement the symmetric Hawkes simulator of Appendix B with GE's 2011-01-03 parameters and T's Table 5 parameters; simulate 10^4 one-minute mid-price returns. Simulate the matched diffusion model (9) with the same parameters via Euler discretization with 60 sub-steps. Compare the two return distributions by Kolmogorov-Smirnov distance, skewness, kurtosis, and 1%/99% quantiles, and tabulate the mean number of jumps per minute. If the KS distance is large (say > 0.05) or the quantiles differ by more than 10%, the Section 3.1 normal approximation fails for these stocks, and the diffusion variance and skewness formulas need an explicit approximation-error bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 asserts that 'a sufficient number of price changes were observed during, e.g., one minute' and then replaces Poisson increments by normal increments. The paper's own parameter estimates contradict this for several analyzed stocks. For GE on 2011-01-03 (Table 4), the stationary intensity is lambda_infinity = mu*beta/(beta-alpha_s-alpha_c) = 0.0067*2.2596/(2.2596-0.4661-1.3576) ≈ 0.0347 per second per direction, i.e., about 4 mid-price changes per minute in total. Table 5 gives roughly 5.6 per minute for T, 12.7 for KO, and 8.4 for VZ. A Poisson sum with mean 4–13 is not close to Gaussian in distribution, so the diffusion SDE (9), the parameter mapping kappa1, kappa2, theta, gamma, phi, and Propositions 4–5 all inherit an unquantified approximation error. The only distributional evidence, Figure 2, uses mu = 0.09 with a 30-second horizon (about 8 events) and reports no error metric. Section 3.3 admits the model is 'not a rigorous mathematical transform' and only 'to provide an intuition not a mathematical proof,' but Section 3.4 then treats the diffusion as the working model. Thus the central 'very close distributional property' claim is not supported for the low-intensity stocks actually used.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a symmetric bivariate Hawkes process for tick-level mid-price dynamics, estimates its parameters by maximum likelihood from ultra-high-frequency NYSE data for ten S&P 500 stocks, and compares the resulting Hawkes-based volatility with two-scale realized volatility (TSRV) and the Robert-Rosenbaum volatility (RRV). The empirical part reports daily and intraday parameter dynamics and finds that Hawkes volatility is generally within 15–25% of TSRV. The paper then proposes a continuous-time diffusion analogue of the Hawkes model, with state variables for price, drift, and variance, and derives closed-form variance and third-moment formulas. Parameter estimates for the diffusion model are obtained by simulated maximum likelihood. The central claim is that the diffusion analogue has 'very close distributional property' to the Hawkes model while offering analytical simplicity.","tokens_in":31861,"tokens_out":5199,"duration_ms":50337,"significance":"If the claims were fully established, the Hawkes-based volatility estimator would be a useful tool for intraday risk measurement using all tick arrivals, and the diffusion analogue would provide a tractable continuous-time model for high-frequency volatility and skewness. The paper has concrete strengths: the Hawkes likelihood estimation is standard and is checked in a simulation study; the variance derivations in the appendices are detailed and checkable; the empirical comparison with TSRV and RRV is extensive and includes an interesting intraday volatility application. The main weakness is that the diffusion analogue—a major advertised contribution—is supported only by a heuristic normal approximation with no error control, and the empirical diffusion estimates do not match the Hawkes volatilities well. The paper is therefore promising but requires substantial additional work before the central diffusion-analogy claim can be accepted.","major_comments":[{"comment":"The normal approximation underlying the diffusion analogue is not justified for the data actually used. Section 3.1 replaces Poisson increments by normal increments on the grounds that 'a sufficient number of price changes were observed during, e.g., one minute.' The paper's own estimates contradict this: for GE on 2011-01-03, Table 4 gives mu=0.0067, alpha_s=0.4661, alpha_c=1.3576, beta=2.2596, so the stationary intensity per direction is about 0.0347 per second, or about 4 mid-price changes per minute total; Table 5 gives roughly 5.6 per minute for T. A Poisson sum with mean 4–13 is not close to Gaussian, and no quantitative error bound is provided. Section 3.3 concedes the derivation is 'not a rigorous mathematical transform' and only 'to provide an intuition,' yet Section 3.4 uses the diffusion model as the working model for estimation. The 'very close distributional property' claim is therefore not supported for the low-intensity stocks in the empirical study; Figure 2 uses a high-intensity parameter setting (mu=0.09, 30-second horizon) and reports no error metric. Please provide a quantitative justification (e.g., a Berry-Esseen bound or a limit theorem with stationarity conditions) or explicitly restrict the claim to regimes where the normal approximation is verified.","section":"Section 3.1 and Section 3.3"},{"comment":"Part of the agreement between the diffusion model and the Hawkes model is by construction. The parameter relation theta = 2bm*delta^2/(b-as-ac) is chosen so that the long-run variance of the diffusion model, b^2 theta t/(S0^2 kappa1^2), exactly equals the long-run variance of the Hawkes model in Remark 2 when m=mu, as=alpha_s, ac=alpha_c, b=beta. Thus the asymptotic variance comparison after Proposition 4 is not independent evidence of distributional closeness; it is baked into the parameter mapping. The finite-time variance in Proposition 4 is a genuine derivation, but its agreement with Proposition 3 is shown in only one figure (Figure 3) and one parameter set. Please clarify in the text that the long-run variance match is by construction and present finite-time comparisons over a range of empirically calibrated parameters.","section":"Section 3.1, parameter mapping after Eq. (9)"},{"comment":"The claimed empirical similarity between the diffusion-model estimates and the Hawkes estimates is not supported by the reported numbers. For GE in January 2011, Table 8 gives diffusion volatilities that are frequently much larger than the Hawkes H.vol in Table 4: on 0105 the diffusion volatility is 0.3661 versus 0.1339 for Hawkes, on 0104 it is 0.1546 versus 0.1139, and on 0110 it is 0.1739 versus 0.1520. Several diffusion estimates also have parameter values very different from the Hawkes estimates (e.g., ac=2.9568 on 0104 vs. alpha_c=1.3941 in Table 4). The text states that the results are 'similar' to the Hawkes model without providing a quantitative comparison, confidence intervals, or an explanation of the systematic discrepancy. Please either provide a proper comparison (e.g., ratio statistics, error bars, or a scatter plot) or temper the claim about the empirical similarity of the two models.","section":"Section 4.5, Tables 8 and 4"}],"minor_comments":[{"comment":"The sentence 'This study examine the theoretical and empirical perspectives' should read 'This study examines'; the same grammatical issue appears in the introduction.","section":"Abstract"},{"comment":"There is a typo 'Haweks' in the paragraph introducing the symmetric model; it should be 'Hawkes'.","section":"Section 2.3"},{"comment":"The reported estimates for 'simulation set 1' in the text appear to actually be from simulation set 2 (alpha_s=0.6590 is close to the true alpha_s=0.65 of set 2, not the true alpha_s=0.4 of set 1). Also, 'alpha_c = 0.0.4864' contains a typo and should be '0.4864'.","section":"Section 2.4"},{"comment":"The caption of Figure 2 says 'histogram of the Hawkes model price by the simulation with 30 seconds (right)' but the panel labels are not visible; please clarify which panel shows the density and which shows the histogram, and add a legend or labels.","section":"Figure 2 and Section 3.2"},{"comment":"Tables 8 and 9 report diffusion parameter estimates without standard errors or any measure of simulation variability; because the simulated likelihood depends on random numbers, at least the seed or a small Monte Carlo standard error should be reported.","section":"Section 4.5"},{"comment":"The likelihood form in Eq. (C.1) should define the log-likelihood function with the parameter vector explicitly; the notation L(θ,T) appears only later, and the text introduces 'the conditional cumulative distributions' with an unnumbered equation that is not referenced.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The diffusion analogy is the weakest link in the paper. The authors should either prove a proper limit theorem (or at least provide a quantitative approximation error bound) or substantially soften the 'very close distributional property' claim and present the diffusion as a standalone heuristic model. The empirical volatility comparison between Tables 8 and 4 needs a serious quantitative treatment before the paper can be accepted. The Hawkes volatility estimation part is solid and could stand alone as a publishable contribution if the diffusion section is appropriately reframed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: the real contribution is the diffusion analogue of the symmetric Hawkes tick model—its closed-form variance, the third-moment condition for leverage, and the intraday volatility estimates. The Hawkes variance formula itself largely overlaps with Da Fonseca and Zaatour (2014a), and the authors say so. What is new is the diffusion construction with explicit parameter mapping and the empirical comparison against TSRV and RRV.\n\nThe paper does several things well. The Hawkes MLE is standard but carefully presented, the simulation study checks finite-sample behavior, and the appendices give checkable algebra for the variance derivations. The empirical tables and figures are useful: volatility estimates from the Hawkes model track TSRV and RRV within 15–25%, and the intraday volatility procedure is a nice practical demonstration. The authors also acknowledge the main weakness of the diffusion analogy in Section 3.3, calling it intuition rather than a rigorous transform.\n\nThe soft spots are real but not fatal. The stress-test concern lands: the normal approximation at the one-minute scale is questionable for the actual stocks used. Using Table 4/5, GE and T have roughly 4–6 mid-price changes per minute, and KO and VZ around 8–13. A Poisson sum with mean 4–13 is not close to Gaussian. So the claimed 'very close distributional property' is not established for the low-intensity stocks; Figure 2 uses about 8 events at 30 seconds and reports no error metric. Also, the diffusion long-run variance is calibrated to match the Hawkes variance through theta, so some agreement between Proposition 4 and Proposition 3 is by construction. The 13–33% empirical gaps are left unexplained, which the authors openly say. No error bars appear on the diffusion estimates, and no code or data are released.\n\nNone of this sinks the paper if read as a heuristic with empirical illustrations. But if the claim is that the diffusion model is very close to Hawkes, the current evidence is qualitative, not quantitative. This is a paper for practitioners and researchers who want a tractable tick-level volatility and skewness model, and for empiricists comparing Hawkes-based volatility with realized measures.\n\nVerdict: worth a serious referee, conditional on revision. Add a quantitative closeness test (for example, a distribution distance between Hawkes simulations and diffusion paths under estimated parameters), standard errors for the diffusion estimates, and release code and data. Then it would be publishable. I would bring it to a reading group and would cite the variance formula if I were working on tick models.\n\nBest,","headline":"Diffusion analogue is a useful idea with honest caveats, but the claimed closeness to Hawkes is not quantitatively established; deserves a serious referee and a revision.","tokens_in":32366,"tokens_out":2522,"would_cite":true,"duration_ms":25912,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The symmetric Hawkes model, estimated from every tick-level mid-price move, gives volatility within about 15–25 percent of realized measures and a diffusion analogue with close distributional properties.","keywords":["Hawkes process","tick dynamics","ultra-high-frequency data","volatility estimation","simulated maximum likelihood","diffusion approximation","market microstructure","realized volatility"],"falsifier":"Take a low-activity stock or day, simulate the fitted symmetric Hawkes process, and compare one-minute mid-price return densities with the diffusion analogue's Kolmogorov-forward density; a significant distributional distance would show the normal approximation fails at that time scale.","tokens_in":31257,"feed_emoji":"📈","tokens_out":10603,"duration_ms":93564,"temperature":0.7,"pith_summary":"The paper tries to establish that a symmetric Hawkes process—the difference of two mutually exciting point processes describing up and down movements of the mid-price—is a practical model for ultra-high-frequency tick dynamics. Combined with maximum likelihood estimation on all price-change arrival times, it yields volatility estimates that track the two-scale realized volatility (TSRV) and the uncertainty-zone realized volatility (RRV) within roughly 15 to 25 percent for ten S&P 500 stocks over 2007–2011. The paper further claims that a diffusion analogue, with parameters explicitly tied to the Hawkes parameters, has a very close distributional property to the Hawkes model, making Itô-calculus tools available for variance, skewness, and leverage. A sympathetic reader would care because the model turns tick-by-tick data directly into volatility estimates, usable on short intraday windows, without first removing market microstructure noise.","feed_headline":"Tick-by-tick Hawkes volatility matches S&P 500 realized measures","feed_subtitle":"Maximum likelihood on every mid-price move yields volatility and a diffusion analogue with close returns.","key_machinery":"The load-bearing object is the symmetric Hawkes process: a bivariate point process $(N_1,N_2)$ with exponential-decay intensities sharing baseline $\\mu$, decay $\\beta$, self-excitation $\\alpha_s$, and mutual excitation $\\alpha_c$, whose difference scaled by the tick size $\\delta$ is the mid-price. Its tractability comes from two facts: fixing $\\beta$ makes the log-likelihood concave in $(\\mu,\\alpha_s,\\alpha_c)$, so the maximum can be located reliably; and the stationarity assumption $\\lambda_1(0)=\\lambda_2(0)=\\mu\\beta/(\\beta-\\alpha_s-\\alpha_c)$ reduces the return variance to a closed formula. The diffusion analogue $dS_t=n_tdt+\\sqrt{V_t}dW^s_t$, $dn_t=-\\kappa_1 n_tdt+\\phi\\sqrt{V_t}dW^s_t$, $dV_t=\\kappa_2(\\theta-V_t)dt+\\gamma\\sqrt{V_t}dW^v_t$ with the parameter identities above carries the second half of the paper: it imports Heston-type square-root variance machinery, a leverage parameter $\\rho$, and simulated maximum likelihood estimation into the tick-level setting.","core_discovery":"On the paper's own terms, the central discovery is that the price process $S_t = S_0 + \\delta(N_1(t)-N_2(t))$, with $N_1,N_2$ counting up and down moves of the mid-price and each intensity following a Hawkes process with common baseline $\\mu$, decay $\\beta$, self-excitation $\\alpha_s$, and mutual excitation $\\alpha_c$, can be estimated by maximum likelihood from ultra-high-frequency data. Under the stationarity condition $\\lambda_1(0)=\\lambda_2(0)=\\mu\\beta/(\\beta-\\alpha_s-\\alpha_c)$, the return variance has the closed form $$\\operatorname{Var}\\left(\\frac{S_t-S_0}{S_0}\\right)=\\frac{2\\$delta^{2}$\\lambda_1(0)}{$S_0^{2}$\\$xi_1^{2}$}\\left\\{\\$beta^{2}$t-\\frac{2(\\alpha_s-\\alpha_c)\\$\\beta$}{\\xi_1}($e^{{\\xi_1 t}}$-1)+\\frac{(\\alpha_s-\\alpha_c)^2}{2\\xi_1}($e^{{2\\xi_1 t}}$-1)\\right\\},$$ with $\\xi_1=-\\beta-\\alpha_c+\\alpha_s$, and this Hawkes volatility tracks TSRV and RRV on ten S&P 500 stocks with mean percentage errors typically in the 15–25 percent range. The companion claim is that replacing Poisson increments by Brownian increments over one-minute intervals produces the diffusion system $dS_t=n_tdt+\\sqrt{V_t}dW^s_t$, $dn_t=-\\kappa_1 n_tdt+\\phi\\sqrt{V_t}dW^s_t$, $dV_t=\\kappa_2(\\theta-V_t)dt+\\gamma\\sqrt{V_t}dW^v_t$, with $\\kappa_1=b-a_s+a_c$, $\\kappa_2=b-a_s-a_c$, $\\theta=2bm\\delta^2/(b-a_s-a_c)$, $\\gamma=\\delta(a_s+a_c)$, $\\phi=a_s-a_c$, and $d[W^s,W^v]_t=\\rho dt$; the paper reports that this diffusion has a very close distributional property to the Hawkes model, giving simpler variance and skewness formulas and a leverage parameter $\\rho$.","pith_inferences":["A natural extension is to estimate the diffusion model on one-minute bins for less liquid names, where the normal approximation is most doubtful; if distributional closeness breaks down there, that would identify the shortest safe sampling scale.","The mean signature plot analysis implies that the sign of the parameter $\\phi=a_s-a_c$ controls how realized variance responds to sampling frequency, and this directional prediction is directly checkable on the same high-frequency data.","If the diffusion analogy holds, Heston-style option pricing formulas could be applied with parameters estimated from tick arrivals, connecting ultra-high-frequency estimation to derivatives pricing.","The paper's 15–25 percent gap between Hawkes volatility and TSRV is left unexplained; allowing the baseline intensity to vary intraday in the likelihood would test whether parameter drift accounts for the gap."],"forward_implications":["Volatility can be estimated from as little as ten minutes of mid-price tick data, so the model produces an intraday volatility curve updated every ten minutes.","The closed-form Hawkes variance formula lets practitioners reparametrize the likelihood directly in terms of annualized volatility and estimate it without first removing microstructure noise.","If the diffusion analogue is accepted, variance, skewness, and leverage formulas follow from Itô calculus, making the tick-level model compatible with standard continuous-time tools.","The empirical comparison shows that the Hawkes volatility tracks TSRV and RRV daily, with mean percentage errors between the two measures staying around 15–25 percent for ten S&P 500 stocks.","The fully characterized Hawkes estimates show the symmetry assumptions are often reasonable, while cases like XOM in 2008 indicate persistence differences between self- and mutual excitation that a symmetric model cannot capture."],"supporting_citations":[{"why":"Supplies the original symmetric bivariate Hawkes setup for order arrivals and price impact on which the price process is built.","marker":"Hewlett (2006)"},{"why":"Supplies the two-scale realized volatility (TSRV) estimator used as the simulation and empirical benchmark.","marker":"Zhang et al. (2005)"},{"why":"Documents the bias of realized variance under microstructure noise, motivating the direct modeling of observed price changes.","marker":"Hansen and Lunde (2006)"},{"why":"Provides moment conditions, autocorrelation functions, and a related mean signature formula for Hawkes microstructure models that the paper extends and compares with.","marker":"Da Fonseca and Zaatour (2014a)"},{"why":"Gives the asymptotic normality of maximum likelihood estimators for stationary point processes, justifying the standard errors and likelihood inference.","marker":"Ogata (1978)"},{"why":"Supplies the simulated likelihood estimation method used for the diffusion analogue.","marker":"Brandt and Santa-Clara (2002)"},{"why":"Supplies the uncertainty-zone realized volatility (RRV) estimator used as the second empirical benchmark.","marker":"Robert and Rosenbaum (2011)"},{"why":"Supplies the square-root variance process structure that the diffusion analogue's variance equation mirrors.","marker":"Heston (1993)"}],"fun_headline_variants":["Symmetric Hawkes model captures ultra-high-frequency price moves","Diffusion analogue of Hawkes matches return distribution closely","Ten S&P 500 stocks: Hawkes volatility close to realized measures","Closed-form variance from Hawkes with diffusion surrogate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that over a one-minute interval enough mid-price changes occur for the Poisson increments in the Hawkes process to be well approximated by normal (Brownian) increments; the paper gives no error bound for this replacement, and the diffusion model's very close distributional claim depends on it.","fun_headline_variants_meta":{"raw":{"variants":["Symmetric Hawkes model captures ultra-high-frequency price moves","Diffusion analogue of Hawkes matches return distribution closely","Ten S&P 500 stocks: Hawkes volatility close to realized measures","Closed-form variance from Hawkes with diffusion surrogate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000741,"raw_usage":{"total_tokens":3408,"prompt_tokens":1149,"completion_tokens":2259,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":765,"completion_tokens_details":{"reasoning_tokens":2201}},"tokens_in":765,"tokens_out":2259,"duration_ms":16856,"temperature":1.0,"reasoning_tokens":2201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:24:21.979132+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a low-activity stock or day, simulate the fitted symmetric Hawkes process, and compare one-minute mid-price return densities with the diffusion analogue's Kolmogorov-forward density; a significant distributional distance would show the normal approximation fails at that time scale.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the original symmetric bivariate Hawkes setup for order arrivals and price impact on which the price process is built."},{"cited_title":"A., and A \\\" t-Sahalia, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the two-scale realized volatility (TSRV) estimator used as the simulation and empirical benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the bias of realized variance under microstructure noise, motivating the direct modeling of observed price changes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the asymptotic normality of maximum likelihood estimators for stationary point processes, justifying the standard errors and likelihood inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the simulated likelihood estimation method used for the diffusion analogue."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the uncertainty-zone realized volatility (RRV) estimator used as the second empirical benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the square-root variance process structure that the diffusion analogue's variance equation mirrors."}],"review_version":1}