{"id":"f228fe5b-1d8b-4edd-9c90-e16e15b08849","arxiv_id":"2504.15944","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Deep neural networks estimating event intensities from mixing covariates achieve a prediction-error rate near the nonparametric optimum, and a structured two-step marked ratio estimator outperforms a single network in simulations.","lead":"The authors prove a mathematical guarantee for how well deep learning can predict when events will occur from historical signals, for a wide class of event models. Simulations and stock order data suggest that splitting the prediction into two smaller steps can beat using one big neural network.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Geometric alpha-mixing assumed in Lemma 5.1 conflicts with the long-memory trade-sign covariate used in the real-data application, so the stated rate bound is not justified for the LOB analysis.","rationale":"I agree with the reader's weakest_assumption. The core oracle inequality is a conditional statement, and its probabilistic engine is the alpha-mixing large-deviation bound; the geometric decay of the mixing coefficients is what makes Lemma 5.1 and the moment bounds in Lemmas 5.3 and 5.4 work. The real-data section uses a covariate that the paper itself describes as long-memory, which is in direct tension with that premise. This does not make the theorem internally inconsistent, but it does mean the headline rate is not transferable to the LOB application without an additional argument. I also note the optimization-error assumption in Theorem 3.1 as a secondary caveat: the rate assumes Delta_T is small, which is nontrivial for non-convex network training, but the mixing gap is the more load-bearing issue. A direct empirical check of the mixing coefficient would settle the concern. The reader's conditional verdict remains appropriate, so I recommend no change.","tokens_in":30867,"tokens_out":12150,"duration_ms":111701,"concrete_test":"Estimate the alpha-mixing coefficients of the trade-sign covariate X1 on the 22-day Total Energies dataset described in Section 4.2, for lags k = 1 through at least 10^4 events, with block-bootstrap confidence intervals. If the estimated coefficients do not decay as C e^{-\\gamma k} (for example, if they decay like a power law or remain above the exponential envelope at large lags), then the mixing premise of Lemma 5.1 fails for the application, and the rate bound in Theorem 3.1 cannot be claimed for the LOB analysis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The rate bound in Theorem 2.4 is driven by Lemma 5.1, a large-deviation estimate for additive functionals of X whose proof uses the geometric alpha-mixing condition on X through the covariance bound (5.23). Section 2 assumes this mixing decay; without it, the O((log T)^2 log N_T) fluctuation term in (2.4) and hence the rate in Theorem 3.1 do not follow. Section 4.2, however, defines X1 as the sign of the last trade and explicitly states that trade signs have long memory. For a stationary binary series, empirically hyperbolic autocorrelation decay is incompatible with geometric alpha-mixing, so the covariate process in the motivating LOB application fails the theorem's premise. This is not an internal inconsistency: the theorem remains valid for processes satisfying the mixing condition, such as the OU covariates in the simulation. It is a load-bearing gap between the theoretical rate and the real-data analysis, where the stated rate cannot be invoked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies nonparametric estimation of the intensity of a multivariate counting process when the intensity is a function of a covariate process, and uses deep ReLU networks as the estimating class. The loss is a general contrast function Ψ_T(a,b), and the risk is the expected contrast evaluated at the true intensity. Under a compatibility condition (2.1) relating the contrast to an L2-type discrepancy, a metric condition (2.3), and a geometric α-mixing assumption on the covariate process, Theorem 2.4 gives an oracle inequality: R_T is bounded by the optimization error Δ_T, the approximation error inf_{FT} h^{-1}E[Ψ_h(a,b)-Ψ_h(a*,b*)], and a complexity term of order (1+F^2)(T^{-1}(log T)^2 log N_T + δ). Theorem 3.1 combines this with Schmidt-Hieber-type network approximation and covering-number estimates to obtain R_T ≤ C φ_T L (log T)^4. The paper then introduces a marked ratio model, studies it in simulations with Ornstein-Uhlenbeck covariates, and applies it to Euronext limit order book data using imbalance, spread, and the sign of the last trade as covariates.","tokens_in":31086,"tokens_out":8341,"duration_ms":85230,"significance":"If the rates hold, this is a useful extension of the i.i.d. deep-learning regression theory of Schmidt-Hieber to point-process intensities under dependence. The oracle inequality cleanly separates optimization error, approximation error, and network complexity, and the proof of Theorem 2.4 is detailed and structurally coherent, using external large-deviation and covariance inequalities in a plausible way. The compatibility condition is checked for the principal examples, and the marked-ratio decomposition is practically motivated. The simulation design is transparent and the real-data application is realistic. The main caveats are the mismatch between the geometric-mixing premise and the long-memory last-trade-sign covariate used in the real-data analysis, and the fact that the multi-output networks used in the simulations are not explicitly covered by the p_L=1 statement in Section 3.","major_comments":[{"comment":"The geometric α-mixing condition on the covariate process X is load-bearing for Theorem 2.4, entering through Lemmas 5.1, 5.3 and 5.4. In the real-data application, Section 4.2 includes X^1_t as the sign of the last trade and states that trade signs have long memory. A stationary binary process with hyperbolically decaying autocorrelation cannot be geometrically α-mixing, so the stated rate bound cannot be invoked for the LOB analysis. The authors should either verify the mixing condition for the actual covariates used (or at least discuss its plausibility empirically), or explicitly state that the real-data section is heuristic and outside the scope of the rate theorem.","section":"Section 4.2 and Theorem 2.4"},{"comment":"The mixing condition is printed as α_X^h(k) ≤ γ^{-1} e^{-γ h} for all k ∈ Z_+, with the right-hand side independent of k. As written, this does not express decay of the α-mixing coefficients in k. The covariance bound (5.23) and the subsequent large-deviation estimates require geometric decay in k; the condition should presumably read γ^{-1} e^{-γ k} (or an equivalent). This is a load-bearing assumption in the theorem statement and should be corrected.","section":"Section 2, mixing coefficient definition"},{"comment":"Section 3 states that p_L=1 for the applications in the article, but the one-step and two-step estimation methods in Section 4 use networks with output dimensions 7 and 3, respectively. The softmax normalization couples the output coordinates, and the theorem as stated with p_L=1 does not explicitly cover this multi-output architecture. Please either state a multi-output version of Theorem 3.1 with the corresponding approximation and covering-number bounds, or explain that the simulation estimates each scalar ratio function separately with p_L=1 networks. This does not affect Theorem 2.4, which is architecture-independent, but it affects the claim that Theorem 3.1 directly justifies the numerical experiments.","section":"Section 3 versus Sections 4.1.2-4.1.3"}],"minor_comments":[{"comment":"The simulation comparison reports means over 20 samples without standard errors or confidence bands. Since the two-step method is claimed to be clearly superior, adding error bars or a formal comparison at each horizon would strengthen the empirical claim.","section":"Section 4.1.5 and Figure 6"},{"comment":"The notation for the horizon is confusing: T ∈ T is sometimes used as a continuous time length and sometimes as an integer in conditions such as T ≥ 3 ∨ log N_T. The proof uses T = T/h; please introduce a separate symbol for the number of blocks or clarify the convention in the theorem statements.","section":"Section 2 and Section 5"},{"comment":"There is a typographical error in the definition of R_T^e: the displayed expression has a mismatched expectation bracket. Please correct it.","section":"Section 5.1, Eq. (5.1)"},{"comment":"The baseline intensity λ_0(t)=1+cos(2πt) has period 1, but the period h of the periodic stationarity is not specified in the simulation. Please state the value of h used in the experiments and its relation to the theoretical block length.","section":"Section 4.1.1"},{"comment":"The minimax-optimality claim is justified through the independent-input submodel. Since the theorem's class contains independent covariates, the lower-bound argument is valid, but the wording could be made more precise by saying the lower bound is obtained for the submodel with periodically independent covariates.","section":"Remark 3.2(ii)"}],"recommendation":"major_revision","confidential_remarks":"The theoretical core appears sound and within the scope of the journal. The main risk is the application-level gap: the rate theorem requires geometric α-mixing, while the real-data covariate process is acknowledged to be long-memory. I would be willing to review a revised version that either verifies the mixing condition for the real covariates or clearly separates the theoretical rate claims from the empirical application. The multi-output network mismatch in Section 3 also needs attention."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: this paper gives a non-asymptotic oracle inequality for deep neural network estimators of point process intensities when the covariate process is alpha-mixing, and then shows that a two-step marked-ratio estimator beats a one-step estimator in simulation. That is a real extension of Schmidt-Hieber's framework, and the proof of Theorem 2.4 is structurally coherent. The large-deviation argument via Merlevède et al. and the covariance bound from Rio are used cleanly; the compatibility condition is checked for the likelihood, ratio and marked examples; the covering-number step goes through as expected. I found no internal contradiction.\n\nThe empirical section is honest in design: 20 replications per horizon, means reported, and a robustness heatmap over network shapes. The two-step advantage shows up consistently in L2, L∞ and empirical risk, and the effect sizes look meaningful. The real-data curves on Euronext order flow are a reasonable demonstration.\n\nThe soft spots concern scope, not the core theorem. First, no error bars or spread across the 20 samples, no baseline comparison to a Hawkes or parametric model, and no training details or code. That limits reproducibility but does not weaken the theoretical claims. Second, and more important, the rate theorem assumes geometric alpha-mixing of the covariate process, while the real-data section uses the sign of the last trade as a covariate, a series the paper itself calls long-memory. A stationary binary series with hyperbolic autocorrelation decay cannot satisfy geometric alpha-mixing, so the rate cannot be invoked for the LOB application. The theorem remains valid for processes like the OU covariates in the simulation; the gap is between the theory and the motivating application. The authors should either weaken the dependence assumption or explicitly label the real-data analysis as heuristic. Third, Theorem 3.1 contains an unverified assumption on the optimization error Delta_T; that is standard in this literature, but it should be flagged. Minor typos (\"nLN2,2\", \"c` agl` ad\") also need cleaning.\n\nVerdict: worth a serious referee. The theorem is likely to be useful to statisticians working on neural point process models, and the two-step structured estimator is a practical contribution. I would send it to review with a request to address the mixing/application gap and to add error bars or code.","headline":"This paper is a solid extension of Schmidt-Hieber's deep learning theory to point process intensities under alpha-mixing, with a genuine two-step marked-ratio estimator, but the stated rate does not cover the long-memory covariate used in the real-data application.","tokens_in":31605,"tokens_out":4214,"would_cite":true,"duration_ms":34884,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M09","60G55","62G05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves an oracle inequality showing that deep neural networks can predict point-process intensities at a near-minimax rate, and demonstrates that a marked ratio model beats a joint model on simulated and real limit order book…","keywords":["deep learning","point process","marked ratio model","oracle inequality","mixing covariates","limit order book","prediction risk","convergence rate"],"falsifier":"Simulate the marked ratio model with an intensity of the same functional form but drive one covariate by a long-memory process such as fractional Ornstein-Uhlenbeck, fit the two-step network, and measure the empirical risk at growing horizons: if the risk decays materially slower than the predicted rate $\\phi_T$ up to logarithmic factors, the geometric-mixing assumption is essential to the bound. Alternatively, estimate the $\\alpha$-mixing coefficients of the real book-imbalance and trade-sign series and check whether they decay exponentially.","tokens_in":30660,"feed_emoji":"📈","tokens_out":7624,"duration_ms":69272,"temperature":0.7,"pith_summary":"Deep learning is usually analysed for i.i.d. data; this paper extends the theory to counting processes whose jump intensity is a bounded function of a possibly mixing covariate process. It proves an oracle inequality that decomposes the expected prediction error of a fitted intensity model into optimization error, approximation error, and network complexity, and then specialises the bound to sparse ReLU networks to obtain a convergence rate that is minimax-optimal up to logarithmic factors. The paper also introduces a deep-learning marked ratio model, in which the intensity factorizes into a baseline, a type intensity, and mark probabilities. Simulation evidence shows that estimating the factorized model in two separate ratio steps beats estimating the eight-dimensional process in one step, and the same two-step model gives accurate fits to real limit order book data.","feed_headline":"Deep nets predict point-process intensities at near-minimax rate","feed_subtitle":"An oracle inequality splits prediction error into optimization, approximation, and complexity; marked ratio models fit real order flow.","key_machinery":"The central object is the contrast function $\\Psi_T(a,b)=-\\int_0^T a(X_t)\\cdot dN_t+\\int_0^T b(X_t)\\,dt$, whose population expectation is used to define the risk. A compatibility condition makes the discrepancy $U(x)=-\\lambda^*(x)\\cdot(a(x)-a^*(x))+b(x)-b^*(x)$ nonnegative and comparable to the squared error, so minimizing $\\Psi_T$ is a legitimate proxy for prediction. The proof then proceeds through a large-deviation lemma for additive functionals of geometrically $\\alpha$-mixing processes, which controls the empirical-process terms, and a $\\delta$-net covering-number bound for the neural network class, which controls complexity. For the applications, the marked ratio model factorizes the intensity into a baseline, a type component, and mark probabilities; the paper contrasts a one-step estimator of the joint marked ratio with a two-step estimator that fits each factor separately.","core_discovery":"The paper establishes that the expected prediction error of the empirical contrast estimator for a point process with periodically stationary, geometrically $\\alpha$-mixing covariates obeys $$R_T \\le 2\\Delta_T + 2\\inf_{(a,b)\\in\\mathcal{F}_T} $h^{{-1}}$\\mathbb{E}[\\Psi_h(a,b)-\\Psi_h(a^*,b^*)] + C_0(1+$F^{2}$)($T^{{-1}}$(\\log T)^2\\log N_T + \\delta),$$ where $\\Delta_T$ is the optimization error, the infimum is the approximation error, and $N_T$ is the covering number of the model class. In the deep-learning case with sparse ReLU networks whose size grows with $T$, this yields $R_T \\le C\\phi_T L(\\log T)^4$, with $\\phi_T = \\max_i T^{-2\\beta_i^*/(2\\beta_i^*+t_i)}$ the effective smoothness rate, under the condition that the optimization error stays at the same order. The same framework covers Cox-type models, marked point processes, and multivariate point processes as special cases. The paper's empirical claim is that when the true intensity has the marked ratio factorization $\\lambda_{i,k_i}(t)=\\lambda_0(t)\\lambda_i(X_t)p^i_{k_i}(Y_t)$, fitting the factors separately gives lower risk than estimating the joint intensity, and this holds both in simulation and on high-frequency limit order book data.","pith_inferences":["The paper stops short of testing the rate theorem on data; a direct check would be to simulate a long-memory covariate process with a known intensity and compare the empirical risk decay to the predicted $\\phi_T$ rate.","The factorized-marked-ratio principle suggests a general design rule for point-process deep learning: encode known multiplicativity in the architecture. One can test this beyond the ratio model, for instance in state-dependent Hawkes or queue-reactive models where a similar factorization is plausible.","The theorem assumes geometric $\\alpha$-mixing, but the paper's real-data covariate, the sign of the last trade, is described as long-memory. Extending the large-deviation lemma to slowly mixing or long-range dependent inputs would be the natural next step and would determine whether the stated rate applies to the motivating application."],"forward_implications":["Any intensity model estimated by minimizing $\\Psi_T$ inherits the same three-term oracle bound, so the theory covers non-network families as well as deep networks.","With sparse ReLU networks of growing size, the risk converges at rate $\\phi_T L(\\log T)^4$, which is minimax-optimal up to logarithmic factors; the extra $(\\log T)^4$ over the i.i.d. case is the price of the mixing dependence.","The two-step marked ratio estimator gives lower $L^2$, $L^\\infty$, and empirical-risk errors than the one-step estimator across simulation horizons, and the advantage persists across many network widths and depths.","On real limit order book data, the fitted deep marked ratio model reproduces known microstructure regularities: imbalance and spread shape the probability of buy versus sell and price-changing versus non-price-changing market orders.","Because the model is nonparametric, the same network architecture can in principle capture nonlinear covariate effects that fixed parametric families such as exponential intensities cannot express."],"supporting_citations":[{"why":"supplies the sparse ReLU network class, the approximation bound for Hölder-composite functions, and the covering-number estimate used to reach Theorem 3.1.","marker":"[22]"},{"why":"provides the Bernstein-type large deviation inequality for strongly mixing sequences that underlies Lemma 5.1 and removes the empirical-process terms.","marker":"[13]"},{"why":"supplies the covariance inequality used to bound the variance term in the large deviation estimate.","marker":"[21]"},{"why":"introduces the Cox-type ratio model for limit order book intensities that the deep-learning ratio estimator generalizes.","marker":"[17]"},{"why":"introduces the marked ratio model and the two-step estimation idea that the simulation and data application compare.","marker":"[18]"},{"why":"supplies an exponential inequality for martingales used to control the second moment of the martingale term $M_T$.","marker":"[24]"}],"fun_headline_variants":["Deep nets hit near-optimal point-process rates","Oracle bound for deep point-process prediction","Marked ratio nets beat joint intensity on order flow","Near-minimax deep learning for point processes","Deep point process model with marked ratio wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the covariate process $X$ is geometrically $\\alpha$-mixing, meaning its dependence at lag $k$ decays like $e^{-\\gamma k}$; if real covariates such as the sign of the last trade have long memory, the stated rate bound does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Deep nets hit near-optimal point-process rates","Oracle bound for deep point-process prediction","Marked ratio nets beat joint intensity on order flow","Near-minimax deep learning for point processes","Deep point process model with marked ratio wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000136,"raw_usage":{"total_tokens":1120,"prompt_tokens":895,"completion_tokens":225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":156}},"tokens_in":511,"tokens_out":225,"duration_ms":2595,"temperature":1.0,"reasoning_tokens":156,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:14:24.067740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the marked ratio model with an intensity of the same functional form but drive one covariate by a long-memory process such as fractional Ornstein-Uhlenbeck, fit the two-step network, and measure the empirical risk at growing horizons: if the risk decays materially slower than the predicted rate $\\phi_T$ up to logarithmic factors, the geometric-mixing assumption is essential to the bound. Alternatively, estimate the $\\alpha$-mixing coefficients of the real book-imbalance and trade-sign series and check whether they decay exponentially.","supporting_citations":[{"cited_title":"In: High dimensional probability V: the Luminy volume, vol","cited_arxiv_id":null,"evidence_quote":"provides the Bernstein-type large deviation inequality for strongly mixing sequences that underlies Lemma 5.1 and removes the empirical-process terms."},{"cited_title":"Springer (2017)","cited_arxiv_id":null,"evidence_quote":"supplies the covariance inequality used to bound the variance term in the large deviation estimate."},{"cited_title":"Quantitative Finance pp","cited_arxiv_id":null,"evidence_quote":"introduces the Cox-type ratio model for limit order book intensities that the deep-learning ratio estimator generalizes."},{"cited_title":"Japanese Journal of Statistics and Data Science 5(1), 1–39 (2022)","cited_arxiv_id":null,"evidence_quote":"introduces the marked ratio model and the two-step estimation idea that the simulation and data application compare."},{"cited_title":"SIAM (2009)","cited_arxiv_id":null,"evidence_quote":"supplies an exponential inequality for martingales used to control the second moment of the martingale term $M_T$."}],"review_version":1}