{"id":"fa03b679-54a8-429c-a637-b23c5cb1746c","arxiv_id":"2607.11177","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Transformer and LSTM models recover FHP parameters (μ,γ,α,β) from inter-arrival sequences with lower MSE than classical MLE on synthetic data and pass distributional checks on AAPL and 911 data.","lead":"Neural nets (LSTM and Transformer) estimate Fractional Hawkes Process parameters from inter-arrival times without likelihoods, beating a weak MLE baseline on synthetic data. The method is then applied to stock trades and 911 calls via predictive simulation checks.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"Headline superiority over MLE rests on an undiagnosed, numerically fragile classical baseline whose MSE of ~2.8 is implausible for parameters in (0,2].","rationale":"The reader correctly isolates the weakest link: the MLE baseline in §4.1/Table 1 is the sole quantitative support for the claim that the neural estimators are superior. All other issues (metric inconsistencies between abstract and Table 1, R^{2} ≈ 0.28 in the ablations, purely qualitative real-data checks) are secondary; they would not overturn the paper if the classical comparator were shown to be competent. Because the concern is already flagged and the recommended verdict remains CONDITIONAL pending a strengthened baseline, no adjustment is required. The concrete multi-start re-optimization test would settle the issue cleanly without requiring new architectures or data.","tokens_in":15562,"tokens_out":571,"duration_ms":20246,"concrete_test":"On the identical 20 % synthetic test set of §4.1, re-optimize the log-likelihood (Eq. 3) with a multi-start L-BFGS or differential-evolution solver that uses a high-precision Mittag-Leffler routine, record the fraction of convergent runs, and recompute MSE only on those runs. If the conditional MLE MSE falls below ~0.5 (or the gap shrinks by more than half), the reported outperformance is an artifact of the original optimizer rather than a genuine advantage of the neural map.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central synthetic claim (Transformer/LSTM MSE 0.1634/0.1752 vs MLE 2.8032 in §4.1 and abstract; Table 1 gives 0.117/0.140 vs 2.901) requires that the “numerically stabilized MLE” be a competent classical estimator. The paper only retains convergent runs after clipping and log stabilization, supplies no multi-start protocol, no Mittag-Leffler evaluation accuracy, no fraction of discarded trajectories, and no recovery diagnostics on the same synthetic draws used for the neural nets. An MSE of ~2.8–2.9 on a four-dimensional parameter vector whose components lie mostly in (0,2] implies average absolute errors larger than the typical parameter scale itself; that magnitude is the signature of optimizer collapse on the non-Markovian Mittag-Leffler likelihood (Eq. 3) rather than of an informative baseline. Because the entire “significantly outperform classical MLE” headline is measured against this single implementation, any systematic numerical failure of the baseline collapses the claimed gap.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes NeuroMem-FHP, a likelihood-free supervised framework that maps fixed-length sequences of inter-arrival times from a fractional Hawkes process (Mittag-Leffler kernel) to the four parameters (μ, γ, α, β) using an LSTM and a Transformer. Synthetic paths are generated by thinning over broad uniform parameter ranges; networks are trained by MSE against ground-truth parameters with positivity/interval constraints on the outputs. On held-out synthetic data the authors report substantially lower MSE than a numerically stabilized MLE baseline (Transformer 0.1634 / LSTM 0.1752 vs MLE 2.8032 in the abstract and §4.1; Table 1 gives related figures). Ablations examine batch size, epochs, hidden size and sample size. On AAPL NBBO and Montgomery County 911 inter-arrival windows the trained Transformer produces parameter estimates from which simulated sequences approximately match empirical ECDF, survival tails and ACF; a window-size sensitivity study is also reported.","tokens_in":15916,"tokens_out":1392,"duration_ms":12416,"significance":"If the synthetic superiority claim is robust, the work supplies a practical, likelihood-free estimator for a non-Markovian self-exciting process whose Mittag-Leffler likelihood is known to be numerically delicate, and it is among the first deep-learning treatments of fractional Hawkes parameter recovery. The real-data predictive checks (ECDF/survival/ACF match under the maintained FHP model) and the window-size analysis are useful applied contributions for high-frequency event modeling. The contribution is primarily empirical and engineering-oriented rather than theoretical; its lasting value hinges on whether the MLE baseline is a competent classical estimator and on clearer reporting of absolute recovery quality (R², per-parameter errors).","major_comments":[{"comment":"The headline claim that both neural models “significantly outperform classical MLE” (§4.1, abstract, Figure 3, Table 1) rests on a single numerically stabilized MLE implementation whose reported MSE of ~2.8–2.9 is implausibly large for a four-dimensional parameter vector whose components mostly lie in (0,2]. The paper retains only convergent runs after clipping and log stabilization, but supplies no multi-start protocol, no Mittag-Leffler evaluation accuracy, no fraction of discarded trajectories, and no recovery diagnostics on the same synthetic draws used for the nets. An MSE of that magnitude is the signature of optimizer collapse on the non-Markovian likelihood (Eq. 3) rather than of an informative baseline. Without a competent classical comparator (or an explicit diagnosis of MLE failure rates), the claimed gap cannot be interpreted as intrinsic superiority of the neural map.","section":null},{"comment":"Ablation metrics in §4.2 / Figure 4 show R² peaking near ~0.28 (and remaining in the 0.22–0.28 range across batch size, epochs, hidden size and sample size). That implies most of the parameter variance remains unexplained even while MSE is framed as low. The abstract and §4.1 lead with MSE alone and do not reconcile the low R² with the claim of accurate recovery of (μ, γ, α, β). Per-parameter bias/variance or R² (especially for the fractional memory parameter β) should be reported so that absolute identification quality can be assessed.","section":null},{"comment":"Internal inconsistencies in experimental protocol undermine reproducibility of the synthetic results. Sequence length is stated as T=50 in the Transformer architecture description (§3.2) but T=100 in data generation and results (§3.3, §4.1); γ ranges differ between §3.3 (U(0.1,2.0)) and §4.1 ([0.10,5.0]); Table 1 MSE numbers (0.117 / 0.140 / 2.901) do not match the abstract/§4.1 figures (0.1634 / 0.1752 / 2.8032); and the LSTM description oscillates between a single FC layer and an intermediate ReLU FC layer. These discrepancies must be resolved and a single, fully specified protocol fixed before the quantitative claims can be trusted.","section":null}],"minor_comments":[{"comment":"Abstract and opening sentence contain grammatical errors (“Experiments on synthetic data that both neural models…”; “The proposed framework is also on two real-world…”).","section":null},{"comment":"§3.1 contains duplicated and contradictory statements of the output activations (Softplus/Sigmoid ranges and the 0.9 scaling appear twice with slightly different wording).","section":null},{"comment":"Figure 3 caption and surrounding text should state whether MSE is averaged over the four parameters, over samples, or both, and whether it is computed on the constrained or unconstrained outputs.","section":null},{"comment":"Real-data preprocessing (log10 transform, exclusion of gaps >100 s) is reasonable but should be justified against the synthetic training distribution so that zero-shot transfer assumptions are explicit.","section":null},{"comment":"Table 1 lists R² in the prose description but the table columns do not include an R² column; align text and table.","section":null},{"comment":"Several references and arXiv preprints are recent and relevant; a short comparison to other likelihood-free or neural Hawkes estimators would help position the contribution.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core idea is publishable after the MLE baseline is either repaired or honestly diagnosed as failing, and after protocol inconsistencies and low R² are addressed. I would not accept the current “significantly outperform MLE” framing without that work. Scope is appropriate for a methods/applications venue in statistical ML or computational statistics; novelty is incremental but useful if the empirical claims are cleaned up."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean methods paper that does one new thing: trains LSTM and Transformer regressors from fixed windows of FHP inter-arrival times to (μ,γ,α,β). That specific application is not in the prior neural-Hawkes or fractional-SDE literature they cite, so the result is new even if the architectures are not.\n\nWhat works. On synthetic data the nets recover parameters with low MSE and the Transformer edges the LSTM, which is the expected pattern for long-memory kernels. The real-data section is honest predictive validation: estimate on AAPL NBBO and MontcoAlert 911 windows, resimulate, and show ECDF, survival, and ACF match. Window-size sensitivity is also done properly. Ablations are standard and readable. Citations cover the right fractional-Hawkes and neural-estimation papers.\n\nSoft spots, in proportion. The abstract and §4.1 claim “significantly outperform classical MLE” with MSE 0.16–0.17 vs 2.8. That MLE number is not credible for parameters mostly in (0,2]; it is the signature of optimizer collapse on the Mittag-Leffler likelihood (Eq. 3). They keep only convergent runs after clipping, give no multi-start protocol, no fraction discarded, and no recovery diagnostics on the same draws. Table 1 also has different MSE numbers from the abstract, and ablation R² tops out near 0.28, so most parameter variance is still unexplained. Real-data checks are qualitative only; no uniqueness or uncertainty. No code or data released. These are fixable presentation and baseline issues, not load-bearing math errors.\n\nWho it is for: people who already work with fractional or long-memory point processes and need a practical estimator. A serious referee should see it; the contribution is real once the MLE comparison is either fixed or demoted. I would engage, cite the synthetic maps if I need an FHP baseline, and ask for code.","headline":"Usable first neural estimator for fractional Hawkes parameters; the headline MLE gap is almost certainly inflated by a broken classical baseline, but the synthetic maps and real-data predictive checks still look real.","tokens_in":16495,"tokens_out":499,"would_cite":true,"duration_ms":7656,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G55","60G22","62M09","68T07"],"pacs":[],"model":"grok-4.5","headline":"LSTM and Transformer networks recover fractional Hawkes parameters from inter-arrival sequences far more accurately than maximum likelihood, without ever evaluating the likelihood.","keywords":["fractional Hawkes process","parameter estimation","Transformer","LSTM","likelihood-free inference","long-range dependence","Mittag-Leffler kernel","point processes"],"falsifier":"Re-optimize the same synthetic 100-event windows with a multi-start or gradient-based MLE that converges on every replicate; if that MLE’s MSE then falls near the neural values of 0.16–0.18, the claimed superiority disappears.","tokens_in":16464,"feed_emoji":"📈","tokens_out":889,"duration_ms":20576,"temperature":0.7,"pith_summary":"The paper introduces NeuroMem-FHP, a supervised deep-learning framework that maps fixed-length windows of inter-arrival times drawn from a fractional Hawkes process directly onto the four parameters that govern baseline rate, decay, branching, and long-memory strength. Classical maximum-likelihood estimation of the same Mittag-Leffler kernel is numerically costly and unstable; the networks avoid that cost entirely. On large synthetic test sets the Transformer reaches an MSE of 0.16 while stabilized MLE remains above 2.8; the same trained models, applied zero-shot to AAPL trade timestamps and Montgomery County 911 calls, generate simulated sequences whose empirical distribution, heavy tails and autocorrelation structure match the real data. A reader who works with high-frequency event streams therefore obtains a practical, likelihood-free route to long-memory point-process modeling.","feed_headline":"Transformers beat MLE at fitting fractional Hawkes processes","feed_subtitle":"Likelihood-free nets recover long-memory parameters and match real trade and 911 data statistics","key_machinery":"NeuroMem-FHP: a supervised neural regression that ingests windowed inter-arrival sequences, encodes them with either gated recurrence or multi-head self-attention, and emits the four FHP parameters under softplus/sigmoid constraints, trained by MSE on thinning-simulated paths.","core_discovery":"When trained on synthetic trajectories of the fractional Hawkes process, both an LSTM and a Transformer learn a direct map from observed inter-arrival windows to the constrained parameter vector (μ, γ, α, β) whose mean-squared error is an order of magnitude lower than that of numerically stabilized maximum-likelihood estimation; the Transformer-estimated parameters further reproduce the ECDF, survival function and ACF of two real high-frequency event series.","pith_inferences":["The dramatic MLE gap may largely reflect numerical fragility of the Mittag-Leffler likelihood rather than pure representational superiority, so a hybrid that uses the network only as a warm start could close the remaining error.","Zero-shot transfer succeeds only when real data lie inside the synthetic parameter box and share the same log-transform and truncation preprocessing.","Extending the output head and simulator to multivariate or marked fractional Hawkes processes would immediately yield a general likelihood-free engine for multi-type event streams."],"forward_implications":["High-frequency event series with long-range dependence can be fitted by a single forward pass of a pre-trained Transformer instead of iterative likelihood maximization.","Once parameters are obtained, realistic synthetic catalogs that preserve empirical tails and autocorrelation become available for stress-testing or scenario generation.","Window length can be treated as a tunable hyper-parameter that trades estimation bias against variance of the recovered memory exponent β.","The same architecture supplies a template for other non-Markovian point processes whose likelihoods are intractable."],"fun_headline_variants":["Transformer cuts FHP param MSE 17× versus MLE","LSTM and Transformer map inter-arrivals to FHP params far better than MLE","Likelihood-free nets recover fractional Hawkes parameters with lower error","Transformers reproduce trade and 911 stats from FHP estimates","NeuroMem-FHP nets beat classical MLE on long-memory Hawkes fitting"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The numerical comparison treats a carefully clipped and filtered maximum-likelihood optimizer as a fair classical baseline rather than an optimizer that systematically fails on the Mittag-Leffler likelihood.","fun_headline_variants_meta":{"raw":{"variants":["Transformer cuts FHP param MSE 17× versus MLE","LSTM and Transformer map inter-arrivals to FHP params far better than MLE","Likelihood-free nets recover fractional Hawkes parameters with lower error","Transformers reproduce trade and 911 stats from FHP estimates","NeuroMem-FHP nets beat classical MLE on long-memory Hawkes fitting"]},"model":"grok-4.5","effort":"low","cost_usd":0.00672,"raw_usage":{"total_tokens":1700,"prompt_tokens":822,"num_sources_used":0,"completion_tokens":98,"cost_in_usd_ticks":67200000,"prompt_tokens_details":{"text_tokens":822,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":780,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":822,"tokens_out":98,"duration_ms":7555,"temperature":1.0,"reasoning_tokens":780,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T06:28:05.356689+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-optimize the same synthetic 100-event windows with a multi-start or gradient-based MLE that converges on every replicate; if that MLE’s MSE then falls near the neural values of 0.16–0.18, the claimed superiority disappears.","supporting_citations":[],"review_version":1}