{"id":"125336e2-119e-4bf8-953a-8debe4cac926","arxiv_id":"2608.13096","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Flow matching with a tick-relative LOB representation and transformer backbone generates realistic, controllable, and cross-instrument limit order book states at low sampling cost on HKEX data.","lead":"FlowLOB is a flow-matching model that generates realistic limit order book trajectories for multiple Hong Kong stocks, and can switch to a new stock without retraining. Flow matching needs only 10 solver steps, far fewer than diffusion, and conditioning can steer the generated books toward high or low volatility, liquidity, and imbalance regimes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim depends on excluding specialized diffusion samplers; DPM-Solver/DDIM at 10–20 NFE could close the 4–120× gap and should be tested before the headline is accepted.","rationale":"The reader identified the same load-bearing spot: the efficiency comparison restricts diffusion to generic fixed-step solvers and omits specialized samplers it cites. My reading of Sections 3.2 and 4.2 confirms the protocol is internally controlled, so this is not an accusation of unfairness within the stated setup; it is a mismatch between the strength of the practical claim (4–120x cheaper) and the breadth of the sweep. The paper's other contributions, matched training, zero-shot transfer, controllability test, and honest reporting of the 10s failure mode, are real and partly independent of this concern. The controllability results contain several near-ties after rounding, and no error bars are given, but the largest realism gaps are large enough that I do not see that as the primary threat. The absence of code and data makes the proposed DPM-Solver/DDIM check impossible without author cooperation, which reinforces the CONDITIONAL stance: the conclusion is plausible and well-executed, but the headline efficiency number should not be treated as established until the specialized-sampler check is run.","tokens_in":13900,"tokens_out":10553,"duration_ms":102366,"concrete_test":"Retrain or obtain the exact diffusion checkpoint used in Sec. 4.2 and sample it with DDIM (eta=0) and DPM-Solver (second- and third-order) at NFE in {10, 20, 50} on the 0.1s and 1s tasks, recomputing W1 price and W1 volume exactly as in Figure 1 and Table 1. If any diffusion configuration matches or beats FlowLOB-Euler-10's W1 on both marginals at NFE <= 20, then the 4–120x claim should be narrowed to the fixed-step solver family and the practical efficiency contribution should be restated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest quantitative claim is Sec. 4.2's result that FlowLOB with 10 Euler steps is 4–120× closer to real marginals than diffusion at matched NFE. The paper is careful to train both objectives with identical data, architecture, and budget, and it samples both through the same Euler/Heun/RK4 family. But the efficiency conclusion is broader than this protocol supports. DDIM, DPM-Solver, and UniPC are cited in Related Work and are designed to integrate the very same diffusion probability-flow ODE with far fewer function evaluations; they are never included in the sweep. In the Introduction these samplers are dismissed as treating the symptom rather than the problem, but for a practitioner, a 10-step DPM-Solver that reaches FlowLOB's fidelity would erase the practical advantage even if flow paths are straighter in principle. The paper also leaves the diffusion noise schedule (alpha_t, sigma_t) and the exact probability-flow ODE conversion unspecified, so the comparison cannot be independently rerun. This is a scoping problem rather than an internal inconsistency: the fixed-step comparison is internally controlled, but the headline efficient-sampling contribution is not yet robust to the standard diffusion acceleration toolbox.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FlowLOB, a conditional flow-matching model that generates limit order book trajectory windows (twenty price levels, thirty-two future steps) from a transformer backbone conditioned on trend, volatility, liquidity, and imbalance, trained jointly on eight HKEX symbols at 0.1s, 1s, and 10s resolutions using a tick-relative representation. The central claims are: (i) flow matching reaches its best quality at 10 Euler steps, giving a 4-120x lower marginal distributional error than an equally trained diffusion model sampled with the same fixed-step ODE solvers; (ii) at this operating point FlowLOB is more realistic than LOB-S5, LOB-GAN, Hawkes, and zero-intelligence baselines in 30 of 32 cells at 0.1s and 1s, including zero-shot transfer to a held-out symbol; and (iii) replacing conditioning values with high- or low-tail values moves generated books toward the corresponding real tail regime in 42 of 48 counterfactual cells. The paper honestly reports that the realism advantage weakens at 10s and that trend control is less reliable, and it concludes with a stated plan to study scaling behavior.","tokens_in":14107,"tokens_out":4912,"duration_ms":48236,"significance":"If the claims hold, the paper makes a practically useful contribution: a single multi-symbol, zero-shot-transferable LOB generator with a low-NFE sampling operating point and a distributional test of counterfactual validity is exactly what practitioners need for large-scale rollout simulation. The matched flow-versus-diffusion setup, with identical data, architecture, optimizer, and solver family, is a strength, as is the tick-relative preprocessing that enables cross-symbol generalization. The controllability criterion, which measures whether the generated statistic actually moves toward the requested real tail regime, is a meaningful improvement over mere conditioning. The disclosed 10s limitation and the clear statement of the operating point chosen for downstream experiments are also positive signs of methodological honesty. The main weakness is that the headline efficiency result is scoped to a single family of fixed-step solvers and lacks the specialized diffusion samplers the paper itself cites; the evaluation also lacks error bars or significance tests, so several 'best in 30 of 32 cells' claims rest on a single run.","major_comments":[{"comment":"The headline efficiency claim rests on a sampling protocol that excludes the specialized diffusion samplers cited in §2 (DDIM, DPM-Solver, UniPC). The statement in §4.2 that 'at the same sampling compute, its marginal distributional error is 4–120× lower than diffusion' is not yet supported as a practical efficiency claim, because those samplers are specifically designed to integrate the probability-flow ODE at 10–20 NFE and could close much of the gap. The paper dismisses such samplers in the Introduction as 'treating the symptom rather than the problem,' but this is a conceptual argument, not an empirical one; for a practitioner comparing models, a 10-step DPM-Solver that reaches FlowLOB's fidelity would erase the practical advantage. Please include at least one of these samplers (or a comparable accelerated diffusion sampler) in the NFE sweep, or explicitly state and justify why they cannot be applied to the chosen noise schedule and probability-flow ODE.","section":"§4.2, §3.2"},{"comment":"The diffusion baseline is not fully specified, which makes the controlled comparison impossible to reproduce independently. In §3.1 the variance-preserving noising path is written as x_t = α_t x_0 + σ_t ε, but the actual schedule (α_t, σ_t) is never given, and in §3.2 the conversion from the noise prediction to the probability-flow ODE field is described only verbally ('converted into the corresponding probability-flow ODE field'). Please provide the explicit schedule, the exact ODE field, and the integration limits used for the diffusion sampler, so that the fairness of the solver comparison and the reproducibility of the 4–120× ratio can be assessed.","section":"§3.1, §3.2"},{"comment":"All realism and controllability claims are based on a single trained checkpoint and a single evaluation pass. Table 1 reports distances without error bars or significance tests, Table 2 contains rounded ties marked with asterisks, and Figure 3 shows no variance information. The claim in §4.5 that the conclusions are 'not artifacts of a lucky training setting' cannot be supported without multiple seeds or bootstrap confidence intervals over the evaluation data. Please add at least seed-variation or bootstrapped intervals for the key comparisons in Tables 1 and 2 and for the NFE curves in Figure 1; otherwise the 30-of-32 and 42-of-48 headline numbers may overstate the robustness of the differences.","section":"§4.1, §4.3, Table 1 and Table 2"}],"minor_comments":[{"comment":"The abstract states that 'flow matching attains its best quality with only 10 ODE-solver steps,' while §4.2 reports that Flow-Euler-10 is the best point in only four of six panels and uses 'best or nearly best' elsewhere; please align the wording with the figure.","section":"§1, §4.2"},{"comment":"The volume W1 values in Table 1 are orders of magnitude larger than the price W1 values; please state explicitly whether the distances are computed on the log-transformed volume representation introduced in §3.3 or on the raw volume scale, and what units the reported numbers have.","section":"§4.3, Table 1"},{"comment":"The adaptation of the ZI and Hawkes baselines from Kawawa-Beaudan et al. is described only as 'calibration procedures'; please specify the fitted marginal distributions, the Hawkes kernel family, and how the event streams are converted to the fixed-grid LOB representation used for evaluation.","section":"§4.1"},{"comment":"The counterfactual test in §4.4 does not report the number of past windows used for the reference and counterfactual samples, nor the size of the 5% tail sets; adding these details is needed to assess the statistical meaning of the 42-of-48 result.","section":"§4.4"},{"comment":"The bottom-row cross-level correlation matrices are dense and difficult to read at print size; consider enlarging each panel or reporting the correlation differences numerically in a supplementary table.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the journal's scope and the 10s-frequency limitation is disclosed honestly, so I would not reject on that basis. My main concern is the efficiency headline: the fixed-step comparison is internally controlled, but the exclusion of specialized diffusion samplers and the missing noise-schedule details mean the central practical claim could change materially after the requested experiments. The reported controllability test is a nice contribution and should be preserved. It would also be worth asking the authors to make the code and trained checkpoints available, given the reproducibility concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read FlowLOB. It is a genuinely new application: first flow-matching model for limit order books, and the paper does the comparison right in many ways. Both flow and diffusion trained on identical data, architecture, and budget, then sampled through the same Euler/Heun/RK4 family. That is a clean setup, and the result that flow matching reaches its best quality at 10 Euler steps while diffusion needs many more NFE is convincing within that protocol. The realism comparison against two learned and two agent-based baselines is also well executed, with the flow model best in 30 of 32 cells at 0.1s and 1s, including zero-shot on a held-out symbol. The controllability test - checking whether changing a scenario condition actually moves the generated distribution toward the real tail - is a nice addition that few LOB papers bother to do. They also honestly report the 10s failure.\n\nThe soft spots are real but mostly addressable. The efficiency headline is scoped too broadly. The paper cites DDIM, DPM-Solver, and UniPC, then dismisses them as treating symptoms, but never benchmarks them. A practitioner who only needs a 10-step DPM-Solver to close the gap would not care about the straight-path argument. That omission is the main reason I would not take the 4-120x figure at face value. Second, the realism numbers are point estimates with no error bars or significance tests; pooling marginals hides whether differences are meaningful. Third, no code or data, which makes the missing noise-schedule details for the diffusion baseline hard to fill in. None of these are fatal; they are the difference between a strong workshop paper and a fully convincing one.\n\nWho should read this: anyone working on generative models for market microstructure, and people thinking about fast sampling for financial simulators. It deserves a serious referee; with the diffusion-sampler benchmark and some uncertainty quantification it could be a solid conference paper. I would send it out rather than desk reject.","headline":"A solid, honest first flow-matching LOB generator; the efficiency claim is real within the tested solver family, but thinner once specialized diffusion samplers are added.","tokens_in":14635,"tokens_out":2187,"would_cite":false,"duration_ms":19663,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Flow matching can generate realistic limit order book trajectories at 10 ODE-solver steps, with better marginal realism than diffusion and four baselines at high frequencies.","keywords":["flow matching","limit order books","generative models","diffusion models","market simulation","controllable generation","zero-shot transfer","sampling efficiency"],"falsifier":"Run the matched diffusion checkpoint with an accelerated sampler such as DDIM or DPM-Solver at 10 to 20 function evaluations and compare Wasserstein-1 distance to FlowLOB's Euler-10 output on the same test windows; if diffusion matches or beats FlowLOB at comparable NFE, the paper's central efficiency claim fails. A second check: re-run the 10s realism comparison with more training symbols or a longer history; if the flow model's marginals still trail zero-intelligence baselines, the stated boundary at coarse frequency is confirmed rather than a training-size artifact.","tokens_in":13699,"feed_emoji":"📈","tokens_out":12321,"duration_ms":99918,"temperature":0.7,"pith_summary":"FlowLOB is built on a single claim: flow matching can generate realistic limit order book trajectories at a fraction of the sampling cost of diffusion, without sacrificing fidelity or the ability to steer scenarios. The paper trains flow and diffusion models with identical data, architecture, and budget, and samples both through the same fixed-step ODE solvers; flow matching reaches its best quality with 10 Euler steps, while diffusion needs many more function evaluations. At that cheap operating point, the distributions of generated prices and volumes are closer to real Hong Kong Exchange data than those from two learned and two agent-based simulators in most comparisons at the 0.1-second and 1-second frequencies, including on a symbol never seen in training. The conditioning variables also pass a distributional counterfactual test: swapping a scenario value for a tail value moves the generated statistic toward the corresponding real tail regime in 42 of 48 settings. If these results hold, practitioners get a simulator that is cheap enough to roll out at scale, generalizes to new instruments without retraining, and can be steered into rare regimes.","feed_headline":"10 Euler steps beat diffusion for limit order books","feed_subtitle":"Flow matching beats diffusion and agent-based baselines on HKEX realism tests using only 10 network evaluations","key_machinery":"The load-bearing object is the flow-matching velocity field learned along a straight-line interpolation path. For noise $z \\sim \\mathcal{N}(0,I)$ and data $x_0$, the interpolation is $x_t=(1-t)z+t x_0$ and the target velocity is the constant $x_0-z$; the network $v_\\theta(x_t,c,t)$ is trained to regress that target, and sampling integrates the ODE $\\frac{dx_t}{dt}=v_\\theta(x_t,c,t)$ from $t=0$ to $t=1$ with conditioning $c$ held fixed. Because this path is nearly straight, coarse Euler integration already tracks it, which is why 10 function evaluations suffice. The supporting machinery is the tick-relative LOB representation—mid-price changes and level gaps measured in ticks (minimum price increments), volumes as $\\log(V+1)$—which strips instrument-specific scale so one model can train across eight HKEX symbols and transfer to a ninth, and the adaLN-Zero transformer that treats each level–time cell as a token and aligns scenario conditioning channels with the corresponding future time steps.","core_discovery":"Under a matched comparison in which flow matching and diffusion share the same dataset, transformer backbone, training recipe, and family of fixed-step ODE solvers, flow matching reaches its best sample quality with only 10 Euler steps, while the diffusion baseline needs many more function evaluations to approach the same fidelity; at equal sampling cost the flow model's marginal distributional error is 4 to 120 times lower across price and volume features. Using Euler with 10 steps as the default operating point, FlowLOB beats LOB-S5, LOB-GAN, Hawkes, and zero-intelligence baselines in 30 of 32 price/volume metric cells at 0.1s and 1s, on both an in-distribution symbol and a held-out symbol, with a median 4.5 times smaller distance than the best non-flow baseline. The counterfactual test replaces one scenario variable with a high or low 5% tail value and asks whether the generated statistic moves toward the corresponding real tail regime; this succeeds in 42 of 48 cases, with liquidity and imbalance reliably steerable at every frequency. The paper further shows zero-shot transfer to a ninth symbol and that the transformer backbone matters most out of distribution, while noting that at the coarsest 10s frequency the realism advantage mostly disappears.","pith_inferences":["The efficiency comparison is a lower bound on diffusion's speed only if specialized samplers such as DDIM and DPM-Solver, which the paper cites but does not run, are excluded; testing FlowLOB's Euler-10 checkpoint against diffusion with those samplers at equal NFE would settle whether the 4 to 120 times gap is intrinsic or an artifact of the solver family.","The tick-relative representation is likely to transfer to other exchanges and asset classes with different tick sizes and price scales, since it removes instrument-specific units; this is a testable extension rather than something the paper demonstrates.","The counterfactual test changes one condition at a time; a natural extension is joint control, such as high volatility and low liquidity together, or control over derived statistics such as return tails, which the paper does not attempt.","Flow matching's near-straight paths could make the model amenable to consistency-style distillation for even fewer evaluations, but the paper does not explore that; it would be a separate speedup on top of the 10-step operating point."],"forward_implications":["At a fixed compute budget, a generator that is accurate after 10 network evaluations can produce many more simulated days, symbols, and random seeds than one requiring tens or hundreds of evaluations, changing what is feasible for stress testing and reinforcement-learning training.","Because one model covers multiple symbols through the tick-relative representation, a new instrument can be simulated zero-shot, without per-symbol retraining or new calibration data.","Conditioning on liquidity and imbalance reliably shifts generated books into the requested tail regime, so the same generator can be used to produce counterfactual scenarios for what-if analysis.","At 0.1s and 1s, FlowLOB's pooled price and volume marginals are closer to real data than the four baselines on nearly all metrics, while at 10s the advantage weakens, limiting the method's use at coarser resolutions in its current form.","Because flow and diffusion were trained with identical data, backbone, and budget and sampled with the same solvers, the measured quality-cost gap reflects the generative dynamics themselves rather than a difference in architecture or training setup."],"supporting_citations":[{"why":"Defines the flow-matching objective and the linear interpolation from noise to data whose straight path FlowLOB learns and integrates.","marker":"[27, 28]"},{"why":"Defines the diffusion objective and probability-flow ODE that FlowLOB is matched against as the baseline sampler.","marker":"[15, 40]"},{"why":"Supplies the Euler/Heun/RK4 fixed-step solver family and the NFE accounting used to compare flow and diffusion at equal cost.","marker":"[22]"},{"why":"The LOB-GAN conditional GAN baseline whose realism FlowLOB is measured against.","marker":"[8]"},{"why":"The LOB-S5 autoregressive state-space baseline whose realism FlowLOB is measured against.","marker":"[30]"},{"why":"The zero-intelligence trader baselines used as agent-based comparisons.","marker":"[12, 13]"},{"why":"The Hawkes-process order-flow baseline used as agent-based comparison.","marker":"[3, 21]"},{"why":"Provides the calibration procedures used to adapt the ZI and Hawkes baselines to the HKEX dataset.","marker":"[23]"},{"why":"The prior counterfactual LOB diffusion model whose conditioning design FlowLOB builds on and extends to multi-symbol pooled training.","marker":"[44]"},{"why":"The adaLN-Zero transformer backbone that aligns the level-time tokens with the scenario conditioning channels.","marker":"[33, 34]"}],"fun_headline_variants":["10-step flow beats diffusion for order books","Flow matching wins with 10 ODE steps for LOBs","Fast flow generator for realistic limit order books","FlowLOB: 10 steps to control and transfer","Zero-shot realism with flow-generated order books"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that measuring diffusion's sampling cost with the same generic fixed-step ODE solvers used for flow matching is a fair comparison; the paper deliberately excludes the accelerated diffusion samplers it cites, so if those close the function-evaluation gap, the claimed 4 to 120 times efficiency advantage would shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["10-step flow beats diffusion for order books","Flow matching wins with 10 ODE steps for LOBs","Fast flow generator for realistic limit order books","FlowLOB: 10 steps to control and transfer","Zero-shot realism with flow-generated order books"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1393,"prompt_tokens":1076,"completion_tokens":317,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":244}},"tokens_in":692,"tokens_out":317,"duration_ms":3640,"temperature":1.0,"reasoning_tokens":244,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:51:31.466022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the matched diffusion checkpoint with an accelerated sampler such as DDIM or DPM-Solver at 10 to 20 function evaluations and compare Wasserstein-1 distance to FlowLOB's Euler-10 output on the same test windows; if diffusion matches or beats FlowLOB at comparable NFE, the paper's central efficiency claim fails. A second check: re-run the 10s realism comparison with more training symbols or a longer history; if the flow model's marginals still trail zero-intelligence baselines, the stated boundary at coarse frequency is confirmed rather than a training-size artifact.","supporting_citations":[{"cited_title":"Towards Realistic Market Simulations: a Generative Adversarial Networks Approach","cited_arxiv_id":"2110.13287","evidence_quote":"The LOB-GAN conditional GAN baseline whose realism FlowLOB is measured against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The prior counterfactual LOB diffusion model whose conditioning design FlowLOB builds on and extends to multi-symbol pooled training."}],"review_version":1}