{"id":"688db10c-64e6-4709-98a5-cf38f10bc4ac","arxiv_id":"2411.17937","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CSF, a two-stage VAE plus spatio-temporal graph convolutional network with a river flow graph as adjacency, reports higher NSE and KGE than STGCN and other baselines for daily streamflow forecasting in the Brazos basin.","lead":"A streamflow forecasting model uses a river network graph as a fixed spatial mask in a graph neural network and reports better scores than several baselines on the Brazos River basin. The paper calls this causal learning, but the graph is a hydrologic prior, not a structure learned from data, and the claimed efficiency advantage over process-based models is not tested.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The manuscript never defines a temporal train/validation/test split, so Table I may reflect in-sample fit rather than forecasting skill; this is the key unverified precondition for the central performance claim.","rationale":"The reader identified the river-flow-graph fidelity (missing dams/reservoirs) as the weakest assumption. I disagree. The graph is a prior; even if it is imperfect, it can still improve accuracy by imposing an upstream-only message-passing structure, so the performance claim does not hinge on the graph being the true causal graph. The performance claim hinges on the evaluation being a genuine out-of-sample forecast. The manuscript's Section V.C never states the temporal split or the number of seeds; the only protocol description is a vague 'rolling training pipeline.' Without knowing whether the test days are strictly future relative to training, the values in Table I could reflect in-sample fit, which would make the comparison irrelevant to forecasting. This is the load-bearing precondition: if the split is missing, no architecture component — graph, VAE, or hierarchy — is validated. The concrete test (explicit split + walk-forward + multi-seed reporting) settles whether the concern lands. The reader's conditional verdict already demands additional robustness information, so the verdict category remains CONDITIONAL, hence 'UNCHANGED.'","tokens_in":12956,"tokens_out":14176,"duration_ms":132809,"concrete_test":"Require the authors to report the exact contiguous split (e.g., train on 1967-1974, validate on 1975, test on 1976-1977) and re-run the full Table I comparison in a walk-forward protocol where every test day is strictly after the last training day, with no overlap and no retraining on test-period observations. Report the metrics for each of at least three random seeds. If the CSF-vs-STGCN margins (0.08 short-range NSE, 0.11 medium, 0.05 long) persist on the future-only test set, the concern is resolved; if the margins shrink or reverse, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section V.D.1, Table I) is that CSF consistently outperforms all baselines. For this to be evidence of forecasting skill, the metrics must be computed on a held-out temporal period. Section V.C ('Experiment Settings') describes only a 'rolling training pipeline' and never states which time periods are used for training, validation, and testing, nor how the 3,650-day record is partitioned. The one-sentence description — 'the model is trained to predict streamflow for the next day, and its prediction is then used as input for forecasting subsequent days' — could describe inference under an expanding-window retraining scheme, an in-sample fit, or a sliding-window validation; the text is ambiguous. If the reported numbers are computed on data that the model was trained on, then the NSE/KGE/VE/rho values in Table I are measures of memorization, not prediction, and the overperformance of CSF over STGCN could be an artifact of model capacity or optimization rather than a genuine forecasting advantage. This concern is distinct from the causal-graph assumption: even if the river-flow graph perfectly encoded true connectivity, the absence of a defined out-of-sample split would leave the headline result unverifiable.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Causal Streamflow Forecasting (CSF), a two-stage model that first learns latent runoff embeddings from meteorological forcings via a variational autoencoder (VAE), then feeds these embeddings into a spatio-temporal graph convolutional network (STGCN) whose spatial aggregation is guided by a river flow graph derived from DEM and HUC data. The model is evaluated on daily streamflow prediction at 73 stations in the Brazos River basin over a 3,650-day period (1967–1977), with short-range (7-day input, 1-day output), medium-range (14-day input, 3-day output), and long-range (28-day input, 7-day output) forecasting tasks. The reported results (Table I) show CSF outperforming MLP, CNN, Conv-LSTM, TCDF, MTGNN, and STGCN on NSE, KGE, VE, and ρ across all horizons; an ablation study (Table II) attributes the gains to the hierarchical network (HN) and river graph (RG) modules.","tokens_in":13120,"tokens_out":3921,"duration_ms":34008,"significance":"If the empirical results are reproducible, the paper makes a useful contribution by demonstrating that a hydrologically informed causal adjacency matrix can guide STGNNs to better streamflow forecasts than purely data-driven graph structures. The two-stage design that disentangles local runoff generation from basin-scale routing is conceptually appealing and mirrors the VIC-CaMa-Flood process chain. The ablation study and the runoff-embedding alignment analysis are valuable components. However, the lack of a clearly defined out-of-sample temporal split, the absence of uncertainty quantification, and the untested efficiency claim substantially limit the strength of the conclusions as currently written.","major_comments":[{"comment":"The manuscript never defines the temporal train/validation/test split. The description of the 'rolling training pipeline' is ambiguous: it could describe expanding-window retraining, sliding-window validation, or in-sample reconstruction, and the text never states which time periods are held out during evaluation. Because the central claim in Section V.D.1 (\"CSF consistently outperforms all baseline models\") rests entirely on the metrics in Table I, these numbers must be computed on a held-out temporal period to be evidence of forecasting skill. Please state explicitly how the 3,650-day record is partitioned, and report results for a fixed test period that is never used for training or validation.","section":"§5.C (Experiment Settings) and Table I"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any table. The reported margins (e.g., short-range NSE 0.86 for CSF vs. 0.78 for STGCN) could lie within run-to-run variation for neural network training, which is stochastic. The claim of consistent outperformance requires either multiple independent runs with variance reporting, or a paired statistical test across stations or time blocks. Without this, the robustness of the comparison cannot be assessed.","section":"§5.D.1 (Forecasting Accuracy) and Table I"},{"comment":"The abstract claims that CSF \"achieves higher computational efficiency compared to traditional simulation methods,\" but no experiment in Section V compares runtime, memory usage, or any efficiency measure against VIC-CaMa-Flood or any process-based model. This is a load-bearing part of the stated contributions; please either add a direct efficiency comparison or remove the claim from the abstract and the contributions list.","section":"Abstract and §2 (Background)"},{"comment":"The river flow graph is treated as the true causal adjacency matrix, but it is generated purely from DEM and HUC data and does not account for artificial structures such as dams, reservoirs, and diversions, which are known to alter actual water routing in the Brazos basin. This assumption is structurally distinct from the performance claim, but it underlies the causal interpretation and the design rationale. The authors should explicitly acknowledge this limitation and, if possible, test sensitivity of the results to the presence of such structures (e.g., by removing or re-routing edges at known dam locations).","section":"§3.B (Causal Adjacency Matrix) and Figure 3b"}],"minor_comments":[{"comment":"The validation of the learned runoff embeddings uses simulated runoff from the VIC model calibrated by the authors' prior work (reference [6]). This introduces a potential circularity because the same prior calibration informs the design. Please clarify that this is a proxy validation, not an independent ground-truth comparison, and consider an alternative validation (e.g., against observed streamflow-derived runoff proxies).","section":"§5.B.1 (Runoff Embedding)"},{"comment":"The total loss in Eq. (4) combines the station-level and prediction losses with a hyperparameter λ, but the manuscript never reports the chosen value of λ or any sensitivity analysis for it. Please provide the hyperparameter settings used in the experiments, including λ, embedding size (mentioned as 8 in Figure 6), and the STGCN architecture details (number of layers, hidden units, learning rate, etc.).","section":"§4.A (Two-Stage Physically Aware Hierarchical Modelling), Eq. (4)"},{"comment":"The description of the rolling training pipeline is insufficiently detailed for reproducing the multi-step forecasts. The sentence \"the model is trained to predict streamflow for the next day, and its prediction is then used as input for forecasting subsequent days\" suggests a recursive strategy, but the medium-range and long-range tasks use output windows of 3 and 7 days, respectively. Please clarify how the 3-day and 7-day outputs are produced (recursive vs. direct multi-step) and whether teacher forcing is used during training.","section":"§5.C (Experiment Settings)"},{"comment":"There are several typographical errors and awkward phrasings, e.g., \"forfor similating\" in Section II, \"hydrologicly\" in Section I, and \"wold\" in Section I. A careful proofreading pass is recommended.","section":"Throughout"},{"comment":"Figure 6 shows the kNN similarity versus embedding size, and Figure 7 shows similarity versus NSE. The text states that embedding size 8 \"provides the best trade-off\" and that higher similarity correlates with better NSE, but no quantitative evidence (e.g., correlation coefficient or standard deviation across runs) is provided. Please add the quantitative relationship or temper the interpretive claims.","section":"§5.D.2 (Runoff Embedding Similarities)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a promising application of causal graph priors to streamflow forecasting, but the experimental reporting is currently insufficient to support the headline claims. The missing temporal split is the most critical issue and is fixable within the scope of the manuscript; I would not recommend rejection if the authors can supply the split definition, add error bars or significance tests, and either test or remove the efficiency claim. I also note that the validation of the runoff embedding relies on the authors' own VIC simulation (reference [6]); while not disqualifying, the authors should be asked to state this dependency explicitly in the main text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's real content is a sensible composite: use a river network graph derived from DEM/HUC as the adjacency prior for an STGCN, with a VAE pretraining stage that produces a runoff-like embedding. The idea is not revolutionary—this group's earlier STCD/STREAMS work and the general STGNN literature point in this direction—but the specific combination on 73 Brazos stations is new, and the ablation table does what it should: removing the river graph or the hierarchical grouping lowers NSE/KGE/VE/rho consistently. Credit where due: the architecture is clearly described, the evaluation metrics are standard for hydrology, and the limitations section is honest about the i.i.d. station assumption, the capped extremes, and the missing human interventions. This is not a sloppy paper.\n\nThe soft spots are real, and one is load-bearing. Nowhere in Section V.C do the authors define which temporal periods are used for training, validation, and test. They describe a 'rolling training pipeline' but never say whether Table I metrics come from held-out days or from the same record the model was trained on. For a forecasting paper, that distinction is the whole ballgame. If the numbers are in-sample fit, the 0.86 vs 0.78 NSE margin is memorization, not skill. The stress-test note is correct and, if the authors cannot clarify the split, the central claim is unverifiable. That alone requires a major revision.\n\nThe secondary issues are proportionate. No error bars or significance tests anywhere, so the margins could be run-to-run noise. No code or data release, which is fixable. The efficiency claim about beating process-based models is asserted, not measured against VIC-CaMa-Flood or any comparable baseline. The 'causal' language is doing more work than the method does: a DEM/HUC-derived flow graph is hydrologic prior knowledge, not a learned or tested causal structure, and the reader's point about dams/reservoirs in the Brazos is exactly the kind of misspecification that undermines the causal interpretation. The VAE embedding is validated against VIC runoff from the authors' own calibrated model, so that alignment number is partly self-referential, though not circular in a damaging way.\n\nWould I referee it? Yes. The paper deserves a serious referee and a revise-and-resubmit, not a desk reject. The right path is: define the temporal split, report variance, release code/data, measure the efficiency claim against a real process model, and soften the causal claims to 'hydrologic prior.' If whatever split they used is held-out, the result will be a solid contribution for the hydrology-ML community; if not, the current version does not support its headline.","headline":"A sensible composite model with an honest ablation, but the missing temporal train/test split makes the headline performance claim unverifiable as written.","tokens_in":13742,"tokens_out":1909,"would_cite":false,"duration_ms":17270,"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 paper reports that feeding a river flow graph into a graph neural network improves streamflow forecasts at all horizons on 73 Texas stations, with short-range NSE rising from 0.78 to 0.86.","keywords":["streamflow forecasting","spatio-temporal graph neural network","causal graph","river flow graph","runoff embedding","variational autoencoder","hydrologic connectivity","Brazos River basin"],"falsifier":"Retrain CSF on the same 73 stations with the causal adjacency matrix randomly rewired while preserving each station's number of upstream neighbors; if short-range NSE stays near 0.86 instead of falling toward the 0.78 of the plain STGCN, the reported improvement is not caused by the river flow graph.","tokens_in":12679,"feed_emoji":"🌊","tokens_out":6198,"duration_ms":53428,"temperature":0.7,"pith_summary":"Streamflow forecasts matter for flood warnings, reservoir operations, and water allocation, but standard graph neural networks treat nearby stations symmetrically and can let downstream information leak into upstream predictions. This paper tries to fix that by using a river flow graph—an explicit map of which stations are upstream of which—as a causal mask inside a spatio-temporal graph neural network. The proposed model, Causal Streamflow Forecasting (CSF), reports the best scores among all tested methods on 73 Brazos River basin stations for short, medium, and long forecast horizons. If the result holds, it gives hydrologists a practical way to inject known physical connectivity into deep learning without giving up computational speed.","feed_headline":"Causal river graph beats standard models at streamflow forecasting","feed_subtitle":"Guidance from upstream-downstream hydrology lifts NSE to 0.86 at 73 Brazos stations, the paper reports.","key_machinery":"The load-bearing object is the river flow graph, a directed adjacency matrix derived from topographic and watershed-boundary data that records which stations are upstream of which. CSF feeds this graph into a two-stage model: a Variational Autoencoder at each station learns a latent 'runoff embedding' from meteorological and static forcing variables, and a Spatio-Temporal Graph Convolutional Network uses the river flow graph as a causal mask so each station aggregates information only from causally relevant upstream neighbors. A hierarchical clustering of stations into sub-drainage groups makes training scalable by batching at the group level.","core_discovery":"The paper claims that streamflow forecasting at a set of river stations improves when a spatio-temporal graph neural network is constrained by a river flow graph rather than by geographic proximity. In the proposed Causal Streamflow Forecasting (CSF) model, the adjacency matrix is built from Digital Elevation Models and Hydrologic Unit Codes so that message passing follows upstream-to-downstream connectivity and ignores stations in different drainage areas. Tested on daily data from 73 stations in the Brazos River basin, CSF reports the highest scores in NSE, KGE, VE, and Pearson correlation across short (1-day), medium (3-day), and long (7-day) forecast horizons; for short range, NSE is 0.86 versus 0.78 for the strongest baseline, STGCN. The ablation study attributes the gain to the combination of the river graph and a hierarchical grouping of stations.","pith_inferences":["Beyond the paper's claims, the river flow graph omits dams, reservoirs, and diversions, so the causal mask may be wrong in regulated basins; a stress test on a dam-heavy river would show whether the method's advantage survives when the graph misstates actual water routing.","Beyond the paper's claims, the runoff embedding is validated only by similarity to VIC-simulated runoff; one could test whether the embedding also predicts observed soil moisture or evapotranspiration to support a physical reading.","Beyond the paper's claims, a direct way to separate the causal-graph contribution from the hierarchical-batching contribution is to rewire the adjacency matrix randomly; if accuracy holds, the gain is not from causality but from regularization or grouping."],"forward_implications":["If the central claim is correct, causal masking with hydrologically defined upstream-downstream edges is a reliable accuracy improvement over proximity-based graphs, not just an interpretability aid.","The reported short-range NSE gain (0.86 vs 0.78) means the design lifts daily forecast skill at a scale useful for water management decisions.","The two-stage decomposition—local runoff embedding plus basin-level routing—offers a faster alternative to full process simulation while retaining physical structure.","The ablation results imply that hierarchical batching and the causal graph are complementary; dropping either one costs accuracy, most visibly at longer horizons.","The same graph-guided recipe should transfer to other basins wherever elevation and watershed-boundary data exist."],"supporting_citations":[{"why":"Defines the STGCN backbone whose spatial and temporal convolutions CSF extends with a causal mask.","marker":"[2]"},{"why":"The VIC model supplies the process-based rainfall-runoff simulation that the station-level VAE stage is designed to mimic.","marker":"[4]"},{"why":"CaMa-Flood supplies the river-routing component of the traditional simulation pipeline that CSF is compared against in spirit.","marker":"[5]"},{"why":"Provides the calibrated VIC simulated runoff used to evaluate the learned runoff embeddings via kNN alignment.","marker":"[6]"},{"why":"TCDF, a causal-discovery baseline, is one of the models CSF must outperform in the forecast comparison.","marker":"[24]"},{"why":"Supplies the variational autoencoder used to learn the latent runoff embeddings at each station.","marker":"[28]"},{"why":"Livneh's dataset provides the daily meteorological forcings that drive both stages of the model.","marker":"[30]"},{"why":"MTGNN is another graph-based baseline whose performance marks the improvement attributable to the river flow graph.","marker":"[37]"}],"fun_headline_variants":["River graph guides neural net to better streamflow forecasts","Causal model uses river network to beat standard AI forecasts","Upstream-downstream links sharpen streamflow predictions","River flow graph boosts streamflow forecast accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The river flow graph is assumed to be the true causal routing of water, even though engineered structures such as dams and diversions can change which stations actually influence which.","fun_headline_variants_meta":{"raw":{"variants":["River graph guides neural net to better streamflow forecasts","Causal model uses river network to beat standard AI forecasts","Upstream-downstream links sharpen streamflow predictions","River flow graph boosts streamflow forecast accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000579,"raw_usage":{"total_tokens":2732,"prompt_tokens":955,"completion_tokens":1777,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":1716}},"tokens_in":571,"tokens_out":1777,"duration_ms":11343,"temperature":1.0,"reasoning_tokens":1716,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:40:39.605326+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain CSF on the same 73 stations with the causal adjacency matrix randomly rewired while preserving each station's number of upstream neighbors; if short-range NSE stays near 0.86 instead of falling toward the 0.78 of the plain STGCN, the reported improvement is not caused by the river flow graph.","supporting_citations":[{"cited_title":"Spatio-temporal graph convolutional net- works: a deep learning framework for traffic forecasting,","cited_arxiv_id":null,"evidence_quote":"Defines the STGCN backbone whose spatial and temporal convolutions CSF extends with a causal mask."},{"cited_title":"A simple hydrologically based model of land surface water and energy fluxes for general circulation models,","cited_arxiv_id":null,"evidence_quote":"The VIC model supplies the process-based rainfall-runoff simulation that the station-level VAE stage is designed to mimic."},{"cited_title":"A physically based description of floodplain inundation dynamics in a global river routing model,","cited_arxiv_id":null,"evidence_quote":"CaMa-Flood supplies the river-routing component of the traditional simulation pipeline that CSF is compared against in spirit."},{"cited_title":"Design principles for engineering wetlands to improve resilience of coupled built and natural water infrastructure,","cited_arxiv_id":null,"evidence_quote":"Provides the calibrated VIC simulated runoff used to evaluate the learned runoff embeddings via kNN alignment."},{"cited_title":"A long-term hydrologically based dataset of land surface fluxes and states for the conterminous united states: Update and extensions,","cited_arxiv_id":null,"evidence_quote":"Livneh's dataset provides the daily meteorological forcings that drive both stages of the model."}],"review_version":1}