{"id":"b573942f-5c7c-4622-9d43-67e57c1d9a24","arxiv_id":"2608.11359","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frozen Chronos-2 foundation model, fed with day-ahead market information and a source-domain gated LoRA adapter, beats zero-shot and vanilla-adapter baselines on four Chinese day-ahead electricity markets under leave-one-market-out evaluation.","lead":"This paper adapts a general time-series AI model to forecast next-day electricity prices in four Chinese markets without training on the target market. A market-information interface plus a small state-dependent adapter cuts average forecast error by a few percent over simpler baselines, offering a template for data-scarce electricity markets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MSMI gains may reflect post-clearing revisions rather than pre-clearing information; the minimum-version policy does not establish that the earliest row precedes clearing.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the day-ahead market-information variables being known and correct over the horizon, with the earliest supply-demand version approximating the pre-clearing information set. I agree this is the most critical point. If the version policy fails, the MSMI interface's large zero-shot gain over Core, and the subsequent Source-LoRA and Gated-LoRA improvements, could be inflated by target-market hindsight information, invalidating the paper's central claim of target-market-free transfer. Other concerns raised by the reader—missing code/data, external validity over four markets, and the marginal pooled p=0.074 for gating—are legitimate but secondary: the gating significance issue is explicitly hedged by the authors, and external validity does not threaten the internal correctness of the main result. The information-leakage concern, by contrast, goes to the validity of the strongest claim itself. The paper's own limitation statement in Section III.C.4 acknowledges the availability assumption but does not prove it holds for the proprietary data; the minimum-version heuristic is a reasonable precaution but not sufficient evidence. Because the reader's verdict is already CONDITIONAL and this concern is among the stated conditions, my stress-test does not move the verdict; it sharpens the condition that should be checked first. The proposed concrete test—verifying publication timestamps and rerunning with provably pre-clearing covariates—would settle whether the concern actually lands. Until that check is performed, the result should not be treated as fully established, which is consistent with the CONDITIONAL verdict.","tokens_in":16271,"tokens_out":4132,"duration_ms":33671,"concrete_test":"Obtain from the data provider the full version history with release/revision timestamps for a random sample of at least 1,000 (market, date, 15-min slot) rows, and verify whether the minimum-version row used in the paper was published before the day-ahead clearing deadline. Then rerun the Table VI MSMI zero-shot and Table II Gated-LoRA configurations using only covariates provably published pre-clearing (e.g., official day-ahead load and renewable forecasts instead of realized values). If the MSMI-vs-Core MAE gain of 7.7% or the Gated-vs-Source gain of 3.05% largely disappears under this substitution, the central transfer claim is an information-leakage artifact rather than transferable skill.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Section III.B.1, Table II) is that the MSMI interface reduces zero-shot MAE from 86.25 (Core) to 79.60 and that gated LoRA reaches 74.63, with the market-information interface identified as the primary source of improvement. This conclusion rests on Section III.A.1's assertion that the deterministic minimum-version policy 'approximates the information that would be available before day-ahead clearing.' That assertion is not validated. The supply–demand table contains load, renewable generation, reserve, maintenance, capacity, and intertie fields, and the authors only state that version 0 is used when available; they provide no timestamp or release-schedule evidence that version 0 precedes the day-ahead clearing deadline. If any of these fields are settled/actual values (e.g., realized load or wind rather than published day-ahead forecasts), then the model receives future target-market information at the forecast origin. This would inflate not only absolute accuracy but also the MSMI-vs-Core ablation gap and the Source-LoRA/Gated-LoRA gains built on MSMI inputs, since the Core interface shares the same leak-prone load/wind/PV fields. Section III.C.4 acknowledges the availability assumption as a limitation, but the assumption is still load-bearing: without it, the headline 'transferable' gain could largely be hindsight leakage rather than forecasting skill.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses target-market-free transfer learning for day-ahead electricity price forecasting. It builds on Chronos-2, a pretrained time-series foundation model, and proposes (i) a multi-source market information (MSMI) interface that feeds seven days of price history plus forecast-horizon supply-demand, reserve, maintenance, capacity, and intertie variables into the frozen backbone; (ii) a source-domain LoRA adapter trained only on the three non-held-out markets under a leave-one-market-out (LOMO) protocol; and (iii) a market-state gate that scales the LoRA update by a linear function of five standardized state scores (net load, renewable share, reserve adequacy, capacity margin, volatility). On four Chinese provincial day-ahead markets, the MSMI interface reduces zero-shot MAE from 86.25 (Core) to 79.60, source-domain LoRA reduces it to 76.98, and the reserve-initialized gated LoRA achieves 74.63 MAE and 137.87 RMSE, a 3.05% MAE improvement over Source-LoRA. The authors also report CRPS, PICP, DM significance tests with HLN correction, an information-interface ablation, and a progressive few-shot adaptation spectrum. They conclude that the task interface is the dominant source of gain and that gated LoRA gives a small but useful additional calibration.","tokens_in":16515,"tokens_out":5672,"duration_ms":50017,"significance":"The paper is significant for the electricity price forecasting community because it tests a concrete, label-free transfer path for data-scarce markets and does so with unusual care for leakage: LOMO folds, a deterministic minimum-version data policy, source-only gate scaler statistics, and multiple control configurations for the gate. The finding that a broader pre-clearing information interface, rather than model-side changes, drives most of the improvement is a useful and falsifiable message. However, the headline numerical gains are only as credible as the assumption that the supply-demand table versions used are truly pre-clearing; the evidence for that assumption is currently absent, and the gated-LoRA advantage over Source-LoRA is modest and only marginally significant. With additional validation of the information timeline and more cautious language, this would be a solid contribution.","major_comments":[{"comment":"The central empirical claim rests on the assertion that the deterministic minimum-version policy 'approximates the information that would be available before day-ahead clearing.' The paper does not provide evidence that version 0 precedes the clearing deadline for any of the four markets, nor that the load, renewable, reserve, maintenance, capacity, and intertie fields are pre-clearing forecasts rather than settled or revised values. Because the MSMI-vs-Core improvement (86.25 to 79.60) and all subsequent Source-LoRA and Gated-LoRA gains are built on these inputs, hindsight leakage in these fields would convert the reported 'transferable' gains into artifacts. Please add release-schedule/timestamp evidence, a lagged-covariate control (feeding information as it would have been known at the forecast origin), or a comparison using only documented pre-clearing variables; without this, the main transferability claim is not established.","section":"III.A.1 and III.C.4"},{"comment":"The evidence that state-dependent gating per se (rather than the initialization prior or small parameterization) explains the improvement is weak. The pooled Gated-vs-Source DM statistic is only marginally significant (p=0.074), Shanxi is not improved (statistic +0.49), and the shuffled-state control (74.873±0.047) is close to the matched reserve-initialized gate (74.633±0.002). The authors acknowledge the small matched-vs-shuffled gap in III.B.1 but elsewhere state that 'reserve-tightness-conditioned calibration can further improve' the adapter as a supported finding. The causal claim should be softened to 'the reserve-initialized gate is a stable, compact calibration that yields a small net improvement,' or additional evidence of state-dependence (e.g., market-level analysis showing larger gate multipliers on tight-reserve days are associated with lower loss) should be provided.","section":"II.D, III.B.2, and Table IV"},{"comment":"The abstract and title foreground the gated-LoRA numbers (74.63/137.87) and the 3.05% reduction, while the body correctly describes the gain as marginally significant and largely limited to LN. This creates a mismatch between the strength of the evidence and the prominence of the claim. Please either report the gating result with explicit significance caveats in the abstract, or reposition the contribution as primarily the MSMI interface plus source-domain LoRA, with gating as a secondary calibration device. The conclusion should also state prominently that gated LoRA degrades interval calibration (PICP@80 falls to 70.23% and Macro-ACD rises to 9.77 p.p.).","section":"Abstract and III.B.3"}],"minor_comments":[{"comment":"The sentence 'These quantities are published or scheduled before market clearing' (the paragraph after Eq. (5)) is an unsupported assertion; please cite market documentation or replace it with a statement about the assumed information set.","section":"II.C"},{"comment":"Table V shows that gated LoRA lowers PICP@80 to 70.23% and increases Macro-ACD to 9.77 p.p.; the paper already notes this, but it should be reflected in the abstract's summary of the gating method.","section":"III.B.3"},{"comment":"The statement that 'representative windows are selected by transparent diagnostic rules' is not sufficient because the rules are not given; please specify them or state explicitly that the panels are illustrative only.","section":"III.B.6"},{"comment":"The gate uses beta=2.0 and clip range [1.0,3.0] in Eq. (12), but these values are introduced only in Section III.A.1; please define them where the gate is first presented.","section":"II.D and III.A.1"},{"comment":"The global-scalar row reports 77.001±0.000, which differs slightly from Source-LoRA (76.98); please explain whether this is a single run or multiple runs collapsed, since a zero standard deviation over three seeds is otherwise surprising.","section":"Table III"},{"comment":"The reported version-0 share (654377 out of 667680 rows) is useful, but the paper should also report the distribution of version numbers by market and by field; this would help readers assess the leakage risk.","section":"III.A.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and technically careful, but the absence of any validation for the pre-clearing information timeline is a serious gap because it underpins the headline transferability claim. I recommend major revision with a request for either timestamp/release-schedule evidence or a control experiment that breaks the suspected leakage (e.g., lagging the covariates). The gating contribution is modest and should be presented with correspondingly modest language."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid, carefully executed transfer-learning study for day-ahead electricity price forecasting, and the leave-one-market-out protocol is the cleanest part of it. But the headline 'transferable' claim rests on an assumption about the data versioning that the paper does not actually validate, and until that is settled I would not treat the MSMI gains as forecasting skill.\n\nWhat's genuinely new: using a frozen Chronos-2 backbone with a richer 'MSMI' input interface (load, renewables, reserve, maintenance, capacity, intertie) under a target-market-free LOMO protocol, plus a six-parameter state-dependent gate that scales a source-trained LoRA. The paper does a lot right: no target-market labels are used for training or model selection; the scaler statistics come from source markets only; the ablations separate interface effects from adapter effects; there are DM tests with HAC and HLN correction; and the authors explicitly say the gating gain is moderate and that interval calibration degrades. That honesty is real.\n\nThe soft spots are in proportion. First and most important, the stress-test note is on target. The 'deterministic minimum-version policy' assumes that version 0 is the pre-clearing version, but the paper provides no timestamp or release-schedule evidence that the earliest row actually precedes the day-ahead clearing deadline. If the supply-demand table contains settled or actual values, then the model sees future target-market information at the forecast origin, and the MSMI-vs-Core gap (86.25 to 79.60) could be leakage rather than transferable skill. This is not a minor detail; it is the load-bearing support for the paper's central claim. The limitation paragraph in III.C.4 mentions that future information is assumed available, but that is not the same as showing the values are pre-clearing.\n\nSecond, there is no code or data. The four Chinese provincial markets are proprietary and selected for comparability, so external validity is genuinely limited. Third, the gated LoRA improvement over Source-LoRA is marginally significant (pooled p=0.074) and concentrated in one volatile market (LN). The paper says as much, so this is a weakness but not a hidden one.\n\nWho is this for? People working on electricity price forecasting and on adapting time-series foundation models to domains with known future covariates. They will get a well-structured evaluation template and a clear warning about information-set leakage.\n\nRecommendation: send it to peer review, but the referee should require evidence about the version policy—ideally release timestamps, or a robustness check that restricts to fields with explicit pre-clearing publication—and preferably code/data for independent verification.","headline":"Careful LOMO study of transferable price forecasting; the MSMI gain hinges on an unvalidated version-policy assumption.","tokens_in":17043,"tokens_out":2405,"would_cite":false,"duration_ms":21727,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62M20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Most of the transferable gain in day-ahead price forecasting comes from feeding a frozen foundation model the pre-clearing supply, demand, reserve, and intertie information; a six-parameter gate adds a smaller state-dependent gain.","keywords":["electricity price forecasting","time-series foundation model","transfer learning","gated LoRA","day-ahead electricity market","leave-one-market-out evaluation","probabilistic forecasting","market information interface"],"falsifier":"Re-run the same leave-one-market-out protocol with the MSMI interface fed the final revised supply–demand rows instead of the earliest available ones, and again with the market-information variables lagged by one day so they could not be known pre-clearing: if either variant reproduces the 7.7% zero-shot MAE gain, the gain is not evidence of using the genuine pre-clearing information set. A complementary check is to deploy the framework on a day-ahead market that publishes only a subset of these fields before clearing; the MSMI advantage should shrink in proportion to the fields actually withheld.","tokens_in":15990,"feed_emoji":"⚡","tokens_out":11308,"duration_ms":84099,"temperature":0.7,"pith_summary":"The paper tries to establish that day-ahead electricity price forecasting can be transferred to a market with no local training labels, and that the thing that makes transfer work is the information fed into the model, not the model itself. The authors feed the frozen Chronos-2 foundation model a 7-day price context plus the pre-clearing supply–demand, reserve, maintenance, generator-capacity, and intertie variables that market participants see before clearing, and fine-tune only about 1% of parameters as a low-rank adapter trained on related markets. Under a leave-one-market-out protocol on four Chinese provincial spot markets, the framework cuts average MAE/RMSE to 74.63/137.87, a 6.24%/7.99% improvement over the market-information-aware zero-shot backbone and a 3.05%/3.52% improvement over the ungated source adapter. The dominant gain is from the market-information interface itself; the state-dependent gate contributes a smaller, volatility-concentrated improvement that survives controls the authors run against global-scalar and shuffled-state alternatives. If correct, this gives newly established or data-scarce markets a deployment path that requires no local labels, and it redirects foundation-model work on market-cleared prices toward interface and adapter-control design rather than backbone redesign.","feed_headline":"Pre-clearing data, not fine-tuning, drive cross-market price forecasts","feed_subtitle":"Giving a frozen model the day-ahead supply, reserve, and intertie picture cut average error by 6.24% on four Chinese markets.","key_machinery":"Two objects carry the argument. The first is the multi-source market-information (MSMI) interface: a task interface that concatenates a 7-day (672-step) price context, its contemporaneous covariates, and the known future values of supply–demand, reserve, maintenance, generator-capacity, and intertie fields over the 1-day (96-step) forecast horizon, exposing the frozen backbone to the same pre-clearing information set a market participant holds. The second is the gated low-rank adapter: the LoRA update $W' = W + \\gamma_\\psi(z_t)\\frac{\\alpha}{r}BA$ is injected into the four attention projections and the patch-output projection of the frozen Chronos-2 weights, and a linear gate maps a five-dimensional standardized market-state vector $z_t$ — net load, renewable share, reserve adequacy, capacity margin, and recent price volatility, signed so that larger values mean tighter operating conditions — to a task-level multiplier clipped to $[1, 3]$. The gate is initialized from the source-trained LoRA checkpoint with only the reserve-tightness weight active, then trained for 300 source-market steps with the LoRA matrices frozen, adding six trainable parameters. The design's work is to let the source adapter strengthen exactly when the operating state tightens, without ever exposing a held-out market's labels.","core_discovery":"The paper's central claim is that a generic pretrained time-series model can forecast day-ahead electricity prices in a market it has never seen, provided the model is handed the economic information set from which those prices are formed. The authors construct a multi-source market-information (MSMI) interface — 672 hours of price history, contemporaneous covariates, and known day-ahead values of load, wind, solar, hydro, nuclear, reserve, ancillary-service, maintenance, generation-capacity, and intertie fields over the 96-step horizon — and show that switching Chronos-2's zero-shot input from a minimal load–wind–PV set to this interface reduces the four-market average MAE from 86.25 to 79.60 without any training. Training a rank-8 LoRA (about 1% of the 120.7M parameters) on the three source markets, with the backbone frozen and the held-out market untouched, brings MAE/RMSE to 76.98/142.90; replacing the adapter's fixed strength with a six-parameter gate that scales the LoRA update by reserve tightness, net load, renewable share, capacity margin, and recent volatility brings it to 74.63/137.87. The authors further claim, from control experiments with a learned global scalar, random gate initialization, and shuffled state–task pairing, that the gate's benefit is genuinely state-dependent rather than a generic calibration artifact. They are explicit that the gated gain is marginal outside the most volatile market and that it degrades 80% interval coverage, recommending a separate interval-calibration layer before operational use.","pith_inferences":["If the interface dominates adaptation, the same recipe likely transfers to other cleared prices with pre-market schedules — real-time and ancillary-service prices, emission allowances, or other auctioned commodities — where the pre-clearing information set is similarly published before the outcome is set.","The 'earliest version' supply–demand policy is only a proxy for the pre-clearing information set; a direct test is to feed the interface the final revised table rows instead. If the revised-version model outperforms the earliest-version model, later revisions carry predictive signal and part of the reported 7.7% interface gain is leakage rather than pre-clearing skill.","The gate's concentration in the volatile fold suggests a threshold hypothesis: the gate's benefit should appear mainly when reserve adequacy or capacity margin falls below a scarcity boundary, so a regime-stratified evaluation (tight vs slack days) would sharpen or refute the reserve-tightness story.","Concretely testable extension: the gate is six scalars; the paper's logic implies that a gate weighting only reserve tightness and net load — the two price-relevant scarcity signals — should match or beat the full five-dimensional gate on sparse data, indicating which market states actually carry transferable information."],"forward_implications":["A newly established or data-scarce market can deploy day-ahead price forecasts immediately with zero local labels: the MSMI interface alone brings Chronos-2 zero-shot average MAE from 86.25 to 79.60.","Domain re-pretraining is unnecessary for this transfer: adapting about 1% of parameters on source markets reaches 76.98/142.90 average MAE/RMSE, and extending the source history from 900 to 1665 days changes results by only −0.24 MAE.","State-dependent adapter calibration works, but its benefit is concentrated: the reserve-initialized gate reaches 74.63/137.87, with the largest drop on the most volatile held-out market and near-zero change on calmer ones.","Probabilistic output is a separate axis: gated LoRA improves CRPS from 62.66 to 61.23 but drops pooled 80% interval coverage from 73.24% to 70.23%, so interval calibration needs an added layer before operational use.","Local few-shot labels are more useful after a source-domain warm start than from the pristine backbone: 30 target-market days give 76.66 average MAE from Source-LoRA versus 78.90 from pretrained Chronos-2."],"supporting_citations":[{"why":"Supplies the pretrained Chronos-2 backbone that the framework freezes and adapts through its interface.","marker":"[19]"},{"why":"Provides the low-rank adaptation mechanism that both the source-LoRA and gated variants build on.","marker":"[28]"},{"why":"Defines the open-access benchmark and rolling-evaluation best practice the experimental protocol follows.","marker":"[4]"},{"why":"Provides the Diebold–Mariano test used for all significance claims on MAE and CRPS differences.","marker":"[34]"},{"why":"Supplies the small-sample Harvey–Leybourne–Newbold correction applied to the Diebold–Mariano statistics.","marker":"[35]"},{"why":"Is the energy-domain foundation-model adaptation the paper contrasts with, since wind power is a physical rather than market-clearing target.","marker":"[25]"},{"why":"Is the closest alternative — a pretrained electricity-price foundation model — whose unexplored adaptation gap this paper targets.","marker":"[24]"},{"why":"Is the original Chronos model, serving as a zero-shot foundation-model baseline in the main comparison.","marker":"[26]"}],"fun_headline_variants":["Market info, not fine-tuning, transfers price forecasts across regions","State-gated LoRA adapts frozen foundation model to new markets","Pre-clearing signals, not retraining, drive cross-market price forecasts","A 1% adapter beats zero-shot in unseen electricity markets","Gated adapter enables transfer of time-series foundation models to power markets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the earliest available supply–demand table row is a truthful, revision-free stand-in for what the market knows before clearing; if load, renewable output, reserves, maintenance, and intertie figures are routinely revised after the fact or only published post-clearing, the large MSMI zero-shot gain would reflect information leakage rather than transferable forecasting skill.","fun_headline_variants_meta":{"raw":{"variants":["Market info, not fine-tuning, transfers price forecasts across regions","State-gated LoRA adapts frozen foundation model to new markets","Pre-clearing signals, not retraining, drive cross-market price forecasts","A 1% adapter beats zero-shot in unseen electricity markets","Gated adapter enables transfer of time-series foundation models to power markets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00087,"raw_usage":{"total_tokens":3885,"prompt_tokens":1179,"completion_tokens":2706,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":795,"completion_tokens_details":{"reasoning_tokens":2616}},"tokens_in":795,"tokens_out":2706,"duration_ms":64973,"temperature":1.0,"reasoning_tokens":2616,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:12:02.386200+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same leave-one-market-out protocol with the MSMI interface fed the final revised supply–demand rows instead of the earliest available ones, and again with the market-information variables lagged by one day so they could not be known pre-clearing: if either variant reproduces the 7.7% zero-shot MAE gain, the gain is not evidence of using the genuine pre-clearing information set. A complementary check is to deploy the framework on a day-ahead market that publishes only a subset of these fields before clearing; the MSMI advantage should shrink in proportion to the fields actually withheld.","supporting_citations":[{"cited_title":"Forecasting day- ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark,","cited_arxiv_id":null,"evidence_quote":"Defines the open-access benchmark and rolling-evaluation best practice the experimental protocol follows."},{"cited_title":"Comparing predictive accuracy,","cited_arxiv_id":null,"evidence_quote":"Provides the Diebold–Mariano test used for all significance claims on MAE and CRPS differences."},{"cited_title":"Testing the equality of prediction mean squared errors,","cited_arxiv_id":null,"evidence_quote":"Supplies the small-sample Harvey–Leybourne–Newbold correction applied to the Diebold–Mariano statistics."},{"cited_title":"Wind power forecasting based on large time series model,","cited_arxiv_id":null,"evidence_quote":"Is the energy-domain foundation-model adaptation the paper contrasts with, since wind power is a physical rather than market-clearing target."},{"cited_title":"Chronos: Learning the language of time series,","cited_arxiv_id":null,"evidence_quote":"Is the original Chronos model, serving as a zero-shot foundation-model baseline in the main comparison."}],"review_version":1}