TRACE：结合官方运营文本的电价预测可复现基准

# A Reproducible Benchmark for Electricity Price Forecasting with Official Operational Text
Source: [https://arxiv.org/html/2610.02256](https://arxiv.org/html/2610.02256)
\\workshoptitle

Foundation Models for Temporal Systems: From Forecasting to World Modeling

## TRACE: A Reproducible Benchmark for Electricity Price Forecasting with Official Operational TextThanks:Accepted to the NeurIPS 2026 Workshop on Foundation Models for Time Series \(FMTS\)\.

Xinyi YiAffiliation:Department of EngineeringAffiliation:University of CambridgeEmail:[xy343@cam\.ac\.uk](mailto:)Moy YuanAffiliation:Department of Computer ScienceAffiliation:University of CambridgeEmail:[zy317@cam\.ac\.uk](mailto:)

###### Abstract

Electricity price forecasting \(EPF\) supports scheduling, bidding, and risk management in electricity markets, yet existing benchmarks focus mainly on numerical inputs, leaving the forecasting value of forecast\-time textual context insufficiently evaluated\. We introduce TRACE, a reproducible benchmark of 7,300 zone–day instances pairing prices from five zones in a major U\.S\. market with official operational text available at the forecast cutoff\. TRACE reconstructs official operational text at each cutoff, preventing post\-cutoff information leakage\. We evaluate TRACE for semantic alignment and forecasting value\. Semantic assessments align with central movement and both tail risks in ground\-truth prices, most consistently for upper\-tail price risk\. Forecasting value is reflected in a median 7\.4% reduction in upper\-tail pinball loss across time\-series foundation models\. A controlled cross\-day text\-mismatch ablation reverses the gains, falling below the no\-text baseline\.

## 1Introduction

Electricity price forecasting \(EPF\) supports bidding, scheduling, and risk management in electricity markets\[[1](https://arxiv.org/html/2610.02256#bib.bib1),[2](https://arxiv.org/html/2610.02256#bib.bib2)\]\. Most methods rely on historical prices and other numerical inputs, possibly augmented with demand, weather, fuel, or calendar variables\[[3](https://arxiv.org/html/2610.02256#bib.bib3)\]\. While these inputs provide strong baselines, the forecasting value of forecast\-time textual context remains underexplored\.

System operators publish official operational text on scarcity, constraints, weather stress, and operational actions\. These official operational records can signal changing conditions before they appear in recent prices\. For example, on 6 January 2022, PJM, a major U\.S\. electricity market operator\[[4](https://arxiv.org/html/2610.02256#bib.bib4)\],111PJM coordinates the wholesale electricity market and bulk power system across 13 states and the District of Columbia\.issued a Cold Weather Alert before 7 January day\-ahead prices were determined; in COMED,222COMED \(Commonwealth Edison\) is a PJM market zone covering northern Illinois, including the Chicago area\.nine hourly prices subsequently exceeded the prior 14\-day maximum\. This case illustrates the potential for official operational text to provide forecasting value beyond recent price history\. Textual context has likewise shown forecasting value beyond numerical inputs in equities\[[5](https://arxiv.org/html/2610.02256#bib.bib5)\]and commodities\[[6](https://arxiv.org/html/2610.02256#bib.bib6)\]\. Electricity markets provide a natural setting for studying textual context beyond numerical inputs: prices exhibit recurring calendar patterns\[[7](https://arxiv.org/html/2610.02256#bib.bib7),[8](https://arxiv.org/html/2610.02256#bib.bib8)\]yet respond sharply to operational stress\.

Recent multimodal time\-series benchmarks pair numerical inputs with both naturally occurring and constructed textual context, but do not target EPF\. Time\-MMD, for example, spans agriculture, climate, gasoline prices, health, and traffic\[[9](https://arxiv.org/html/2610.02256#bib.bib9)\], whereas other benchmarks construct forecasting tasks and textual context so that the context is informative by design\[[10](https://arxiv.org/html/2610.02256#bib.bib10)\]\. Related work also studies irregular multimodal observations in healthcare, climate, and finance\[[11](https://arxiv.org/html/2610.02256#bib.bib11)\], and data\-leakage\-aware context construction\[[12](https://arxiv.org/html/2610.02256#bib.bib12)\]\. In EPF, concurrent work NSW\-EPNews\[[13](https://arxiv.org/html/2610.02256#bib.bib13)\]reports limited gains from textual context, but uses third\-party market commentary and GPT\-4o\-generated summaries, complicating forecast\-time reconstruction and making benchmark construction harder to reproduce\.

We introduce TRACE \(Temporal Reconstruction of Available Context for Electricity Price Forecasting\), to our knowledge the first reproducible EPF benchmark pairing prices\[[14](https://arxiv.org/html/2610.02256#bib.bib14)\]with official operational text\[[15](https://arxiv.org/html/2610.02256#bib.bib16)\]reconstructed at each forecast cutoff\. Across 1,460 days and five zones, TRACE uses publication time, cutoff\-time cancellation status, and applicability intervals announced at publication to construct 7,300 zone–day instances without post\-cutoff information leakage\. Its construction applies unchanged as new official operational records arrive\. Figure[1](https://arxiv.org/html/2610.02256#S1.F1)shows one instance\.

Figure 1:TRACE instance from COMED \(target: 7 January 2022\)\. A Cold Weather Alert was issued four hours before the forecast cutoff \(b\); 9 prices exceeded the prior 14\-day maximum \(a,c\)\.TRACE makes three contributions\. First, it provides a reproducible benchmark pairing prices from five PJM zones with official operational text reconstructed at the forecast cutoff\. Second, it evaluates semantic alignment: semantic assessments align with central movement and both tail risks, most consistently for upper\-tail price risk\. Third, it evaluates forecasting value: forecast\-time textual context reduces upper\-tail pinball loss by a median7\.4%7\.4\\%across time\-series foundation models, while a controlled cross\-day text\-mismatch ablation reverses gains, falling below the no\-text baseline\.

## 2TRACE benchmark construction

For delivery daydd, the forecast cutoffτd\\tau\_\{d\}is 11:00 ET on dayd−1d\-1, when PJM’s day\-ahead bid period closes\[[16](https://arxiv.org/html/2610.02256#bib.bib15)\]\. TRACE deterministically aligns zone\-specific price history and official operational text available atτd\\tau\_\{d\}with the target for daydd\. Figure[2](https://arxiv.org/html/2610.02256#S2.F2)summarizes the construction\.

Figure 2:TRACE construction\.Deterministic forecast\-time textual context reconstruction\.TRACE uses PJM’s Emergency Procedures archive\[[15](https://arxiv.org/html/2610.02256#bib.bib16)\]for precise forecast\-time reconstruction and broad coverage \(Appendix[A\.1](https://arxiv.org/html/2610.02256#A1.SS1)\)\. We fix a 14\-day window before evaluation, retaining87\.6%87\.6\\%of unlimited\-lookback coverage while limiting stale context \(Appendix[A\.2](https://arxiv.org/html/2610.02256#A1.SS2)\)\. Records are filtered by publication and cutoff\-time cancellation status; announced applicability intervals must overlap the target day\. Thus, only information available at the forecast cutoff enters the benchmark, preventing post\-cutoff information leakage\.

Numerical inputs\.Numerical inputs use the same 14\-day window of zone\-specific prices\[[14](https://arxiv.org/html/2610.02256#bib.bib14)\]\. Following common DST normalization\[[17](https://arxiv.org/html/2610.02256#bib.bib17)\], spring\-forward gaps are interpolated and one fall\-back observation retained, yielding 336 hourly inputs\.

Cross\-modal alignment and benchmark splits\.Cross\-modal inputs align by\(d,τd\)\(d,\\tau\_\{d\}\): prices use the preceding 14\-day prices, while text includes records published and uncancelled byτd\\tau\_\{d\}and applicable todd\. TRACE contains 7,300 five\-zone instances over 1,460 days \(2021\-10\-02–2025\-09\-30\), chronological split 821/366/273 train/validation/test; 702 records cover 400/146/131 days\.

## 3Evaluating semantic alignment and forecasting value in TRACE

### 3\.1Semantic alignment of textual context

We test whether official operational text available at the forecast cutoff contains information about subsequent prices\. For each delivery daydd, an LLM semantic assessor reads all retained official operational text available by its forecast cutoffτd\\tau\_\{d\}and outputs three assessments: central movementC^d\\hat\{C\}\_\{d\}, upper\-tail price riskS^dU\\hat\{S\}\_\{d\}^\{U\}, and lower\-tail price riskS^dL\\hat\{S\}\_\{d\}^\{L\}\. All use\{−1,0,\+1\}\\\{\-1,0,\+1\\\}, denotingdown/neutral/upforC^d\\hat\{C\}\_\{d\}anddecrease/no clear change/increasefor the tails\.

Ground\-truth central movement and tail risks\.Let𝒟\\mathcal\{D\}be training and validation days with textual context\. Letyd,hy\_\{d,h\}be the ground\-truth price at hourhhon daydd, andy¯d\\bar\{y\}\_\{d\}its daily mean\. Using the previous\-day reference, defineDd=y¯d−y¯d−1D\_\{d\}=\\bar\{y\}\_\{d\}\-\\bar\{y\}\_\{d\-1\},rd,h=yd,h−yd−1,hr\_\{d,h\}=y\_\{d,h\}\-y\_\{d\-1,h\}, andδ=0\.1​medd∈𝒟​\|Dd\|\\delta=0\.1\\,\\operatorname\{med\}\_\{d\\in\\mathcal\{D\}\}\|D\_\{d\}\|\. The ground\-truth central\-movement direction and day\-level tail\-risk levels are

Cd=\{\+1,Dd\>δ,0,\|Dd\|≤δ,−1,Dd<−δ,Ud=Q0\.9​\(\{rd,h\}h\),Ld=−Q0\.1​\(\{rd,h\}h\)\.C\_\{d\}=\\begin\{cases\}\+1,&D\_\{d\}\>\\delta,\\\\ 0,&\|D\_\{d\}\|\\leq\\delta,\\\\ \-1,&D\_\{d\}<\-\\delta,\\end\{cases\}\\qquad U\_\{d\}=Q\_\{0\.9\}\(\\\{r\_\{d,h\}\\\}\_\{h\}\),\\qquad L\_\{d\}=\-Q\_\{0\.1\}\(\\\{r\_\{d,h\}\\\}\_\{h\}\)\.To account for recurring calendar patterns\[[7](https://arxiv.org/html/2610.02256#bib.bib7),[8](https://arxiv.org/html/2610.02256#bib.bib8)\], we repeat the construction using four\-week same\-weekday medians,medj=14​y¯d−7​j\\operatorname\{med\}\_\{j=1\}^\{4\}\\bar\{y\}\_\{d\-7j\}andmedj=14​yd−7​j,h\\operatorname\{med\}\_\{j=1\}^\{4\}y\_\{d\-7j,h\}, in place ofy¯d−1\\bar\{y\}\_\{d\-1\}andyd−1,hy\_\{d\-1,h\}, respectively\.

Alignment metrics\.We compare the LLM assessments with ground\-truth central\-movement directionCdC\_\{d\}and upper\- and lower\-tail risksUdU\_\{d\}andLdL\_\{d\}\. Central alignment uses balanced accuracyACA\_\{C\}\(chance1/31/3\), while tail alignment uses Spearman correlation\[[18](https://arxiv.org/html/2610.02256#bib.bib34)\]across days,AUA\_\{U\}andALA\_\{L\}\(null00\):

AC=\\displaystyle A\_\{C\}=13​∑c∈\{−1,0,\+1\}Prd∈𝒟⁡\(C^d=c∣Cd=c\),\\displaystyle\\frac\{1\}\{3\}\\sum\_\{c\\in\\\{\-1,0,\+1\\\}\}\\Pr\_\{d\\in\\mathcal\{D\}\}\(\\hat\{C\}\_\{d\}=c\\mid C\_\{d\}=c\),AU=\\displaystyle A\_\{U\}=ρS​\(\(S^dU\)d∈𝒟,\(Ud\)d∈𝒟\),AL=ρS​\(\(S^dL\)d∈𝒟,\(Ld\)d∈𝒟\)\.\\displaystyle\\rho\_\{S\}\\\!\\left\(\(\\hat\{S\}\_\{d\}^\{U\}\)\_\{d\\in\\mathcal\{D\}\},\(U\_\{d\}\)\_\{d\\in\\mathcal\{D\}\}\\right\),A\_\{L\}=\\rho\_\{S\}\\\!\\left\(\(\\hat\{S\}\_\{d\}^\{L\}\)\_\{d\\in\\mathcal\{D\}\},\(L\_\{d\}\)\_\{d\\in\\mathcal\{D\}\}\\right\)\.Here,AC∈\[0,1\]A\_\{C\}\\in\[0,1\]andAU,AL∈\[−1,1\]A\_\{U\},A\_\{L\}\\in\[\-1,1\], with larger values indicating stronger positive alignment\.

Experimental results\.We use Claude Opus 5\[[19](https://arxiv.org/html/2610.02256#bib.bib18)\]and DeepSeek V4\-Flash\[[20](https://arxiv.org/html/2610.02256#bib.bib19)\]as LLM semantic assessors with the same frozen prompt \(Appendix[B\.1](https://arxiv.org/html/2610.02256#A2.SS1)\)\. Assessments are frozen and shared across zones; alignment is zone\-specific\. Significance uses1,0001\{,\}000joint day\-level permutations, one\-sidedpp\-values, and Holm–Bonferroni correction over three dimensions, two references within each zone–LLM pair\[[21](https://arxiv.org/html/2610.02256#bib.bib20)\]\. Positive, significant alignment \(AC\>1/3A\_\{C\}\>1/3orAU,AL\>0A\_\{U\},A\_\{L\}\>0; adjustedp<0\.05p<0\.05\) occurs in 17/20 central, 20/20 upper\-tail, and 16/20 lower\-tail settings\. Upper\-tail risk is most consistent, matching the strongest and most consistent forecasting value of textual context below \(Table[1](https://arxiv.org/html/2610.02256#S3.T1)\)\.

Table 1:Cross\-zone semantic alignment\. Full results are in Appendix[B\.2](https://arxiv.org/html/2610.02256#A2.SS2)\.A\. Positive, Holm\-significant alignment Positive:AC\>1/3A\_\{C\}\>1/3,AU,AL\>0A\_\{U\},A\_\{L\}\>0; significant: Holmp<0\.05p<0\.05

ZoneCentralACA\_\{C\}UpperAUA\_\{U\}LowerALA\_\{L\}COMED2/44/42/4AEP3/44/42/4DOM4/44/44/4PSEG4/44/44/4BGE4/44/44/4All17/2020/2016/20
B\. Upper\-tail alignmentAUA\_\{U\} All entries:AU\>0A\_\{U\}\>0, Holmp<0\.05p<0\.05

ClaudeDeepSeekZonePrev\.Same\-wd\.Prev\.Same\-wd\.COMED0\.1640\.4290\.1770\.415AEP0\.2470\.4020\.2630\.390DOM0\.2820\.3630\.3010\.371PSEG0\.3470\.4490\.3740\.432BGE0\.2930\.3980\.2840\.398

### 3\.2Forecasting value of textual context

We evaluate TRACE forecasting in two settings:direct text fusion\[[9](https://arxiv.org/html/2610.02256#bib.bib9),[22](https://arxiv.org/html/2610.02256#bib.bib21)\]combines encoded official operational text with numerical features, whilestructured semantic fusion\[[23](https://arxiv.org/html/2610.02256#bib.bib22)\]uses semantic assessments\. Forτ∈\{0\.1,0\.5,0\.9\}\\tau\\in\\\{0\.1,0\.5,0\.9\\\}, we use pinball loss\[[24](https://arxiv.org/html/2610.02256#bib.bib23),[25](https://arxiv.org/html/2610.02256#bib.bib24)\],PBτ=\|Ω\|−1∑\(d,h\)∈Ω\(τ−𝟏\{yd,h<qd,hτ\}\)\(yd,h−qd,hτ\)\\mathrm\{PB\}\_\{\\tau\}=\|\\Omega\|^\{\-1\}\\sum\_\{\(d,h\)\\in\\Omega\}\(\\tau\-\\mathbf\{1\}\\\{y\_\{d,h\}<q\_\{d,h\}^\{\\tau\}\\\}\)\(y\_\{d,h\}\-q\_\{d,h\}^\{\\tau\}\), whereΩ\\Omegaindexes test day–hour pairs,yd,hy\_\{d,h\}is the ground\-truth price, andqd,hτq\_\{d,h\}^\{\\tau\}the predictedτ\\tau\-quantile; lower is better\. Seven zero\-shot time\-series foundation models serve as baselines: TimesFM 2\.5\[[26](https://arxiv.org/html/2610.02256#bib.bib25)\], TimesFM 3\.0\[[27](https://arxiv.org/html/2610.02256#bib.bib26)\], Chronos\-Bolt\-Tiny/Base\[[28](https://arxiv.org/html/2610.02256#bib.bib27)\], Chronos\-2\[[29](https://arxiv.org/html/2610.02256#bib.bib28)\], Moirai\-2\[[30](https://arxiv.org/html/2610.02256#bib.bib29)\], and Sundial\[[31](https://arxiv.org/html/2610.02256#bib.bib30)\], using 336 hours of history to forecast 24 hours\.

Fordirect text fusion\(Figure[3](https://arxiv.org/html/2610.02256#S3.F3)A\), following MM\-TSFlib\[[9](https://arxiv.org/html/2610.02256#bib.bib9)\], trainable PatchTST\[[32](https://arxiv.org/html/2610.02256#bib.bib31)\]encodes the 14\-day zone\-specific price history and frozen BERT\[[33](https://arxiv.org/html/2610.02256#bib.bib32)\]encodes chronologically concatenated official operational text\. The resulting representations feed a quantile head trained jointly \(except BERT\) with equal\-weight pinball loss overτ∈\{0\.1,0\.5,0\.9\}\\tau\\in\\\{0\.1,0\.5,0\.9\\\}; results average five seeds\.NumOnlyis trained identically with text features zeroed; gains are reported relative toNumOnly\.

Forstructured semantic fusion\(Figure[3](https://arxiv.org/html/2610.02256#S3.F3)B\), frozen LLMs provide a 12\-dimensional vector of three one\-hot semantic assessments and confidence scores\. For each forecast hour, the MLP takes this vector, three baseline quantile forecasts from the frozen time\-series foundation model, and four time\-of\-day features as 19 inputs, and outputs quantile forecasts\. The→→319\\\!\\rightarrow\\\!64\\\!\\rightarrow\\\!3adaptor is trained with equal\-weight pinball loss while preserving quantile order\.Cnumis trained identically with the vector zeroed; gains overCnumisolate the semantic contribution beyond numerical recalibration\.

Table[2](https://arxiv.org/html/2610.02256#S3.T2)summarizes upper\-tail forecasting gains\. For each model family and zone, it reports the member with the lowest frozenPB0\.9\\mathrm\{PB\}\_\{0\.9\}; full seven\-model results are in Appendix[C](https://arxiv.org/html/2610.02256#A3)\. Direct text fusion reducesPB0\.9\\mathrm\{PB\}\_\{0\.9\}in all five zones by3\.7–9\.1%versusNumOnly\. Atτ=0\.9\\tau=0\.9, structured semantic fusion improves all70/70zone–model–LLM settings overCnum, with a median7\.4%reduction \(range1\.41\.4–19\.3%19\.3\\%\), and by11\.3%versus the frozen baselines \(range2\.82\.8–24\.3%24\.3\\%\); atτ=0\.5\\tau=0\.5,64/70settings improve\. Correct pairing beats the mean over 1,000 context re\-pairings in all5/5direct\-text zones and70/70semantic settings, with median gains of20\.7%and13\.9%, respectively\.

Figure 3:Direct text fusion \(A\) and structured semantic fusion \(B\)\.Table 2:Forecasting results of direct text fusion and structured semantic fusion\.A\. Direct text fusion atτ=0\.9\\tau=0\.9

ZoneDirect↓%\\downarrow\\%vs\.NumOnly↓%\\downarrow\\%vs\. ShuffleCOMED4\.89736\.4%19\.6%AEP4\.43807\.2%20\.6%DOM8\.48727\.9%20\.7%PSEG5\.20733\.7%23\.8%BGE6\.92719\.1%27\.4%
B\. Semantic fusion

τ\\tauImproved settings0\.970/70 \(100%\)0\.564/70 \(91%\)0\.124/70 \(34%\)
C\. Structured semantic fusion atτ=0\.9\\tau=0\.9

Claude Opus 5DeepSeek V4\-FlashZoneBaselineFrozenCnumFusion\(↓%\\downarrow\\%vs\.Cnum\)ShuffleFusion\(↓%\\downarrow\\%vs\.Cnum\)ShuffleCOMEDTimesFM 3\.03\.7223\.6053\.403 \(5\.6%\)3\.8613\.416 \(5\.2%\)3\.872Chronos\-Bolt\-Base3\.6673\.5303\.396 \(3\.8%\)3\.7133\.410 \(3\.4%\)3\.761Moirai\-23\.8633\.7653\.465 \(8\.0%\)4\.0923\.506 \(6\.9%\)4\.163Sundial4\.7663\.8623\.606 \(6\.6%\)4\.0693\.626 \(6\.1%\)3\.975AEPTimesFM 3\.03\.8713\.7703\.432 \(9\.0%\)3\.8033\.508 \(6\.9%\)3\.805Chronos\-Bolt\-Base3\.7913\.6283\.330 \(8\.2%\)3\.7783\.376 \(7\.0%\)3\.825Moirai\-23\.9413\.8223\.533 \(7\.5%\)3\.9473\.535 \(7\.5%\)3\.966Sundial4\.6434\.0573\.823 \(5\.8%\)4\.1393\.908 \(3\.7%\)4\.121DOMTimesFM 3\.06\.9566\.9146\.573 \(4\.9%\)7\.6196\.758 \(2\.3%\)7\.938Chronos\-27\.2257\.0266\.521 \(7\.2%\)7\.2436\.604 \(6\.0%\)7\.403Moirai\-27\.3166\.9236\.490 \(6\.3%\)7\.5496\.617 \(4\.4%\)7\.910Sundial8\.7197\.5447\.180 \(4\.8%\)8\.1307\.334 \(2\.8%\)8\.395PSEGTimesFM 3\.04\.1584\.0993\.446 \(15\.9%\)4\.4593\.517 \(14\.2%\)4\.244Chronos\-23\.9863\.6893\.300 \(10\.5%\)4\.2653\.427 \(7\.1%\)4\.368Moirai\-24\.1734\.0403\.432 \(15\.0%\)4\.3573\.492 \(13\.6%\)4\.251Sundial5\.1514\.4563\.928 \(11\.8%\)4\.8513\.972 \(10\.9%\)4\.806BGETimesFM 3\.06\.1316\.0825\.524 \(9\.2%\)6\.7065\.670 \(6\.8%\)6\.931Chronos\-25\.9865\.9025\.488 \(7\.0%\)6\.3655\.534 \(6\.2%\)6\.462Moirai\-26\.4546\.2555\.590 \(10\.6%\)7\.0255\.744 \(8\.2%\)7\.366Sundial7\.6346\.6445\.956 \(10\.4%\)7\.1366\.009 \(9\.6%\)7\.278

## 4Discussion and outlook

We propose TRACE, the first reproducible EPF benchmark with official operational text\. Semantic assessments show strongest alignment with upper\-tail risk, where forecasting gains are strongest\.

## References

- \[1\]P\. Joskow and J\. Tirole\(2007\)Reliability and competitive electricity markets\.The RAND Journal of Economics38\(1\),pp\. 60–84\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p1.1)\.
- \[2\]W\. W\. Hogan\(2013\)Electricity scarcity pricing through operating reserves\.Economics of Energy & Environmental Policy2\(2\),pp\. 65–86\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p1.1)\.
- \[3\]Y\. Wang, H\. Wu, J\. Dong, G\. Qin, H\. Zhang, Y\. Liu, Y\. Qiu, J\. Wang, and M\. Long\(2024\)Timexer: empowering transformers for time series forecasting with exogenous variables\.Advances in neural information processing systems37,pp\. 469–498\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p1.1)\.
- \[4\]PJM Interconnection, L\.L\.C\.\(2026\)PJM Website\.Note:PJM InterconnectionAccessed August 26, 2026External Links:[Link](https://www.pjm.com/)Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p2.1)\.
- \[5\]P\. C\. Tetlock\(2007\)Giving content to investor sentiment: the role of media in the stock market\.The Journal of finance62\(3\),pp\. 1139–1168\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p2.1)\.
- \[6\]L\. A\. Smales\(2014\)News sentiment in the gold futures market\.Journal of Banking & Finance49,pp\. 275–286\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p2.1)\.
- \[7\]R\. Weron\(2014\)Electricity price forecasting: a review of the state\-of\-the\-art with a look into the future\.International journal of forecasting30\(4\),pp\. 1030–1081\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p2.1),[§3\.1](https://arxiv.org/html/2610.02256#S3.SS1.p2.2)\.
- \[8\]J\. Lago, G\. Marcjasz, B\. De Schutter, and R\. Weron\(2021\)Forecasting day\-ahead electricity prices: a review of state\-of\-the\-art algorithms, best practices and an open\-access benchmark\.Applied Energy293,pp\. 116983\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p2.1),[§3\.1](https://arxiv.org/html/2610.02256#S3.SS1.p2.2)\.
- \[9\]H\. Liu, S\. Xu, Z\. Zhao, L\. Kong, H\. Kamarthi, A\. B\. Sasanur, M\. Sharma, J\. Cui, Q\. Wen, C\. Zhang,et al\.\(2024\)Time\-mmd: multi\-domain multimodal dataset for time series analysis\.Advances in Neural Information Processing Systems37,pp\. 77888–77933\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p3.1),[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1),[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p2.1)\.
- \[10\]A\. R\. Williams, A\. Ashok, É\. Marcotte, V\. Zantedeschi, J\. Subramanian, R\. Riachi, J\. Requeima, A\. Lacoste, I\. Rish, N\. Chapados,et al\.\(2024\)Context is key: a benchmark for forecasting with essential textual information\.arXiv preprint arXiv:2410\.18959\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p3.1)\.
- \[11\]C\. Chang, J\. Hwang, Y\. Shi, H\. Wang, W\. Wang, W\. Peng, and T\. Chen\(2025\)Time\-imm: a dataset and benchmark for irregular multimodal multivariate time series\.Advances in Neural Information Processing Systems38\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p3.1)\.
- \[12\]Z\. Xu, W\. Cai, X\. Dai, Z\. Deng, and Q\. Xu\(2025\)Fidel\-ts: a high\-fidelity multimodal benchmark for time series forecasting\.arXiv preprint arXiv:2509\.24789\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p3.1)\.
- \[13\]Z\. Bi, L\. Huang, H\. Jin, Q\. Zeng, and H\. Chen\(2025\)NSW\-epnews: a news\-augmented benchmark for electricity price forecasting with llms\.arXiv preprint arXiv:2506\.11050\.Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p3.1)\.
- \[14\]PJM Interconnection, L\.L\.C\.\(2026\)Data Miner\.Note:PJM InterconnectionAccessed August 26, 2026External Links:[Link](https://www.pjm.com/markets-and-operations/etools/data-miner-2.aspx)Cited by:[§1](https://arxiv.org/html/2610.02256#S1.p4.1),[§2](https://arxiv.org/html/2610.02256#S2.p3.1)\.
- \[15\]PJM Interconnection, L\.L\.C\.\(2026\)Emergency Procedures\.Note:PJM InterconnectionAccessed August 26, 2026External Links:[Link](https://www.pjm.com/markets-and-operations/etools/emerg-procedure.aspx)Cited by:[§A\.1](https://arxiv.org/html/2610.02256#A1.SS1.p2.1),[§1](https://arxiv.org/html/2610.02256#S1.p4.1),[§2](https://arxiv.org/html/2610.02256#S2.p2.1)\.
- \[16\]PJM Interconnection\(2026\)PJM manual 11: energy & ancillary services market operations\.PJM Interconnection\.Note:Revision 137, effective July 28, 2026External Links:[Link](https://www.pjm.com/-/media/DotCom/documents/manuals/m11.pdf)Cited by:[§2](https://arxiv.org/html/2610.02256#S2.p1.1)\.
- \[17\]N\. Hauzenberger, M\. Pfarrhofer, and L\. Rossini\(2025\)Sparse time\-varying parameter vecms with an application to modeling electricity prices\.International Journal of Forecasting41\(1\),pp\. 361–376\.Cited by:[§2](https://arxiv.org/html/2610.02256#S2.p3.1)\.
- \[18\]C\. Spearman\(1904\)The proof and measurement of association between two things\.The American Journal of Psychology15\(1\),pp\. 72–101\.External Links:[Document](https://dx.doi.org/10.2307/1412159)Cited by:[§3\.1](https://arxiv.org/html/2610.02256#S3.SS1.p3.1)\.
- \[19\]Anthropic\(2026\)Claude Opus 5 System Card\.Note:Released July 24, 2026External Links:[Link](https://www-cdn.anthropic.com/b514064af1408018e64b1ad24e7d5e75850b4ffd/Claude%20Opus%205%20System%20Card.pdf)Cited by:[§B\.1](https://arxiv.org/html/2610.02256#A2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2610.02256#S3.SS1.p4.1)\.
- \[20\]A\. Xu, B\. Lin, B\. Xue, B\. Wang, B\. Xu, B\. Wu, B\. Zhang, C\. Lin, C\. Dong, C\. Ling,et al\.\(2026\)Deepseek\-v4: towards highly efficient million\-token context intelligence\.arXiv preprint arXiv:2606\.19348\.Cited by:[§B\.1](https://arxiv.org/html/2610.02256#A2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2610.02256#S3.SS1.p4.1)\.
- \[21\]S\. Holm\(1979\)A simple sequentially rejective multiple test procedure\.Scandinavian journal of statistics,pp\. 65–70\.Cited by:[§3\.1](https://arxiv.org/html/2610.02256#S3.SS1.p4.1)\.
- \[22\]R\. Huang, Z\. Zhang, and Y\. Wang\(2025\)Cross\-moe: an efficient temporal prediction framework integrating textual modality\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 29927–29938\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[23\]L\. Li, Y\. Wang, J\. Yan, W\. Zhang, J\. Deng, H\. Sun, Z\. Han, and Y\. Gong\(2026\)From text to forecasts: bridging modality gap with temporal evolution semantic space\.arXiv preprint arXiv:2603\.12664\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[24\]B\. Lim, S\. Ö\. Arık, N\. Loeff, and T\. Pfister\(2021\)Temporal fusion transformers for interpretable multi\-horizon time series forecasting\.International journal of forecasting37\(4\),pp\. 1748–1764\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[25\]C\. Cornell, N\. T\. Dinh, and S\. A\. Pourmousavi\(2024\)A probabilistic forecast methodology for volatile electricity prices in the australian national electricity market\.International Journal of Forecasting40\(4\),pp\. 1421–1437\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[26\]A\. Das, W\. Kong, R\. Sen, and Y\. Zhou\(2023\)A decoder\-only foundation model for time\-series forecasting\.arXiv preprint arXiv:2310\.10688\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[27\]Google Research\(2026\)TimesFM\.Note:GitHub repositoryTimesFM 3\.0 release, accessed September 7, 2026External Links:[Link](https://github.com/google-research/timesfm)Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[28\]A\. F\. Ansari, L\. Stella, C\. Turkmen, X\. Zhang, P\. Mercado, H\. Shen, O\. Shchur, S\. S\. Rangapuram, S\. P\. Arango, S\. Kapoor,et al\.\(2024\)Chronos: learning the language of time series\.arXiv preprint arXiv:2403\.07815\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[29\]A\. F\. Ansari, O\. Shchur, J\. Küken, A\. Auer, B\. Han, P\. Mercado, S\. S\. Rangapuram, H\. Shen, L\. Stella, X\. Zhang,et al\.\(2025\)Chronos\-2: from univariate to universal forecasting\.arXiv preprint arXiv:2510\.15821\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[30\]C\. Liu, T\. Aksu, J\. Liu, X\. Liu, H\. Yan, Q\. Pham, S\. Savarese, D\. Sahoo, C\. Xiong, and J\. Li\(2025\)Moirai 2\.0: when less is more for time series forecasting\.arXiv preprint arXiv:2511\.11698\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[31\]Y\. Liu, G\. Qin, Z\. Shi, Z\. Chen, C\. Yang, X\. Huang, J\. Wang, and M\. Long\(2025\)Sundial: a family of highly capable time series foundation models\.arXiv preprint arXiv:2502\.00816\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p1.1)\.
- \[32\]Y\. Nie, N\. H\. Nguyen, P\. Sinthong, and J\. Kalagnanam\(2022\)A time series is worth 64 words: long\-term forecasting with transformers\.arXiv preprint arXiv:2211\.14730\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p2.1)\.
- \[33\]J\. Devlin, M\. Chang, K\. Lee, and K\. Toutanova\(2019\)Bert: pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 \(long and short papers\),pp\. 4171–4186\.Cited by:[§3\.2](https://arxiv.org/html/2610.02256#S3.SS2.p2.1)\.
- \[34\]PJM Interconnection, L\.L\.C\.\(2026\)PJM Inside Lines\.Note:PJM InterconnectionAccessed August 26, 2026External Links:[Link](https://insidelines.pjm.com/)Cited by:[§A\.1](https://arxiv.org/html/2610.02256#A1.SS1.p2.1)\.

## Appendix ABenchmark construction details

### A\.1Official operational text source selection

We select the text source before any semantic or forecasting evaluation\. For delivery daydd, letτd\\tau\_\{d\}denote the forecast cutoff and letpip\_\{i\}be the publication time of recordii\. Under unlimited lookback, letpd⋆=maxi:pi<τdpip\_\{d\}^\{\\star\}=\\max\_\{i:p\_\{i\}<\\tau\_\{d\}\}p\_\{i\}denote the latest publication available before the forecast cutoff for delivery daydd, and definead=τd−pd⋆a\_\{d\}=\\tau\_\{d\}\-p\_\{d\}^\{\\star\}as its age\. Thus, “Withinkkdays” is the fraction of delivery days whose latest pre\-cutoff publication is no more thankkdays old\.

Both Emergency Procedures\[[15](https://arxiv.org/html/2610.02256#bib.bib16)\]and Inside Lines\[[34](https://arxiv.org/html/2610.02256#bib.bib33)\]have at least one pre\-cutoff publication for all 1,460 delivery days\. Emergency Procedures is more recent: its medianada\_\{d\}is 1\.15 days, compared with 3\.00 days for Inside Lines \(Table[3](https://arxiv.org/html/2610.02256#A1.T3)\)\. It also provides the publication and cancellation timestamps required for forecast\-time reconstruction\. Inside Lines provides update timestamps but does not preserve superseded text, so its state at an earlier forecast cutoff cannot be reconstructed\. We therefore use the PJM Emergency Procedures archive as the source of official operational text\.

Table 3:Recency of candidate PJM text sources under unlimited lookback\.SourceMedian ageWithin 1 dayWithin 3 daysWithin 7 daysEmergency Procedures1\.15 days44\.5%77\.4%94\.9%Inside Lines3\.00 days21\.3%50\.1%79\.9%
### A\.2Context window and coverage

After selecting Emergency Procedures, we fix the textual context window before any semantic or forecasting evaluation\. Letpminp\_\{\\min\}be the earliest publication in the collected archive \(2021\-09\-04\)\. For delivery dayddwith forecast cutoffτd\\tau\_\{d\}, aWW\-day window considers records published in

\[max⁡\{pmin,τd−W\},τd\)\.\[\\max\\\{p\_\{\\min\},\\,\\tau\_\{d\}\-W\\\},\\,\\tau\_\{d\}\)\.The full\-archive setting removes the finite lookback limit and therefore uses\[pmin,τd\)\[p\_\{\\min\},\\tau\_\{d\}\)\. Thus, benchmark days near the start of the evaluation period may use records published before the first delivery day\. Within each window, records are filtered by publication and cutoff\-time cancellation status; if an applicability interval was announced at publication, it must overlap delivery daydd\. Let𝒵d​\(W\)\\mathcal\{Z\}\_\{d\}\(W\)denote the retained records\. A day is context\-bearing when\|𝒵d​\(W\)\|\>0\|\\mathcal\{Z\}\_\{d\}\(W\)\|\>0\.

Table[4](https://arxiv.org/html/2610.02256#A1.T4)shows the resulting coverage\. The full archive yields 773 context\-bearing days\. A 14\-day window retains 677 of them \(87\.6%87\.6\\%\); extending to 28 days adds only 42 days while doubling the maximum record age\. We therefore setW=14W=14days\.

Table 4:Sensitivity of forecast\-time context coverage to the lookback window\. Coverage is relative to the 773 context\-bearing days obtained from the full archive\.Context windowContext\-bearing daysCoverage1 day32642\.2%7 days62180\.3%14 days67787\.6%28 days71993\.0%Full archive773100\.0%

## Appendix BSemantic evaluation details

### B\.1Frozen semantic assessment prompt

Claude Opus 5\[[19](https://arxiv.org/html/2610.02256#bib.bib18)\]and DeepSeek V4\-Flash\[[20](https://arxiv.org/html/2610.02256#bib.bib19)\]independently assess all 546 context\-bearing training and validation days using the same frozen prompt\. For each day, an LLM receives only the official operational text retained by forecast\-time reconstruction within the 14\-day context window, including record type, publication time, announced applicability start time, and body text\. It returns three categorical semantic assessments and corresponding confidence scores: central movement, upper\-tail price risk, and lower\-tail price risk\. The LLM receives no price history, numerical forecasts, ground\-truth prices, or previous\-day or same\-weekday reference values\. All outputs are frozen before alignment with ground\-truth prices is computed\.

#### Frozen prompt\.

The prompt below preserves the semantic task and label definitions while omitting model\-specific API formatting and output boilerplate\.

> You are evaluating official electricity\-market information for a day\-ahead forecast\. Make three independent assessments of what the supplied context implies for the target delivery day\. You have no access to any ground\-truth price, historical price, numerical forecast, model output, or baseline\. Assess only the economic or operational pressure conveyed by the supplied context\. Do not use generic seasonal, weekday, or calendar priors\. Central price direction\.Does the context imply net upward, net downward, or no clear directional pressure on the price level for the target delivery day? UP: net upward pressure\. DOWN: net downward pressure\. NEUTRAL\_MIXED: no sufficiently clear net direction, materially offsetting effects, or evidence too weak or indirect to support a direction\. Upper\-tail risk change\.Does the context make unusually high price outcomes on the target delivery day more likely, less likely, or neither? INCREASE,DECREASE, orNO\_CLEAR\_CHANGE\. Lower\-tail risk change\.Does the context make unusually low price outcomes on the target delivery day more likely, less likely, or neither? INCREASE,DECREASE, orNO\_CLEAR\_CHANGE\. Treat the three assessments independently: do not derive either tail assessment from the central direction, or vice versa\. Consider the magnitude, timing, persistence, freshness, and operational relevance of the supplied events\. When the evidence is weak or conflicting, use the no\-clear\-signal labels rather than forcing a prediction\. Return one label and confidence score for each assessment\.

### B\.2Full five\-zone semantic alignment

Table[5](https://arxiv.org/html/2610.02256#A2.T5)reports semantic alignment for every zone–LLM pair\. For each pair, we jointly permute\(C^d,S^dU,S^dL\)\(\\hat\{C\}\_\{d\},\\hat\{S\}\_\{d\}^\{U\},\\hat\{S\}\_\{d\}^\{L\}\)across days1,0001\{,\}000times while holding ground\-truth prices fixed, and recomputeACprev,AUprev,ALprev,ACcal,AUcal,ALcalA\_\{C\}^\{\\mathrm\{prev\}\},A\_\{U\}^\{\\mathrm\{prev\}\},A\_\{L\}^\{\\mathrm\{prev\}\},A\_\{C\}^\{\\mathrm\{cal\}\},A\_\{U\}^\{\\mathrm\{cal\}\},A\_\{L\}^\{\\mathrm\{cal\}\}\. The resulting permutationpp\-values are Holm–Bonferroni corrected across these six statistics within each zone–LLM pair; the minimum attainable correctedpp\-value is0\.0060\.006\. Alignment uses 546 context\-bearing training and validation days for the previous\-day reference and 542 for the same\-weekday reference, as four days lack sufficient prior same\-weekday observations\. Because textual context is shared across zones, semantic assessments are fixed across zones, while ground\-truth quantities and alignment statistics are zone\-specific\.

Upper\-tail alignment is positive and significant for every zone–LLM pair under both references\. Central movement is significant under both references for seven of ten pairs, and lower\-tail price risk for six of ten\. Under the previous\-day reference, central movement and lower\-tail price risk are nonsignificant for COMED, and lower\-tail price risk is also nonsignificant for AEP\. Upper\-tail price risk is the only semantic dimension significant under both references for every zone–LLM pair\.

Table 5:Five\-zone semantic alignment on context\-bearing training and validation days\. Entries are observed statistics with Holm\-corrected permutationpp\-values in parentheses; bold denotes positive alignment withpHolm<0\.05p\_\{\\mathrm\{Holm\}\}<0\.05\.ACA\_\{C\}is balanced accuracy \(chance1/31/3\), whileAUA\_\{U\}andALA\_\{L\}are Spearman correlations \(null00\)\.CentralACA\_\{C\}Upper tailAUA\_\{U\}Lower tailALA\_\{L\}ZoneLLMprev\. daysame wkdayprev\. daysame wkdayprev\. daysame wkdayCOMEDClaude0\.369 \(0\.140\)0\.402\(0\.012\)0\.164\(0\.006\)0\.429\(0\.006\)−0\.008\-0\.008\(0\.661\)0\.339\(0\.006\)DeepSeek0\.361 \(0\.250\)0\.394\(0\.012\)0\.177\(0\.006\)0\.415\(0\.006\)0\.014 \(0\.457\)0\.341\(0\.006\)AEPClaude0\.413\(0\.006\)0\.384 \(0\.060\)0\.247\(0\.006\)0\.402\(0\.006\)0\.033 \(0\.272\)0\.336\(0\.006\)DeepSeek0\.413\(0\.006\)0\.399\(0\.006\)0\.263\(0\.006\)0\.390\(0\.006\)0\.069 \(0\.069\)0\.310\(0\.006\)DOMClaude0\.411\(0\.006\)0\.419\(0\.006\)0\.282\(0\.006\)0\.363\(0\.006\)0\.107\(0\.012\)0\.330\(0\.006\)DeepSeek0\.412\(0\.006\)0\.429\(0\.006\)0\.301\(0\.006\)0\.371\(0\.006\)0\.139\(0\.006\)0\.317\(0\.006\)PSEGClaude0\.419\(0\.006\)0\.426\(0\.006\)0\.347\(0\.006\)0\.449\(0\.006\)0\.139\(0\.006\)0\.316\(0\.006\)DeepSeek0\.420\(0\.006\)0\.422\(0\.006\)0\.374\(0\.006\)0\.432\(0\.006\)0\.174\(0\.006\)0\.314\(0\.006\)BGEClaude0\.405\(0\.006\)0\.390\(0\.026\)0\.293\(0\.006\)0\.398\(0\.006\)0\.089\(0\.026\)0\.312\(0\.006\)DeepSeek0\.411\(0\.006\)0\.389\(0\.016\)0\.284\(0\.006\)0\.398\(0\.006\)0\.121\(0\.006\)0\.281\(0\.006\)

## Appendix CFull five\-zone forecasting results

### C\.1Direct text fusion

Table[6](https://arxiv.org/html/2610.02256#A3.T6)reports all three quantile levels againstNumOnly, a separately trained control using the same architecture and training procedure but with text features zeroed during training\.PB0\.9\\mathrm\{PB\}\_\{0\.9\}decreases in all five zones, whereasPB0\.1\\mathrm\{PB\}\_\{0\.1\}increases in all five andPB0\.5\\mathrm\{PB\}\_\{0\.5\}in four of five\. This mirrors the semantic\-alignment results: the forecasting value of textual context is concentrated in upper\-tail price risk rather than broad distributional improvement\.

Shuffle mean is the meanPB0\.9\\mathrm\{PB\}\_\{0\.9\}over1,0001\{,\}000test\-time permutations of textual context across the 131 held\-out context\-bearing test days\. Trained models are reused; only the day–context pairing is permuted, while numerical inputs and targets remain fixed\. Correct forecast\-time day–context pairing outperforms the Shuffle mean in all five zones, with a median reduction of 20\.7%\. Moreover, the Shuffle mean is worse thanNumOnlyin all five zones, showing that mispaired textual context can degrade forecasts rather than merely fail to improve them\. Correct forecast\-time day–context pairing is significantly better than random re\-pairing under the aggregate permutation test, outperforming all1,0001\{,\}000permuted pairings \(p=1/1001p=1/1001\)\. Together with the semantic\-alignment results, this supports the forecasting value of correctly paired official operational text beyond price history alone\.

Table 6:Direct text fusion against theNumOnlycontrol and shuffled\-context control, five zones, 131 context\-bearing test days each \(pinball loss, USD/MWh, at the released four\-decimal precision\)\.↓\\downarrowcolumns are relative reductions \(%\) of Direct fromNumOnlyand from Shuffle mean atτ=0\.9\\tau=0\.9\.τ=0\.1\\tau=0\.1τ=0\.5\\tau=0\.5τ=0\.9\\tau=0\.9ZoneNumOnlyDirectNumOnlyDirectNumOnlyDirectShuffle mean↓\\downarrowNumOnly↓\\downarrowShuffleCOMED3\.19883\.43576\.82417\.15175\.23484\.89736\.09416\.419\.6AEP2\.94753\.29396\.24346\.63214\.78354\.43805\.58767\.220\.6DOM4\.13924\.516511\.023810\.75759\.21398\.487210\.70067\.920\.7PSEG3\.30893\.54897\.09887\.44535\.40755\.20736\.83183\.723\.8BGE3\.58404\.14849\.46869\.55867\.61966\.92719\.54239\.127\.4
### C\.2Structured semantic fusion

Table[7](https://arxiv.org/html/2610.02256#A3.T7)reports all 70 zone–model–LLM settings\. Atτ=0\.9\\tau=0\.9, structured semantic fusion improves every setting relative to the separately trainedCnumcontrol, with a median reduction of7\.4%7\.4\\%\(range1\.41\.4–19\.3%19\.3\\%\)\. The effect is weaker at the other quantiles: 64/70 settings improve atτ=0\.5\\tau=0\.5, whereas only 24/70 improve atτ=0\.1\\tau=0\.1\. Forecasting value is therefore strongest and most consistent for upper\-tail price risk\.

Table[7](https://arxiv.org/html/2610.02256#A3.T7)also reports the shuffled\-context control\. For each trained model, textual context is randomly re\-paired across the 131 held\-out context\-bearing test days1,0001\{,\}000times, while numerical inputs and targets remain fixed\. Correct forecast\-time day–context pairing outperforms the Shuffle mean in all 70/70 settings, with a median reduction of13\.9%13\.9\\%\. The aggregate statistic outperforms all1,0001\{,\}000random re\-pairings \(p=1/1001p=1/1001\), supporting that the gains depend on correctly matched forecast\-time textual context rather than text\-derived features alone\.

Table 7:Structured semantic fusion for all 70 zone–model–LLM settings on 131 context\-bearing test days \(pinball loss, USD/MWh; lower is better\)\.Frozenis the frozen numerical baseline andCnumthe separately trained control with the semantic vector zeroed during training\. Fusion uses the correct forecast\-time day–context pairing; Shuffle is the mean over1,0001\{,\}000test\-time context re\-pairings\. The final two columns report Claude Fusion minusCnumat the other quantiles; negative favors structured semantic information\.NumericalClaude Opus 5DeepSeek V4\-FlashClaude, otherτ\\tauZoneBaselineFrozenCnumFusionShuffleFusionShuffleΔ0\.1\\Delta\_\{0\.1\}Δ0\.5\\Delta\_\{0\.5\}COMEDTimesFM 2\.54\.12334\.02773\.63584\.20013\.67324\.2970\+\+0\.1222−\-0\.0354TimesFM 3\.03\.72223\.60473\.40273\.86093\.41603\.8716\+\+0\.1203−\-0\.0326Chronos\-Bolt\-Tiny3\.88873\.78813\.52553\.90293\.52293\.9579−\-0\.0052−\-0\.0433Chronos\-Bolt\-Base3\.66713\.53013\.39623\.71253\.41003\.7610\+\+0\.0560−\-0\.0312Chronos\-23\.74383\.61543\.46403\.84743\.47653\.8603−\-0\.1128\+\+0\.0238Moirai\-23\.86333\.76543\.46464\.09213\.50584\.1632\+\+0\.0494−\-0\.0374Sundial4\.76613\.86173\.60594\.06913\.62573\.9750\+\+0\.1192−\-0\.0259AEPTimesFM 2\.54\.24114\.05163\.62874\.14313\.77864\.1358−\-0\.0564−\-0\.0148TimesFM 3\.03\.87063\.77023\.43183\.80263\.50843\.8051−\-0\.0048−\-0\.0060Chronos\-Bolt\-Tiny4\.07153\.94533\.50924\.01593\.50964\.0646−\-0\.0153−\-0\.0545Chronos\-Bolt\-Base3\.79083\.62803\.33003\.77763\.37583\.8253\+\+0\.0052−\-0\.0343Chronos\-23\.79893\.65243\.28083\.73833\.32583\.7722−\-0\.0439−\-0\.0126Moirai\-23\.94133\.82163\.53313\.94683\.53493\.9660−\-0\.0082\+\+0\.0012Sundial4\.64254\.05673\.82314\.13853\.90804\.1207\+\+0\.0342−\-0\.0024DOMTimesFM 2\.58\.11947\.53177\.21758\.56937\.42598\.9055\+\+0\.0800−\-0\.1190TimesFM 3\.06\.95586\.91406\.57287\.61906\.75847\.9382−\-0\.0864−\-0\.1371Chronos\-Bolt\-Tiny7\.84177\.45406\.99458\.08887\.19688\.4098\+\+0\.0556−\-0\.1564Chronos\-Bolt\-Base7\.44837\.12916\.71427\.40506\.76717\.6506−\-0\.0380−\-0\.0267Chronos\-27\.22477\.02636\.52107\.24316\.60397\.4030−\-0\.0275−\-0\.1157Moirai\-27\.31626\.92266\.48977\.54896\.61727\.9101−\-0\.0104−\-0\.1236Sundial8\.71927\.54457\.18058\.13007\.33458\.3948−\-0\.0180−\-0\.0692PSEGTimesFM 2\.54\.74624\.50283\.63164\.84763\.84474\.8606\+\+0\.0600−\-0\.0637TimesFM 3\.04\.15774\.09903\.44584\.45933\.51724\.2444\+\+0\.0954−\-0\.0416Chronos\-Bolt\-Tiny4\.54234\.32443\.71564\.84843\.79084\.7851\+\+0\.0365−\-0\.0459Chronos\-Bolt\-Base4\.21543\.96503\.62894\.26073\.67534\.2254−\-0\.0340−\-0\.0001Chronos\-23\.98603\.68923\.30034\.26543\.42714\.3684\+\+0\.2009−\-0\.0529Moirai\-24\.17344\.04013\.43224\.35743\.49154\.2509\+\+0\.0242−\-0\.0126Sundial5\.15084\.45573\.92844\.85083\.97154\.8057\+\+0\.0845−\-0\.0346BGETimesFM 2\.57\.34847\.05466\.05227\.39686\.06987\.6222\+\+0\.0285−\-0\.1063TimesFM 3\.06\.13096\.08235\.52446\.70565\.66996\.9312−\-0\.0462−\-0\.0553Chronos\-Bolt\-Tiny6\.96776\.67355\.83727\.21935\.93107\.4830\+\+0\.0722−\-0\.1170Chronos\-Bolt\-Base6\.37116\.45235\.88826\.89945\.89696\.9826−\-0\.0193−\-0\.0548Chronos\-25\.98625\.90185\.48776\.36475\.53386\.4618\+\+0\.0325−\-0\.0782Moirai\-26\.45366\.25495\.59057\.02535\.74447\.3657\+\+0\.0411−\-0\.1066Sundial7\.63416\.64455\.95557\.13576\.00887\.2778\+\+0\.0038−\-0\.0667