arxiv: 2603.12451 · v3 · submitted 2026-03-12 · 💻 cs.LG

Recognition: no theorem link

Overcoming the Modality Gap in Context-Aided Forecasting

Vincent Zhihao Zheng , \'Etienne Marcotte , Arjun Ashok , Andrew Robert Williams , Lijun Sun , Alexandre Drouin , Valentina Zantedeschi

Authors on Pith no claims yet

Pith reviewed 2026-05-15 12:26 UTC · model grok-4.3

classification 💻 cs.LG

keywords context-aided forecastingtime seriesmultimodal modelsdata augmentationsemi-synthetic datadataset quality

0 comments

The pith

Semi-synthetic contexts close the performance gap in context-aided time series forecasting.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Context-aided forecasting tries to boost predictions by adding descriptive information beyond raw numbers, but multimodal models have not beaten simpler numerical models in practice. The authors trace this to low-quality contexts in current datasets that fail to provide useful complementary information. They develop a semi-synthetic augmentation technique to create large volumes of contexts that accurately reflect time series dynamics while adding verifiable extra value. This produces the CAF-7M dataset of seven million examples, on which models can be pre-trained to actually use the context when evaluated on real data. The work concludes that improving dataset quality, rather than changing model architectures, is what will advance the field.

Core claim

By introducing a semi-synthetic data augmentation method that generates contexts both descriptive of temporal dynamics and verifiably complementary to numerical histories, the authors create CAF-7M, a corpus of 7 million context-augmented time series windows with a verified test set. Pre-training on this data transfers effectively to real-world evaluation, with models demonstrating clear use of the context, indicating that dataset quality rather than architectural limitations has been the primary bottleneck.

What carries the argument

semi-synthetic data augmentation method that generates contexts descriptive of temporal dynamics and verifiably complementary to numerical histories

If this is right

Pre-training on the semi-synthetic CAF-7M dataset transfers to improved performance on real-world forecasting tasks.
Trained models exhibit measurable utilization of the provided context information.
Future progress in context-aided forecasting depends primarily on creating higher-quality datasets rather than new model designs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the method scales, similar semi-synthetic approaches could address data quality issues in other multimodal domains such as vision-language tasks.
Real-world datasets might benefit from targeted synthetic augmentation to fill gaps in context quality.
The verified test set allows direct measurement of context utilization, which could become a standard benchmark for future work.

Load-bearing premise

The semi-synthetic contexts are descriptive of temporal dynamics and verifiably complementary to numerical histories, so that training on them improves real-world performance.

What would settle it

If multimodal models pre-trained on CAF-7M show no improvement over unimodal baselines on real-world evaluation sets, or if context utilization cannot be verified through controlled tests.

read the original abstract

Context-aided forecasting (CAF) holds promise for integrating domain knowledge and forward-looking information, enabling AI systems to surpass traditional statistical methods. However, recent empirical studies reveal a puzzling gap: multimodal models often fail to outperform their unimodal counterparts. We hypothesize that this underperformance stems from poor context quality in existing datasets, as verification is challenging. To address these limitations, we introduce a semi-synthetic data augmentation method that generates contexts both descriptive of temporal dynamics and verifiably complementary to numerical histories. This approach enables massive-scale dataset creation, resulting in CAF-7M, a corpus of 7 million context-augmented time series windows, including a rigorously verified test set. We demonstrate that semi-synthetic pre-training transfers effectively to real-world evaluation, and show clear evidence of context utilization. Our results suggest that dataset quality, rather than architectural limitations, has been the primary bottleneck in context-aided forecasting.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper's main move is a semi-synthetic pipeline that builds a 7M-scale dataset with verified complementary contexts, reframing the modality gap as a data issue rather than an architecture one.

read the letter

The paper claims the modality gap in context-aided forecasting is mostly a data-quality problem and backs it with a semi-synthetic augmentation method that produces CAF-7M, a 7-million-example corpus plus a rigorously checked test set. They report that pre-training on this data transfers to real-world evaluation and that models show clear context use. That scale and the verification step are the concrete advances here; generating contexts that are both descriptive of the series dynamics and independently informative at this volume is not trivial and could be useful for anyone trying to add external signals to forecasting models. The transfer result, if it holds up, would be the part worth testing in practice. The soft spot is the absence of numbers. The abstract gives no effect sizes, no ablation on how much context helps versus baselines, and no direct comparison to prior CAF datasets, so the central claim that dataset quality was the bottleneck stays hard to evaluate from the summary alone. The stress-test worry about artificial complementarity is worth checking in the methods: if the generation rule pulls too much from patterns already in the numerical windows, the apparent utilization could be an artifact rather than genuine cross-modal reasoning. This is for people working on multimodal time series or applied forecasting who have run into the gap before. A reader looking for a large, verifiable dataset and a data-first hypothesis will get something concrete to try. It deserves peer review because the scale and the hypothesis are worth proper scrutiny even if the current evidence is light.

Referee Report

2 major / 2 minor

Summary. The paper claims that the modality gap in context-aided forecasting arises primarily from poor context quality in existing datasets rather than model architectures. It introduces a semi-synthetic data augmentation procedure to generate contexts that are both descriptive of temporal dynamics and verifiably complementary to numerical time series, enabling creation of the CAF-7M corpus (7 million context-augmented windows with a rigorously verified test set). The work reports that pre-training on this semi-synthetic data transfers effectively to real-world evaluation and yields clear evidence of context utilization, concluding that dataset quality has been the central bottleneck.

Significance. If the empirical claims hold, the result would redirect research in multimodal time-series forecasting from architectural innovation toward scalable, verifiable data augmentation and curation. A large, transferable pre-training corpus could enable more reliable integration of domain knowledge and forward-looking signals, addressing a persistent empirical puzzle in the field.

major comments (2)

[Abstract / Experiments] Abstract and Experiments: the assertions of effective transfer, context utilization, and dataset quality as the primary bottleneck are presented without quantitative metrics, baseline comparisons, ablation studies, or verification statistics for complementarity. This absence prevents assessment of whether the semi-synthetic contexts are independently descriptive or merely aligned by construction with the numerical windows.
[Method] Method: the semi-synthetic augmentation procedure must demonstrate that generated contexts are not derived from latent patterns already present in the time-series windows; without explicit, independent verification steps, the reported transfer success cannot rule out artificial complementarity that would not generalize to noisy or externally sourced contexts in prior datasets.

minor comments (2)

[Dataset] Provide the exact composition of CAF-7M (number of real vs. synthetic windows, context length distribution, and verification criteria for the test set) to allow reproducibility.
[Preliminaries] Clarify notation for context tokens versus numerical history tokens when describing the multimodal input format.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript accordingly to provide the requested quantitative validation and verification details.

read point-by-point responses

Referee: [Abstract / Experiments] Abstract and Experiments: the assertions of effective transfer, context utilization, and dataset quality as the primary bottleneck are presented without quantitative metrics, baseline comparisons, ablation studies, or verification statistics for complementarity. This absence prevents assessment of whether the semi-synthetic contexts are independently descriptive or merely aligned by construction with the numerical windows.

Authors: We agree that the current presentation lacks sufficient quantitative support. In the revised manuscript we will expand the Experiments section to include explicit baseline comparisons (multimodal vs. unimodal), ablation studies on context quality levels, and verification statistics (e.g., independence tests, complementarity scores, and transfer metrics) that demonstrate the contexts are independently descriptive rather than aligned solely by construction. revision: yes
Referee: [Method] Method: the semi-synthetic augmentation procedure must demonstrate that generated contexts are not derived from latent patterns already present in the time-series windows; without explicit, independent verification steps, the reported transfer success cannot rule out artificial complementarity that would not generalize to noisy or externally sourced contexts in prior datasets.

Authors: We acknowledge this limitation in the current method description. We will add explicit independent verification steps to the Methods section, including statistical tests for independence from latent time-series patterns, cross-checks against external real-world contexts, and metrics confirming that complementarity is not artificial. These additions will directly address generalizability concerns. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical semi-synthetic augmentation is self-contained

full rationale

The paper introduces a semi-synthetic data augmentation procedure as a new contribution to create CAF-7M, with the central claim resting on experimental transfer results rather than any derivation chain. No equations, fitted parameters renamed as predictions, self-definitional constructs, or load-bearing self-citations appear in the described method. The augmentation is presented as independently generating descriptive and complementary contexts, and the evaluation on real-world transfer is external to the generation rule itself. This is the standard case of an empirical ML paper whose claims are falsifiable via the reported benchmarks and do not reduce to their inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unverified assumption that the generated contexts are verifiably complementary and that pre-training on them transfers to real data; no free parameters or invented entities are mentioned.

axioms (1)

domain assumption Semi-synthetic contexts can be generated to be both descriptive of temporal dynamics and verifiably complementary to numerical histories
This is the load-bearing premise stated in the abstract for the method's validity and transfer results.

pith-pipeline@v0.9.0 · 5471 in / 1137 out tokens · 33510 ms · 2026-05-15T12:26:33.363226+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 3 internal anchors

[1]

Jialin Chen, Aosong Feng, Ziyu Zhao, Juan Garza, Gaukhar Nurbek, Cheng Qin, Ali Maatouk, Leandros Tassiulas, Yifeng Gao, and Rex Ying

URLhttps://openreview.net/forum?id=gerNCVqqtR. Jialin Chen, Aosong Feng, Ziyu Zhao, Juan Garza, Gaukhar Nurbek, Cheng Qin, Ali Maatouk, Leandros Tassiulas, Yifeng Gao, and Rex Ying. MTBench: A multimodal time series benchmark for temporal reasoning and question answering.arXiv preprint arXiv:2503.16858,

work page arXiv
[2]

T2s: High-resolution time series generation with text-to-series diffusion models.arXiv preprint arXiv:2505.02417,

Yunfeng Ge, Jiawei Li, Yiji Zhao, Haomin Wen, Zhao Li, Meikang Qiu, Hongyan Li, Ming Jin, and Shirui Pan. T2s: High-resolution time series generation with text-to-series diffusion models.arXiv preprint arXiv:2505.02417,

work page arXiv
[3]

The Llama 3 Herd of Models

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

work page internal anchor Pith review Pith/arXiv arXiv
[4]

Haoxin Liu, Shangqing Xu, Zhiyuan Zhao, Lingkai Kong, Harshavardhan Kamarthi, Aditya B

URLhttps://arxiv.org/abs/2310.01728. Haoxin Liu, Shangqing Xu, Zhiyuan Zhao, Lingkai Kong, Harshavardhan Kamarthi, Aditya B. Sasanur, Megha Sharma, Jiaming Cui, Qingsong Wen, Chao Zhang, and B. Aditya Prakash. Time-MMD: A new multi-domain multimodal dataset for time series analysis, 2024a. Haoxin Liu, Shangqing Xu, Zhiyuan Zhao, Lingkai Kong, Harshavardha...

work page arXiv 2024
[5]

URLhttps://arxiv.org/abs/2404.11757. OpenAI. GPT-5.2 API.https://platform.openai.com/,

work page arXiv
[6]

Zijie Pan, Yushan Jiang, Sahil Garg, Anderson Schneider, Yuriy Nevmyvaka, and Dongjin Song.s2 IP-LLM: Semantic space informed prompt learning with LLM for time series forecasting

Accessed: 2026-01-27. Zijie Pan, Yushan Jiang, Sahil Garg, Anderson Schneider, Yuriy Nevmyvaka, and Dongjin Song.s2 IP-LLM: Semantic space informed prompt learning with LLM for time series forecasting. InForty-first International Conference on Machine Learning,

work page 2026
[7]

Layer by Layer: Uncovering Hidden Representations in Language Models

Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv. Layer by Layer: Uncovering hidden representations in language models.arXiv preprint arXiv:2502.02013,

work page internal anchor Pith review arXiv
[8]

Context is Key: A benchmark for forecasting with essential textual information

11 Andrew Robert Williams, Arjun Ashok, Étienne Marcotte, Valentina Zantedeschi, Jithendaraa Subramanian, Roland Riachi, James Requeima, Alexandre Lacoste, Irina Rish, Nicolas Chapados, and Alexandre Drouin. Context is Key: A benchmark for forecasting with essential textual information. InProceedings of the 2025 International Conference on Machine Learnin...

work page 2025
[9]

URLhttps://openreview.net/forum?id=ih2WuBT1Fn. Poster. Wenfa Wu, Guanyu Zhang, Zheng Tan, Yi Wang, and Hongsheng Qi. Dual-forecaster: A multimodal time series model integrating descriptive and predictive texts.arXiv preprint arXiv:2505.01135,

work page arXiv
[10]

FinMultiTime: A four-modal bilingual dataset for financial time-series analysis.arXiv preprint arXiv:2506.05019,

Wenyan Xu, Dawei Xiang, Yue Liu, Xiyu Wang, Yanxiang Ma, Liang Zhang, Chang Xu, and Jiaheng Zhang. FinMultiTime: A four-modal bilingual dataset for financial time-series analysis.arXiv preprint arXiv:2506.05019,

work page arXiv
[11]

Qwen3 Technical Report

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

work page internal anchor Pith review Pith/arXiv arXiv
[12]

Does multimodality lead to better time series forecasting?arXiv preprint arXiv:2506.21611,

Xiyuan Zhang, Boran Han, Haoyang Fang, Abdul Fatir Ansari, Shuai Zhang, Danielle C Maddix, Cuixiong Hu, Andrew Gordon Wilson, Michael W Mahoney, Hao Wang, et al. Does multimodality lead to better time series forecasting?arXiv preprint arXiv:2506.21611,

work page arXiv
[13]

to appear

URLhttps://arxiv.org/ab s/2502.04395. to appear. Xin Zhou, Weiqing Wang, Francisco J Baldán, Wray Buntine, and Christoph Bergmeir. MoTime: A dataset suite for multimodal time series forecasting.arXiv preprint arXiv:2505.15072,

work page arXiv
[14]

Block structure and ordering.A DualT5 decoder block consists of the following sublayers in sequence: (i) masked self-attention, (ii) standard Chronos encoder–decoder cross-attention layer, overets, (iii) DualT5 cross-attention layer, overectx, and (iv) a feed-forward network (FFN). Each sublayer uses the standard T5 pre-norm convention: layer normalizatio...

work page 2025
[15]

Unless otherwise stated, DoubleCast uses the structured template in all experiments. " <i n f o> f o r e c a s t_s t a r t_d a t e =2025 -06 -25 0 0 : 0 0 : 0 0 f r e q u e n c y= D s c a l e_f a c t o r =0.1234 </ i n f o> <c o n t e x t> Background : The t i m e s e r i e s e x h i b i t s s e a s o n a l f l u c t u a t i o n s . . . </ c o n t e x t> ...

work page 2025
[16]

For a window with horizon lengthT, let ˆy1:T be the Chronos point forecast andy1:T the ground truth

We apply the same per-window capc = 5and report the mean over a split: nMAE = 1 M MX i=1 min{nMAEi, c}.(17) 25 MASE.We also compute the Mean Absolute Scaled Error (MASE) to characterize per-window forecast difficulty. For a window with horizon lengthT, let ˆy1:T be the Chronos point forecast andy1:T the ground truth. Let z1:L denote the historical context...

work page 2025
[17]

broadens the modality set by coupling vision-style encodings of temporal patterns with generated textual descriptions, then leveraging a frozen vision-language backbone to provide augmented representations that particularly help in few/zero-shot regimes. D.2 Prompt-Based Methods Another line of work prompt LLMs with both numerical and contextual inputs to...

work page 2023
[18]

In parallel, AutoTimes (Liu et al., 2024d) repurposes decoder-only LLMs as autoregressive forecasters by projecting time series into the language token space; Wang et al

learns prompts in a joint semantic space that aligns pre-trained LLM embeddings with time-series features, improving transfer while reducing heavy fine-tuning. In parallel, AutoTimes (Liu et al., 2024d) repurposes decoder-only LLMs as autoregressive forecasters by projecting time series into the language token space; Wang et al. (2024) use LLM agents to r...

work page 2024
[19]

For zero-shot evaluation, we use the publicly released checkpoints and settings reported by Williams et al

architecture using the authors’ official codebase (https: //github.com/KimMeen/Time-LLM). For zero-shot evaluation, we use the publicly released checkpoints and settings reported by Williams et al. (2025). For in-domain adaptation on CAF-7M, we further train Time-LLM on ourindomain-train split. Concretely, we train for400,000optimization steps with BF16 m...

work page 2025
[20]

context helps

aggregate outcomes by dataset and report: counts of judge-passed and DP-accepted windows; the judge pass rate (passed divided by total); and the “context helps” rate (accepted divided by passed). zeroshot-test(hard). 29 0 200 400 600 800 1000 1200 1400 1600 Count Processed Passed Judge Valid 1,602 (100.0%) 1,135 (70.8%) 490 (30.6%) Sample Acceptance Funne...

work page 2000
[21]

A plausible cause for this change is a supply chain disruption, which would lead to stockouts and lost sales, resulting in significantly lower profits

Scenario: Starting from 1994-04-21 00:00:00, a major change occurs, with the profit values dropping to 0 for several weeks, followed by intermittent returns to the usual value of 53.48. A plausible cause for this change is a supply chain disruption, which would lead to stockouts and lost sales, resulting in significantly lower profits. This disruption is ...

work page 1994
[22]

This downturn led to a reduction in production capacity, causing values to be approximately 20% lower than the previous year

Scenario: Suppose a major economic downturn occurred from 1991 onwards, resulting in a substantial decline in industrial production. This downturn led to a reduction in production capacity, causing values to be approximately 20% lower than the previous year. The impact of this economic downturn is expected to persist, with production levels continuing to ...

work page 1991
[23]

After this period, sales returned to the usual low levels

Scenario: A major supplier change occurred from 2013-03-28 00:00:00 to 2013-04-30 00:00:00, resulting in a significant increase in product availability, leading to sales values being 2-5 times higher than usual during this period. After this period, sales returned to the usual low levels. Constraint: The sales values are generally bounded above by

work page 2013
[24]

History Future DC (noctx) DC Chronos Figure 37Examples of windows in zeroshot-test (hard) where the context was found to be helpful when forecasting using DoubleCast. 1970-04-0100:00 1970-07-0100:00 1970-10-0100:00 1971-01-0100:00 1971-04-0100:00 1971-07-0100:00 1971-10-0100:00 1972-01-0100:00 1972-04-0100:00 Time 600 800 1000 1200 1400 1600 1800Value Con...

work page 1970
[25]

This change is expected to persist throughout the forecast horizon

Scenario: A major regulatory change occurred from 1971-04-30 00:00:00 onwards, resulting in stricter financial regulations and a subsequent decline in financial activities, leading to values approximately 30% lower than the preceding months. This change is expected to persist throughout the forecast horizon. History Future DC (noctx) DC Chronos 2003-09-01...

work page 1971
[26]

This downturn could be attributed to a reduction in consumer spending and business investments, leading to lower employment figures

Scenario: A significant economic downturn occurred from 2006-01-31 00:00:00 to 2006-11-30 00:00:00, resulting in a sustained decline of approximately 50% in the indicator values compared to the previous year. This downturn could be attributed to a reduction in consumer spending and business investments, leading to lower employment figures. The impact of t...

work page 2006
[27]

This shift is expected to result in a sustained decline of approximately 20% in tourism metrics over the next few years

Scenario: A significant change in the forecast horizon, starting from 1998-12-31 00:00:00, can be attributed to a shift in global travel preferences, with tourists increasingly opting for alternative destinations due to rising concerns over environmental sustainability and cultural sensitivity. This shift is expected to result in a sustained decline of ap...

work page 1998
[28]

However, from 1994-09-08 00:00:00 onwards, the product is expected to be discontinued, leading to zero profits

Scenario: A major supplier for this product is expected to increase production capacity from 1994-06-30 00:00:00 to 1994-09-08 00:00:00, resulting in a temporary surge in profit, approximately 30% higher than the previous stable value, due to increased demand and reduced production costs. However, from 1994-09-08 00:00:00 onwards, the product is expected ...

work page 1994
[29]

This pattern is expected to persist in the forecast horizon

History Future DC (noctx) DC Chronos 1992-10-0100:00 1993-01-0100:00 1993-04-0100:00 1993-07-0100:00 1993-10-0100:00 1994-01-0100:00 1994-04-0100:00 1994-07-0100:00 Time 0 5 10 15 20 25 30 35Value Context: Background: The historical time series shows a pattern of fluctuating profit values, with some weeks having zero profit and others having varying level...

work page 1992
[30]

Historically, the data shows significant fluctuations, with values ranging from 2500 to 5800 units. Scenario: Suppose a major economic restructuring occurs from 1985-12-31 12:00:00 to 1987-12-31 12:00:00, resulting in a significant shift in production priorities and a subsequent 40% decline in output. This restructuring leads to a sustained decrease in pr...

work page 1985
[31]

Scenario: Suppose a new variant of the virus emerges around May 5, 2020, leading to a slight increase in death counts

This suggests a gradual spread of the pandemic. Scenario: Suppose a new variant of the virus emerges around May 5, 2020, leading to a slight increase in death counts. This variant is more contagious but not significantly more lethal, resulting in a sustained, moderate rise in death counts. The impact of this new variant is expected to be around 10% to 20%...

work page 2020
[32]

This downturn is expected to last for several years, affecting the entire forecast period

Scenario: Suppose a major economic downturn occurs from 1989-12-31 00:00:00 onwards, resulting in reduced demand for industrial products and leading to a sustained decline in production, with values expected to be around 20% lower than the previous year's average. This downturn is expected to last for several years, affecting the entire forecast period. H...

work page 1989
[33]

The historical data shows fluctuating sales with occasional peaks, indicating potential seasonality or event-driven changes. Scenario: A major supplier of car parts announces a significant price reduction from 2001-03-01 00:00:00 to 2001-08-01 00:00:00, leading to a surge in sales as customers take advantage of the discounted prices. This price reduction ...

work page 2001
[34]

Notably, there are periods of zero sales, which could be due to various factors such as inventory management, supplier issues, or changes in demand

The historical data shows fluctuating sales with occasional peaks, indicating possible seasonal or event-driven trends. Notably, there are periods of zero sales, which could be due to various factors such as inventory management, supplier issues, or changes in demand. Scenario: Suppose a major automotive manufacturing plant in the region undergoes a signi...

work page 2001
[35]

History Future DC (noctx) DC Chronos Figure 52Examples of windows in zeroshot-test (easy) where the context was found to not be helpful when forecasting using DoubleCast. 1976-01-0100:00 1980-01-0100:00 1984-01-0100:00 1988-01-0100:00 1992-01-0100:00 1996-01-0100:00 Time 2500 5000 7500 10000 12500 15000 17500 20000 22500Value Context: Background: This dat...

work page 1976
[36]

This downturn could lead to a decline in passenger numbers, potentially 15% lower than the previous year's peak

Scenario: Suppose a global economic downturn occurs from 1992-12-31 00:00:00 to 1995-12-31 00:00:00, resulting in reduced travel due to financial constraints. This downturn could lead to a decline in passenger numbers, potentially 15% lower than the previous year's peak. The economic recovery starting from 1996-12-31 00:00:00 could then lead to an increas...

work page 1992
[37]

This downturn affects the entire year of

Scenario: Suppose a major economic downturn occurs from 2010-01-01 00:00:00 to 2010-12-31 00:00:00, resulting in reduced demand and a subsequent 20% decrease in production. This downturn affects the entire year of

work page 2010
[38]

History Future DC (noctx) DC Chronos Figure 53Examples of windows in zeroshot-test (easy) where the context was found to not be helpful when forecasting using DoubleCast

The production levels are expected to rebound in the following years as the economy recovers. History Future DC (noctx) DC Chronos Figure 53Examples of windows in zeroshot-test (easy) where the context was found to not be helpful when forecasting using DoubleCast. 38 1978-01-0100:00 1980-01-0100:00 1982-01-0100:00 1984-01-0100:00 1986-01-0100:00 1988-01-0...

work page 1978