[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-81667-en":3,"doc-seo-81667-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},81667,16904993612988,"Olivia Brown","https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd",8,"Research & Report","CITRAS-FM: Tiny Time Series Foundation Model for Covariate-Informed Zero-Shot Forecasting","Pretrained time series foundation models enable zero-shot forecasting, yet existing approaches are often computationally heavy and limited in handling diverse variable types, especially when exogenous covariates drive target variability. CITRAS-FM proposes a compact 7M-parameter TSFM for univariate, multivariate, and covariate-informed zero-shot forecasting with real-time CPU inference. A patch-based decoder-only Transformer incorporates Shifted Attention for covariate usage across the horizon, while CovSynth synthesizes realistic covariates from decomposed target components. Experiments on fev-bench (100 tasks) show leading sub-10M accuracy with sub-0.1-second CPU inference.","CITRAS-FM: Tiny Time Series Foundation Model for Covariate-Informed Zero-Shot Forecasting  \nYosuke Yamaguchi, Issei Suemitsu, Yuki Kajihara, Wenpeng Wei  \nResearch & Development Group, Hitachi Ltd., Tokyo, Japan  \n{yosuke.yamaguchi.fy, issei.suemitsu.rj, [yuki.kajihara.fj](yuki.kajihara.fj), [wenpeng.wei.bo](wenpeng.wei.bo}@hitachi.com)[}](wenpeng.wei.bo}@hitachi.com)[@hitachi.com](wenpeng.wei.bo}@hitachi.com)  \narXiv :2606 . 10798v2 [ cs .LG] 10 Jul 2026  \nAbstract—Pretrained time series foundation models (TSFMs) have enabled zero-shot forecasting on unseen target series. However, existing TSFMs often incur high computational cost and provide limited support for diverse variable types, often failing to account for covariates that exogenously influence target variability. To address these challenges, we propose CITRAS-FM, a tiny 7M-parameter TSFM that supports univariate, multivariate, and covariate-informed zero-shot forecasting with realtime CPU inference. Built on a patch-based, decoder-only Transformer, CITRAS-FM introduces Shifted Attention into the cross-variate module to effectively exploit known covariates accessible throughout the forecast horizon. Moreover, to enable covariate-aware pretraining despite the scarcity of covariaterich corpora, we propose CovSynth, which synthesizes realistic covariates from decomposed components of target series. Experiments on fev-bench, spanning 100 tasks across various settings, demonstrate that CITRAS-FM achieves state-of-the-art zero-shot accuracy among sub-10M TSFMs while delivering sub-0.1-second CPU inference, offering a strong balance between forecasting accuracy and real-time deployability. The code is available at [https://github.com/hitachi-ais/citras-fm](https://github.com/hitachi-ais/citras-fm).  \nIndex Terms—Time Series, Foundation Model, Forecasting, Covariate, Transformer  \nI. INTRODUCTION  \nTime series forecasting—predicting the future values of target variables—is widely used across domains such as server load forecasting [1] and energy demand forecasting [2] . Recently, time series foundation models (TSFMs) [3] have gained traction. These models are trained on massive datasets to learn generalizable temporal dynamics, enabling zero-shot forecasting on unseen datasets without task-specific training. This paradigm is particularly promising in industrial environments where data distributions shift continually and collecting sufficient historical data for each deployment is impractical. However, two shortcomings limit the practical use of existing TSFMs.  \nFirst, most leading TSFMs on popular benchmarks are computationally expensive. For example, TimesFM-2.5 [3] and Chronos-2 [4] contain over a hundred million parameters, resulting in slow inference on limited computational resources. Many real-world systems—server clusters or sensor-dense manufacturing equipment—require on-device low-latency operation [5] . Although lightweight TSFMs exist, they typically lag behind larger models in forecasting accuracy [6] and often require fine-tuning to achieve competitive performance [7] .  \nAccepted to EUSIPCO 2026 .  \nSecond, the majority of existing TSFMs are restricted to relatively simple univariate settings. In practice, applications frequently demand simultaneous forecasting of multiple targets (multivariate), or the incorporation of covariates that represent exogenous signals. For instance, in server metric forecasting, it is crucial to maintain consistency between CPU temperature and power consumption [1]; in energy demand forecasting, considering the impact of temperature and holiday calendar is essential [8] . However, TSFM architectures that can accommodate such diverse variables remain largely underexplored. Furthermore, large-scale pretraining corpora containing rich covariates are scarce [9] . Consequently, as summarized in Table I, most TSFMs support only a limited subset of variable types in the zero-shot setting. To the best of our knowledge, Chronos-2 [4] is the o","cbCailikHzZ8oRFs","https://ap.wps.com/l/cbCailikHzZ8oRFs","pdf",1018695,3,1,6,"English","en",105,"# Introduction\n## Computational cost and latency limits\n## Limited variable types and scarce covariates\n# Proposed Approach: CITRAS-FM\n## Shifted Attention for covariate-informed prediction\n## CovSynth for covariate-aware pretraining\n# Experimental Evaluation\n## Accuracy on fev-bench and comparisons\n## Inference efficiency","[{\"question\":\"What limitations of existing time series foundation models does CITRAS-FM address?\",\"answer\":\"It targets high computational cost and limited support for diverse variable types, particularly inadequate incorporation of exogenously influencing covariates.\"},{\"question\":\"How does CITRAS-FM incorporate covariates during zero-shot forecasting?\",\"answer\":\"It uses a patch-based decoder-only Transformer with a Shifted Attention layer in the cross-variate module to exploit known covariates across the forecast horizon.\"},{\"question\":\"What is CovSynth, and why is it introduced?\",\"answer\":\"CovSynth synthesizes pseudo-covariates from decomposed components of the target series, enabling covariate-aware pretraining despite limited covariate-rich corpora.\"}]",1784175298,15,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"citras-fm-tiny-time-series-foundation-model-for-covariate-informed-zero-shot-forecasting","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,50],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":20},"https://docshare.wps.com/document/research-report/",{"item":51,"name":13,"@type":43,"position":52},"https://docshare.wps.com/document/citras-fm-tiny-time-series-foundation-model-for-covariate-informed-zero-shot-forecasting/81667/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-26","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What limitations of existing time series foundation models does CITRAS-FM address?","Question",{"text":75,"@type":76},"It targets high computational cost and limited support for diverse variable types, particularly inadequate incorporation of exogenously influencing covariates.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does CITRAS-FM incorporate covariates during zero-shot forecasting?",{"text":80,"@type":76},"It uses a patch-based decoder-only Transformer with a Shifted Attention layer in the cross-variate module to exploit known covariates across the forecast horizon.",{"name":82,"@type":73,"acceptedAnswer":83},"What is CovSynth, and why is it introduced?",{"text":84,"@type":76},"CovSynth synthesizes pseudo-covariates from decomposed components of the target series, enabling covariate-aware pretraining despite limited covariate-rich corpora.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,114,119,122,127,130,134],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":22,"doc_module":4,"doc_module_name":46,"category_name":111,"show_sort_weight":112,"slug":113},"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":46,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":46,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":46,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":46,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]