← PROJECTS
2026 Preprint · Deep Learning

TimeBraid:Unifying Time Series and Language for Understanding and Forecasting

Understand, forecast, and shape time series with language.

Xinyue Wang · Jiacheng Pang · Kun Zhou · Kexin Zhang · Defu Cao · Fan Feng · Faisal · Songyao Jin · Yan Liu · Biwei Huang

The idea

Time and language,
in their native forms.

TimeBraid brings pretrained language and time-series models into one framework. It connects language knowledge and reasoning with continuous-signal perception and forecasting, so the same model can understand a signal, explain it in words, and generate what comes next.

TimeBraid's overview: the shared interface where time series and text both enter and leave the model; twelve application tasks across forecasting, control, reconstruction, perception, understanding and reasoning; and results across six benchmarks.
Figure 1. One set of weights covers the whole map. Series and text enter as patches and tokens and leave the same way; twelve tasks span forecasting, control, reconstruction and temporal understanding; the strip below reports six benchmarks, where SFT marks TSAQA-specific fine-tuning.
Understand & explain ↗Condition & generate ↗Forecast across domains ↗Reason & write ↗

Method

Three roles. One shared context.

Language tokens and numerical time-series patches pass through their pretrained backbones. Interleaved global residual attention connects the streams at multiple depths: each token can draw on earlier context from both modalities, and the updates return to each backbone through residual connections.

Perception, reasoning, and forecasting experts interact through interleaved global residual attention over language tokens and time-series patches, with a detail view of the shared causal attention step.
Figure 3. Native processing within each modality, with global causal attention across the mixed sequence. The detail on the right follows one residual attention layer: each expert projects its own queries, keys and values, the streams align by global position, then regroup and return to their backbones. The perception and forecasting roles share a time-series backbone in the dual-tower design.

Perception

Encodes observed time-series segments and the structure within them.

Reasoning

Processes instructions and context, and generates language responses.

Forecasting

Generates future time-series segments from numerical and textual context.

One interface for signals and language.

TimeBraid treats a task as an ordered sequence of text and time-series segments. Each output draws on the segments before it. Multiple variables enter as separate signals, each bound to a <ts></ts> marker. Text becomes language tokens; signals become patches of 32 continuous values.

Understand multiple signals

User Equity observed series Stress observed series

How do they relate?

Assistant Stress leads each equity drawdown.

Generate a conditioned forecast

User CPU load observed series

A deployment just shipped. Forecast the next 96 hours.

Assistant future numerical patches

Illustrative prompt structures from Section 3.2. Context time-series segments carry their length, mean and standard deviation. Numerical payloads remain separate from the text transcript.

Generate continuous values directly.

The forecasting expert predicts numerical patches autoregressively, with a central trajectory and nine quantiles for uncertainty. Text responses use next-token prediction. The two output types learn together through separate language and forecasting losses.

Training recipe

Align first.
Learn diverse tasks next.

A two-stage curriculum connects temporal structure with language, then extends that foundation to instruction following and real-world tasks.

Stage 1 · Alignment

2.23M paired examples

Descriptions, analysis and forecasting examples teach the model how signals relate to language and how different conditions lead to different futures.

30% understanding · 70% forecasting

Stage 2 · Instruction tuning

4.88M samples

A diverse mixture covers forecasting, question answering, reasoning and captioning across energy, climate, healthcare, finance, sensing and industrial monitoring.

Domain tasks · Unimodal data · Alignment retention

Explore the training data mixture

Understanding pairs combine morphology and context-rich captions, attribute-first synthetic signals, structural causal models and programmable questions. Forecast-control pairs give one history several possible futures, each described by its own condition.

Forecast annotations are created with access to the future and can describe the target interval; future numerical values are excluded from the model input. These pairs teach conditional generation under stated scenarios.

The two-stage TimeBraid data recipe: alignment pairs followed by instruction tuning across forecasting, reasoning, captioning and unimodal tasks.
Figure 2. Stage 2 contains 4,881,583 examples, including alignment retention. Its percentages describe training sampling shares; the counts show the dataset inventory.
Model scales, attention and training setup
Model family
TimeBraid-1.2B, TimeBraid-2.5B and TimeBraid-6.7B are named for their total parameter counts: 1.205B, 2.497B and 6.676B. They pair Qwen3-0.6B, Qwen3-1.7B and Qwen3-4B language towers, respectively, with TimesFM-derived time-series towers and cross-modal attention layers.
Cross-modal attention
Stream-specific projections map each backbone into a shared causal attention space. Zero-initialized output projections make the new layers start as identity maps. Local signal positions restart within each segment; global attention follows the full mixed sequence.
Signal normalization
Observed segments use their full context statistics. Targets use the history prefix and rolling causal normalization. Length, mean and standard deviation carry scale in the text, while normalized patches carry temporal shape.
Joint objective
Next-token cross-entropy trains text. Multi-patch prediction trains numerical targets with a point loss and quantile losses at 0.1–0.9. Their weight in the joint objective is 0.5.
Optimization
All parameters train with AdamW, learning rate 2 × 10−5, a 500-step warmup, batch size 128 and BF16. Alignment runs for one epoch at context length 2,560; instruction tuning runs for 30k steps at context length 4,096.

Paper benchmarks

Understanding and forecasting, together

Aggregate results under the paper’s evaluation protocols, across temporal understanding, contextual forecasting and language.

TSAQA

Understand patterns, anomalies and temporal relationships.

ModelOverall ↑Accuracy (%)
TimeBraid-2.5B78.31
TimeBraid-6.7B80.65
LLaMA3.1-8B (SFT)85.26
Closed-source reference · not ranked
Gemini-2.5-Flash65.08
GPT-5.463.10

Official test-set accuracy. The SFT baseline is finetuned on TSAQA training data.

TimeSeriesExam

Reason about patterns, noise, anomalies, similarity and causality.

ModelOverall ↑Accuracy (%)
TimeBraid-2.5B60.05
TimeBraid-6.7B63.14
TS-Reasoner-7B54.26
Closed-source reference · not ranked
GPT-4.1 (vision)67.89
GPT-5.467.83

Overall multiple-choice accuracy. GPT-4.1 uses the vision input setting.

CaTS-Bench

Describe time series with accurate language and numbers.

ModelDeBERTa F1 ↑Score [0, 1]SimCSE ↑Score [0, 1]Numeric Fidelity ↑Score [0, 1]
TimeBraid-2.5B0.7080.8850.670
TimeBraid-6.7B0.7120.8860.684
Idefics 2 (finetuned)0.7130.8850.748
Qwen3-8B0.6700.8650.678
LLaMA3.1-8B-Instruct0.6670.8590.697
Closed-source reference · not ranked
Gemini 2.5 Flash0.6940.8870.802

Human-rewritten split. The text-only baselines receive series and metadata without plots; Idefics 2 is finetuned on CaTS-Bench.

TemporalBench

Connect temporal perception, prediction and contextual reasoning.

ModelOverall ↑Macro accuracy (%)
TimeBraid-2.5B35.74
TimeBraid-6.7B36.58
Closed-source reference · not ranked
GPT-5.438.50
GPT-4.136.91
Gemini-2.5-Flash34.61

Macro accuracy over 16 tasks in four domains. Forecasting errors and task breakdowns are in the full results.

Selected models and headline metrics from each benchmark. TimeBraid model sizes count all parameters. Bold marks the best displayed value. Following the paper, closed-source commercial models are listed as reference points below the divider and stay outside the ranking. ± denotes standard error.

How the evaluation relates to training

Evaluation examples are held out at the split level. TemporalBench MCQ, CiK, and ETT/Weather are outside the training mixture. TSAQA, CaTS, TimeMMD, CGTSF and CAF contribute training-split data.

TimeSeriesExam and Ctrl-F share their data-generating process with independently generated alignment examples. Part of GIFT-Eval’s training split provides seed data for generation. These settings test different forms of generalization.

Ablations & findings

What makes joint learning work?

Six research questions test the interface, architecture, optimization, curriculum and use of text. The results explain the choices behind TimeBraid.

RQ1 · Interaction space

Meet in a shared space.

Residual attention reaches substantially lower language loss than projecting time-series features directly into the language space with an MLP connector.

RQ2 · Expert separation

Two roles can share one backbone.

Separate perception and forecasting towers yield nearly identical training-loss curves to a shared time-series tower in the tested alignment runs.

RQ3 · Optimization stability

Keep regression spikes in check.

Nonstationary targets can produce large forecasting losses that also disrupt language learning. Robust normalization compresses extreme values; response-level loss capping limits their contribution.

RQ4 · Modality balance

Give forecasting more of the data.

The 30:70 understanding-to-forecasting mixture improves forecasting loss over an even split without worsening language loss. The final recipe weights the time-series loss by 0.5.

Compare the architecture and optimization ablations
Ten-thousand-step training-loss comparisons of residual attention and an MLP connector, learning rate, normalization, data ratio and shared versus separate time-series towers.
Figure 4. Each panel compares 10k-step alignment runs that differ in one factor. These are training-loss comparisons; downstream alignment results follow below.

RQ5 · Alignment before instruction tuning

Alignment helps when language matters.

Aligned initialization stays ahead on TSAQA and CAF throughout instruction tuning. TimeMMD finishes at similar error with either initialization, showing that the benefit depends on the task and its use of text.

Aligned initialization compared with direct instruction tuning over time: TSAQA accuracy improves, CAF forecast error decreases, and TimeMMD ends at similar error.
Figure 5. Aligned versus direct SFT with the same instruction-tuning recipe. TSAQA uses a 10% test subset; TimeMMD reports MSE and CAF reports normalized qCRPS.

RQ6 · Text strength at inference

Control how much the forecast follows text.

Run the same model with history alone and with history plus text, then mix the two point forecasts. The weight λ moves from the history-only prediction at 0 to the fully conditioned prediction at 1.

Forecast(1 − λ) × history-onlyλ × text-conditioned

In the TimeMMD sweep, the effect of text varies with the domain and history length. Security favors history alone; Climate and Social Good change with history length. The TimeMMD comparison uses one validation-selected λ = 0.3 across all domains.

Explore text guidance across nine domains
Sensitivity to text-guidance weight across nine TimeMMD domains and input histories from 8 to 128 steps.
Figure 6. Relative MSE change against history-only forecasting. This is a sensitivity analysis; the released text is not certified as available at each historical forecast date. The sweep is separate from validation-based model selection.
Current limitations and directions

Numerical detail. Patch compression can lose exact local values and make long, high-frequency or noisy waveforms difficult to read. CaTS captions still miss numerical details compared with stronger captioning baselines.

Forecast control. Abrupt changes that contradict the history are followed less reliably. Complex or weakly related text remains difficult to translate into numerical consequences.

Training coverage. The recipe moves from alignment directly to instruction tuning. Broader interleaved pretraining data and better language-retention mixtures remain open directions.

Explore TimeBraid

See the model in action.

Ask about temporal patterns, reason across signals, or change the conditions of a forecast. Explore language and time series through the same model.

Selected field cases · 2024–2026

Forecasting the world as it happens.

Freshly collected public series across markets, the macro economy, air quality, weather, power grids and the web. These selected cases show TimeBraid outperforming the baselines displayed here; see the paper benchmarks for aggregate results.

TimeBraid sees nothing but each series’ own past values: no name, no unit, no dates. The baselines receive exactly the same numbers.

Equities · US equities

Nasdaq Composite

Forecast Jun 2 – Aug 31, 2026 · next 63 trading days
Input: 512 past observations only

Nasdaq Composite: forecasts against what happened19.8k23.7k27.6k31.5k-188-138-88-371363index pointstrading days from forecast originHistoryGemini 3.8 FlashClaude Sonnet 5GPT-5.5TimeBraidWhat happened
TimeBraidGPT-5.5Claude Sonnet 5Gemini 3.8 FlashWhat happened

Series: Yahoo Finance chart API ↗

What contextual text changes

Public health · Time-MMD

Follow seasonal influenza activity

Same history. Add the context.
One shared input 36 observations
Follow seasonal influenza activity0.223.95.7-35-180ILI ratio (%)WeekShared history
∿TimeBraid All 36 future values
Same history, with and without context02.95.88.8-7210192736ILI ratio (%)WeekShared history · last observationsWith contextual textWithout contextual textObserved future
With contextual textWithout contextual textObserved future

AWith contextual text

Forecast the next 36 values of the time series.
Domain: public health.

Context information:
Available facts are as follows: 2019-02-04: The Centers for Disease Control and Prevention (CDC) recommends vaccination against hepatitis A for certain populations at increased risk, including the civilian population of the United States. [Source: www.cdc.gov]; Influenza vaccine effectiveness in the United States during the 2015–2016 season was studied, and the results were published in the New England Journal of Medicine in 2017. [Source: www.ncbi.nlm.nih.gov]; Rapid diagnostic testing for influenza is available, and the CDC provides information on currently available tests and approved respiratory specimens. [Source: www.cdc.gov]; Influenza infections are responsible for substantial numbers of hospitalizations and deaths each year, and the US recommends vaccination for all individuals >6 months of age. [Source: www.ncbi.nlm.nih.gov]; Neuraminidase inhibition contributes to influenza A virus protection, and this is a work of the U.S. Government. [Source: pubmed.ncbi.nlm.nih.gov]; The CDC shows increasing flu activity for the United States, and vaccination is recommended to prevent influenza virus infection. [Source: www.apphealthcare.com] 2019-01-28: The 2009 pandemic saw a significant increase in influenza-like illness in the United States. [Source: www.ncbi.nlm.nih.gov]; By the close of week 40 in the season, influenza A strains encompassed 73.4% of influenza cases. [Source: www.ncbi.nlm.nih.gov]; In 2007, FDA approved the first U.S. vaccine for people against an avian influenza A(H5N1) virus. [Source: www.cdc.gov]; Human infection with a novel influenza virus was added to the Nationally Notifiable Diseases Surveillance System in 2007. [Source: www.cdc.gov] 2019-01-21: Influenza A(H1N1) virus illness is a significant concern among pregnant women in the United States. [Source: www.ncbi.nlm.nih.gov] 2019-01-14: The 1918 H1N1 pandemic was not tracked by any state or federal public health agency prior to the pandemic. [Source: www.sciencedirect.com]; Influenza A virus infection induces muscle wasting via IL-6 and inflammatory endocrine responses. [Source: pubmed.ncbi.nlm.nih.gov]; Reducing the number of influenza A virus infected pigs at weaning is critical to minimize IAV spread to other farms. [Source: journals.plos.org];

System instruction

You are a reliable assistant for time-series forecasting.

BWithout contextual text

Forecast the next 36 values of the time series.
Domain: public health.

System instruction

You are a reliable assistant for time-series forecasting.

Weather & calendar context

Home energy

Forecast household use with weather and calendar context

You 1 signal

London household electricity · previous five days
London household electricity · previous five days0.030.070.120.16-239-191-143-96-480Electricity usage30-minute stepLondon household electricity · previous five days

Forecast the next 48 values of the time series.

Context:
This sequence records electricity usage at a household in London, United Kingdom, with a collection granularity of 30 minutes. The target date for prediction is Sunday, August 11, 2013. It is a weekend with light drizzle and gentle breeze. The minimum temperature is 15 degrees, and the maximum temperature is 21 degrees. The sun will rise at 5:40 and set at 20:31.

∿TimeBraid Time series

Next full day
Next full day0.030.060.10.13-23-95203448Electricity usage30-minute stepInputTimeBraidObserved future
InputTimeBraidObserved future
ComparisonAstra · Sol
InputTimeBraidObserved futureAstraSol
Next full day0.030.060.10.13-23-95203448Electricity usage30-minute stepInputTimeBraidObserved futureAstraSol

Market & economic reports

Energy · Time-MMD

Forecast fuel prices with supply context

Same history. Add the context.
One shared input 36 observations
Forecast fuel prices with supply context3.23.94.65.3-35-180USD / gallonWeekShared history
∿TimeBraid All 36 future values
Same history, with and without context33.84.55.3-7210192736USD / gallonWeekShared history · last observationsWith contextual textWithout contextual textObserved future
With contextual textWithout contextual textObserved future

AWith contextual text

Forecast the next 36 values of the time series.
Domain: energy.

Context information:
Available facts are as follows: 2022-06-20: The average gas price in Georgia was reported on June 22, 2022. [Source: gasprices.aaa.com]; The average gas price in Florida was reported on June 23, 2022. [Source: gasprices.aaa.com]; The price of gas has increased by almost $2 per gallon in America since Putin began amassing troops on the border of Ukraine. [Source: www.whitehouse.gov]; The United States produced 12 million barrels of oil per day in June 2022, the highest level under the current president. [Source: www.whitehouse.gov]; U.S. refining capacity has sunk to near decade low, according to the EIA. [Source: oilprice.com] 2022-06-13: The United States barely uses Russian oil, and the country is the world's largest oil producer. [Source: www.billtrack50.com]; The target population for EIA Weekly On-Highway Diesel Fuel Price is all active retail on-highway diesel fuel outlets in the contiguous United States. [Source: www.eia.gov]; The cost of global crude oil accounts for 61% of the average cost of gasoline in the United States. [Source: www.stout.com]; The United States is the world's largest oil producer, but all U.S. companies together only produce about 40% of the oil they refine. [Source: www.courthousenews.com] 2022-06-06: The price of crude oil has reached $120 per barrel, and global refining capacity has shrunk since COVID. [Source: oilprice.com]; The price of crude oil accounted for 56% of the price of gasoline in the decade through 2020. [Source: www.investopedia.com]; The national average price of gasoline has topped $5 per gallon, up 60 cents from a month ago. [Source: www.nytimes.com]; The spike in gas, oil, and diesel prices has saddled businesses with higher costs, forcing them to raise prices on their customers. [Source: www.washingtonpost.com] 2022-05-30: The US is a major crude oil exporter, and stopping exports could increase gas prices, not lower them. [Source: axpc.org]; The price of crude oil accounts for more than half of the price of gas at the pump. [Source: www.kiplinger.com]; As of June 1, 2022, the US average price of regular gasoline was $4.67. [Source: yaleclimateconnections.org]; The US residual fuel oil wholesale/resale price by refiners has been fluctuating over the years, with a recorded price of $0.603 in January 1983. [Source: www.eia.gov];

System instruction

You are a reliable assistant for time-series forecasting.

BWithout contextual text

Forecast the next 36 values of the time series.
Domain: energy.

System instruction

You are a reliable assistant for time-series forecasting.

Explore hypothetical events

Electricity

Anticipate demand around an event

Same history. Change the text.
One shared input 244 observations
Observed electricity demand2.6k3.7k4.8k6k-243-1220Electricity demand30-minute stepShared history
∿TimeBraid All 48 future values
One history, different conditional futures2.5k3.4k4.4k5.4k1204048Electricity demand30-minute stepShared history · last observationsDuring a sporting eventWithout the event
During a sporting eventWithout the event

ADuring a sporting event

Forecast the next 48 values.
Forecast interval: 2010-11-21 22:30:00 to 2010-11-22 22:00:00.
Background: The dataset captures regional power consumption across five Australian states, offering insights into overall energy usage patterns in Australia. Historically, the time series shows daily fluctuations with higher demand during the day and lower demand at night.

Scenario: Suppose a major sporting event occurs from 2010-11-22 08:00:00 to 2010-11-22 18:00:00, resulting in increased electricity consumption due to heightened activity in homes, businesses, and event venues. Electricity demand is elevated during the specified event hours, then returns toward its usual daily pattern after the event.

Constraint: The forecast values are assumed to be bounded above by the maximum historical value.

ComparisonAstra · Sol
InputTimeBraidAstraSol
Next 48 values2.4k3.6k4.8k6k-23-95203448Electricity demand30-minute stepInputTimeBraidAstraSol

BWithout the event

Forecast the next 48 values.
Forecast interval: 2010-11-21 22:30:00 to 2010-11-22 22:00:00.
No additional future scenario is provided.

ComparisonAstra · Sol
InputTimeBraidAstraSol
Next 48 values2.5k3.5k4.6k5.6k-23-95203448Electricity demand30-minute stepInputTimeBraidAstraSol

Compose a future trajectory

Controlled generation · Ctrl-F

Combine constraints for hospital activity

Same history. Change the text.
One shared input 22 observations
Observed recording17.726.936.145.3-21-110Metric valueStepShared history
∿TimeBraid All 12 future values
One history, different conditional futures22.129.737.344.9-7-314812Metric valueStepShared history · last observationsA sequence of eventsA bounded pulseA sustained higher level
A sequence of eventsA bounded pulseA sustained higher level

AA sequence of events

Use this continuation condition: The trajectory begins with a dip at future step 2, followed by a recovery to a local peak at future step 4. It then drops to a local minimum at future step 6 before surging to the highest future point at future step 7. The sequence concludes with a decline followed by a recovery at the end.

ComparisonAstra · Sol
InputTimeBraidAstraSol
Next 12 values17.726.936.145.3-21-14-8-1512Metric valueStepInputTimeBraidAstraSol

BA bounded pulse

Use this continuation condition: Create a pronounced upward pulse in future steps 3 to 8. Cap upper peaks in future steps 1 to 12.

ComparisonAstra · Sol
InputTimeBraidAstraSol
Next 12 values17.726.936.145.3-21-14-8-1512Metric valueStepInputTimeBraidAstraSol

CA sustained higher level

Use this continuation condition: Maintain a persistently higher level across future steps 1 to 12.

ComparisonAstra · Sol
InputTimeBraidAstraSol
Next 12 values17.627.236.846.4-21-14-8-1512Metric valueStepInputTimeBraidAstraSol

Read the abstract

We present TimeBraid, a series of unified time-series and language models that align pretrained language models and pretrained time-series foundation models through interleaved global residual attention layers. Each model inherits knowledge, instruction following, and reasoning from one side, continuous-signal perception and zero-shot forecasting from the other, and fuses the two in a shared representation space where both modalities are understood and generated. We study the design choices that make such unified modeling work: where to align the two representation spaces, how to ground language in temporal structure, how to balance understanding with generation, and how to keep joint optimization stable. The resulting recipe combines a unified prompting scheme for diverse time-series and text tasks, stabilized joint training, and supervision from 2.2M curated series–text pairs and 4.9M instruction-tuning samples. Across benchmarks spanning time-series perception, understanding, reasoning, and both context-aided and unimodal forecasting, TimeBraid remains competitive with far larger general-purpose models and task-specific counterparts.

Authors & affiliations

Xinyue Wang1,3† · Jiacheng Pang2 · Kun Zhou3 · Kexin Zhang2 · Defu Cao2 · Fan Feng1 · Faisal2 · Songyao Jin1,3† · Yan Liu2 · Biwei Huang1,3

  • 1 University of California San Diego
  • 2 University of Southern California
  • 3 Aether AI

† Part of the work was completed during an internship at Aether AI

Cite this work
@misc{wang2026timebraidunifyingtimeseries,
      title={TimeBraid: Unifying Time Series and Language for Understanding and Forecasting},
      author={Xinyue Wang and Jiacheng Pang and Kun Zhou and Kexin Zhang and Defu Cao and Fan Feng and Faisal and Songyao Jin and Yan Liu and Biwei Huang},
      year={2026},
      eprint={2609.29792},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.29792},
}