Source-linked AI summary
Time-MQA: Time Series Multi-Task Question Answering with Context Enhancement
Yaxuan Kong, Yiyuan Yang, Yoontae Hwang, Wenjie Du, Stefan Zohren, Zhangyang Wang, Ming Jin, Qingsong Wen
TL;DR
Time-series methods and datasets have largely centered on limited numerical tasks, while paired language-and-time-series resources for dialogue and reasoning remain scarce. The paper introduces Time-MQA and TSQA, a multi-task natural-language framework and approximately 200k-pair dataset, and reports improved temporal interpretation after continual pre-training. The authors also identify coverage, long-sequence, hallucination, and evaluation limitations.
Problem
Existing time-series methods and datasets focus narrowly on tasks such as forecasting or anomaly detection, while paired language-and-time-series datasets for dialogue and reasoning are lacking.
Method
Time-MQA unifies forecasting, imputation, anomaly detection, classification, and open-ended reasoning through natural-language questions over time series and context, supported by the TSQA dataset.
Results
Continual pre-training Mistral 7B, Llama 8B, and Qwen-2.5 7B on TSQA enhanced temporal-pattern interpretation beyond numeric tasks and produced contextually rich insights.
Takeaways & Limitations
TSQA provides a resource for advancing time-series question answering and reasoning in LLMs while making temporal information more accessible and context-aware.
Takeaways & Limitations
TSQA may omit highly irregular or domain-specific series, financial data are limited to earnings-call transcripts, long numerical sequences may cause hallucinations, and open-ended reasoning evaluation remains challenging.
Abstract
from arXiv · showhide
Time series data are foundational in finance, healthcare, and energy domains. However, most existing methods and datasets remain focused on a narrow spectrum of tasks, such as forecasting or anomaly detection. To bridge this gap, we introduce Time Series Multi-Task Question Answering (Time-MQA), a unified framework that enables natural language queries across multiple time series tasks - numerical analytical tasks and open-ended question answering with reasoning. Central to Time-MQA is the TSQA dataset, a large-scale dataset containing $\sim$200k question-answer pairs derived from diverse time series spanning environment, traffic, etc. This comprehensive resource covers various time series lengths and promotes robust model development. We further demonstrate how continually pre-training large language models (Mistral 7B, Llama-3 8B, and Qwen-2.5 7B) on the TSQA dataset enhanced time series reasoning capabilities, moving beyond mere numeric tasks and enabling more advanced and intuitive interactions with temporal data. The complete TSQA dataset, models, user study questionnaires for evaluation, and other related materials have been open-sourced.
1 Introduction
Existing time-series research and datasets largely emphasize single numerical tasks, leaving broader natural-language reasoning underdeveloped. Time-MQA addresses this gap with a unified framework and the approximately 200k-pair TSQA dataset, while continual pre-training improves time-series knowledge and reasoning in representative LLMs.
- Most pre-LLM time-series research focused on numerical tasks such as forecasting and anomaly detection, typically under single-objective methods.
- Existing context-enhanced approaches commonly target one task, leaving broader multitask reasoning and inference capabilities insufficiently addressed.
- Time-MQA unifies classical time-series tasks and open-ended questions, allowing users to query analyses through natural language.
- TSQA contains approximately 200k question-answer pairs across diverse domains, time-series lengths, tasks, and open-ended reasoning questions.
- Continual pre-training on TSQA enables Mistral 7B, Llama-3 8B, and Qwen-2.5 7B to acquire time-series knowledge and reasoning beyond basic numeric handling.
2 Related Works
Classical time-series work spans forecasting, imputation, classification, generation, augmentation, and decomposition, while newer methods incorporate textual information and multimodal integration. Question answering adds natural-language reasoning but retains challenges in ambiguity, accuracy, and efficiency.
- Classical time-series analysis addresses temporal-pattern challenges across applications including finance, healthcare, climate, electricity, AIOps, and industrial maintenance.
- Forecasting predicts future time points from historical values and features, with short-term and long-term horizons capturing different temporal dependencies.
- Imputation repairs missing or corrupted sequence data to support downstream analysis, using methods ranging from interpolation to deep generative models.
- Time-series research also includes classification, regression, generation, augmentation, and decomposition for real-world applications.
- Text-enhanced methods use descriptions, metadata, reports, alignment, attention, or contrastive learning to combine textual and time-series information.
- Natural-language question answering seeks accurate generated responses but must handle ambiguous queries, answer reliability, processing efficiency, and complex reasoning.
3 Methodology
Time-MQA unifies classical time-series objectives and open-ended reasoning under natural-language queries, using contextual information and shared representations. TSQA supplies approximately 200k question-answer pairs across twelve domains and five task types, including diverse reasoning formats.
- The Time-MQA Framework: Time-MQA maps a time series, auxiliary context, and a natural-language question to answers such as predictions, labels, anomalous timestamps, or explanations.Its model uses continued pre-training of Mistral 7B, Llama-3 8B, and Qwen-2.5 7B on TSQA.
- The Time-MQA Framework: The framework unifies forecasting, imputation, classification, anomaly detection, and open-ended queries instead of requiring separate task-specific pipelines.Shared representations support knowledge sharing across question types.
- The Time-MQA Framework: Time-MQA integrates auxiliary context to resolve ambiguities and improve robustness beyond methods relying solely on the time series.Context can include textual metadata, domain knowledge, or other modalities.
- Dataset Composition and Categorization: TSQA contains 192,843 question-answer pairs spanning twelve domains and five task types.The task types are forecasting, imputation, anomaly detection, classification, and open-ended reasoning.
- Dataset Composition and Categorization: Open-ended reasoning QA includes 6,919 true/false, 11,281 multiple-choice, and 12,510 open-ended questions.These questions cover trends, seasonality, cycles, summarization, volatility, anomalies, and structural breaks.
- Comparison with Existing Datasets: Compared with prior text-enhanced datasets, TSQA supports both classical numerical and cognitive tasks, covers twelve domains, and combines real and synthetic data.The paper reports approximately 200k instances versus approximately 10k–16k for prior datasets and a maximum of nine domains.
4 Experiment and Result
Experiments evaluate fine-tuned models across numerical and open-ended tasks, finding improved performance across multiple tasks and stronger reasoning or forecasting behavior in specific settings. User feedback further distinguishes model strengths in numerical accuracy, explanation clarity, and overall preference.
- Main Results: Fine-tuned models improved performance across multiple tasks, including open-ended reasoning, forecasting, and imputation.The evaluation sampled 50 QA pairs for each task type or question format.
- Main Results: 82% accuracy was achieved by fine-tuned Qwen on judgment questions, while fine-tuned Mistral reached 64% on multiple-choice questions.These are reported results for open-ended reasoning QA.
- Main Results: Fine-tuned Mistral outperformed GPT-4o on forecasting despite relatively high MSE values associated with long time series.The long sequences were described as challenging for models to handle.
- Forecasting with Long Time Series: Fine-tuned Mistral forecasts outperformed GPT-4o on a long time-series example and included rationales addressing complex patterns and seasonal trends.The model also answered follow-up questions about its reasoning in natural language.
- TSQA-Tuned vs. Zero-Shot Model: The TSQA-tuned Mistral model performed slightly better than the original zero-shot Mistral model on open-ended reasoning QA.The comparison covered judgment and multiple-choice questions.
- User Study: In the user study, Mistral received 70.5% of overall preference votes, while Qwen received 32.1%.Mistral and Qwen were identified as strongest for accuracy-driven tasks, with Mistral especially strong for thorough explanations.
5 Conclusion
Time-MQA unifies diverse time series analyses under natural-language question answering, while TSQA supplies broad training data for time series reasoning. Continual pre-training on TSQA improves models’ ability to interpret temporal patterns and generate contextually rich insights.
- Time-MQA integrates forecasting, imputation, anomaly detection, classification, and open-ended reasoning in one question-answering framework.
- TSQA contains approximately 200k question-answer pairs across twelve domains and varying time series lengths.
- Continual pre-training Mistral 7B, Llama 8B, and Qwen-2.5 7B on TSQA enhances temporal-pattern interpretation beyond numeric tasks.
- The framework makes temporal information more accessible and context-aware through natural-language interaction.
Limitations
The paper identifies scope, modeling, evaluation, and computational limitations. TSQA may not represent highly irregular or domain-specific series, while long numerical sequences, open-ended evaluation, and large-model fine-tuning remain challenging.
- TSQA may not cover all real-world scenarios, particularly highly irregular or domain-specific time series.
- The financial data currently consist only of earnings call transcripts, excluding other financial time series and textual sources.
- Pre-training LLM architectures may not be optimized for long numerical sequences, which may cause hallucination problems.
- Automated metrics may not fully capture open-ended reasoning quality, and high fine-tuning costs may limit some real-time applications.
Ethics Statement
The work uses publicly available datasets and language models, constructing TSQA from established benchmarks and synthetic data generated under ethical guidelines. These measures aim to minimize bias and support data quality.
- All datasets and language models used in the work are publicly available.
- TSQA was constructed from well-established, publicly accessible time series benchmarks.
- Synthetic data generation followed ethical guidelines to minimize biases and ensure data quality.
A Examples of TSQA Dataset
TSQA examples translate varied time series into natural-language tasks, including forecasting, imputation, classification, anomaly detection, and open-ended interpretation. The examples span EEG, influenza surveillance, web traffic, activity recognition, and cloud monitoring.
- An EEG example asks the model to predict the next nine time series points from cortical-potential recordings sampled at 256 Hz.
- The EEG forecasting answer provides nine predicted values ranging from -11.69 to -7.59.
- An influenza example combines weighted ILI rates, patient demographics, provider counts, and internet search results to predict a future rate.
- The influenza answer estimates a weighted ILI rate of 1.83453 for June 04, 2014 and characterizes the situation as moderate to severe.
- Other examples request missing-value imputation for Wikipedia traffic, activity classification from accelerometer data, and anomaly detection in AWS CloudWatch metrics.
A.5 Open-Ended Reasoning QA.
The open-ended reasoning QA examples cover trend, structural-break, and seasonality questions, with answers explaining patterns in the supplied sequences. Some pairs were generated by GPT-4o using real time-series signals.
- Open-Ended Reasoning QA: The dataset asks users to summarize overall movement, including an initial increase, a peak, and a later decline.The example answer describes the middle portion as a peak with higher fluctuations and a later slight decline.
- Open-Ended Reasoning QA: The structural-break example labels the sequence as having no structural break because it changes gradually and then flattens.The explanation distinguishes gradual variability from an abrupt change in level or direction.
- Open-Ended Reasoning QA: The seasonality example is labeled false because the fluctuations do not form a clear cyclical pattern.The answer characterizes the variations as more random than seasonally driven.
- Open-Ended Reasoning QA: GPT-4o partially generated open-ended questions across topics such as trends, seasonality, volatility, anomalies, and structural breaks.Figure 7 presents the prompt used for this generation process.
C High-Level Data Statistics for the TSQA dataset
TSQA is documented through high-level statistics and standardized question-answer formatting. The materials describe dataset sources and model-specific sequence delimiters.
- High-Level Data Statistics: Table 5 summarizes QA counts, task and domain distributions, real versus synthetic data, and data sources.The listed sources include UTSD, TimeMMD, WISDM, FOG, NAB, MIT-BIH Arrhythmia, UCR, MGAB, SensorScope, KPI, and Yahoo.
- Formatting: Question-answer pairs use the template <QUE> {Question} <ANS> {Answer} </END>.This template explicitly separates each question from its answer and marks the sample ending.
- Formatting: Continual pre-training adds model-specific boundary tokens for Llama, Mistral, and Qwen samples.Llama uses beginning and end tokens, Mistral uses <s> and </s>, and Qwen appends <|endoftext|>.
E Experiments on Model Reliance on Time Series Data
The reliance experiments alter or remove time-series inputs while preserving questions, testing whether models use numeric and temporal structure. Results indicate that altered inputs change responses, whereas removing data produces unreliable outputs.
- Experimental Design: The experiments targeted seasonality, abrupt changes, trends, and anomalies using fine-tuned Mistral 7B and Qwen-2.5 7B.The evaluation included 50 multiple-choice and 50 true-or-false questions, with N/A used for inconsistent or refused outputs.
- Experimental Design: Removing the time series caused incomplete or fabricated answers that could not be consistently scored.This setup supplied only the question text and omitted all numeric data.
- Results: Randomized series often produced accurate generic answers because they lacked meaningful seasonality or discernible trends.The passage cautions that higher true-or-false accuracy did not necessarily reflect deeper understanding.
- Results: Shuffling values removed evident trends or abrupt breaks, and models consequently adjusted their answers.The shuffled setup retained the same values but destroyed their temporal order.
- Conclusion: The findings suggest that models adapt responses when numeric context is altered or removed.The authors interpret the combined ablations as evidence that models reflect on time-series inputs.
F User Study
The user study collected 78 responses evaluating model outputs for accuracy, completeness, clarity, and preference. Results favored Mistral overall, while Qwen was preferred for trend analysis and Mistral for several numerical questions.
- Evaluation Criteria: The study questionnaire separately assessed answer accuracy, explanation clarity, logical consistency, and completeness.The listed questions ask whether answers are accurate, understandable, logically connected to the data, and complete.