TSQA: Time Series Question Answering Overview

Updated 8 July 2026

TSQA is defined as natural-language question answering based on numerical time series data, supporting tasks like forecasting, imputation, and anomaly detection.
It decomposes the reasoning process into skills such as temporal scale selection, localization, and cross-interval integration for detailed temporal analysis.
Benchmark datasets like Time-MQA, QuAnTS, and MMTS-BENCH advance TSQA research by assessing model capabilities across diverse temporal tasks.

Searching arXiv for papers on TSQA and related benchmarks/methods. arxiv_search query: TSQA time series question answering benchmark reasoning agentic PATRA T3LLM QuAnTS ChatTime MMTS-BENCH Time-MQA TimeART ARFBench TS-Skill TDBench TSQA is an overloaded acronym spanning several question-answering traditions, but in recent multimodal and foundation-model research it most often denotes time series question answering: a setting in which a model answers natural-language questions grounded in one or more numerical time series rather than in ordinary text alone. In that formulation, the task is written either as $f:(\mathbf{X},\mathbf{C},Q)\rightarrow A$ , where $\mathbf{X}$ is a time series, $\mathbf{C}$ is contextual information, and $Q$ is a question, or as $A=f(Q,S)$ for a question $Q$ and time-series set $S$ (Kong et al., 26 Feb 2025, Lu et al., 26 Feb 2026). Across this literature, TSQA serves as a unified interface for forecasting, imputation, anomaly detection, classification, pattern interpretation, and reasoning over temporal evidence; the same acronym also appears in the unrelated senses Tabular Scenario-Based Question Answering and Time-Sensitive Question Answering (Li et al., 2021, Kim et al., 4 Aug 2025).

1. Definition, scope, and problem formulation

In the time-series literature, TSQA is defined as natural-language question answering grounded in temporal signals. The central claim is that a model must answer by reading values, patterns, and structures in one or more time series, not by retrieving symbolic facts already expressed as language. This distinguishes TSQA from text QA, where evidence is already symbolic, and from many vision-language settings, where semantics are often spatially organized in images or videos (Han et al., 23 May 2026, Divo et al., 7 Nov 2025).

The task is broad in answer form. In QuAnTS, answers may be binary, multiple-choice, or open-ended natural language; in Time-MQA, the answer may be a forecasted numeric sequence, an imputed sequence, a class label, an anomaly judgment, or a free-form reasoning response (Divo et al., 7 Nov 2025, Kong et al., 26 Feb 2025). T3LLM makes the same unification explicit by treating forecasting, imputation, classification, anomaly detection, and pattern-oriented questions such as trend, volatility, seasonality, and outlier understanding as instances of TSQA (Su et al., 27 Dec 2025).

The distinctive difficulty is that time-series evidence is numeric, dense, multivariate, temporally extended, and not inherently symbolic. TS-Skill emphasizes that answers may depend on the correct temporal granularity, the correct time interval, or evidence distributed across disjoint temporal regions (Han et al., 23 May 2026). PATRA further argues that time series carry intrinsic dynamics, especially trend and seasonality, and that treating them merely as text tokens or as generic multimodal patches is insufficient for deep alignment and reasoning (Lu et al., 26 Feb 2026).

2. Reasoning structure and task taxonomies

A recurrent theme in TSQA research is that benchmark labels such as “forecasting” or “anomaly detection” do not fully specify the underlying reasoning operations. TS-Skill therefore decomposes TSQA into three composable analytical skills: temporal scale selection (SK1), temporal localization (SK2), and cross-interval integration (SK3). Its argument is that aggregate benchmark scores can obscure which signal-level capabilities a model actually possesses (Han et al., 23 May 2026).

MMTS-BENCH proposes a broader capability hierarchy. It organizes time-series understanding into structural awareness, feature analysis, temporal reasoning, sequence matching, and cross-modal understanding, and reports 286 fine-grained composite subtasks. In that framework, higher-level reasoning is treated as a composition of lower-level capabilities such as local/global structural awareness and feature perception (Yin et al., 9 Feb 2026).

Application-specific benchmarks further refine this decomposition. ARFBench, which targets software incident response, organizes 8 categories into 3 tiers: Tier I covers anomaly presence; Tier II covers characterization of a single series through identification, start time, end time, magnitude, and categorization; Tier III covers paired-series reasoning through correlation and leading/lagging indicator questions (Xie et al., 23 Apr 2026). This tiering makes explicit that multi-series temporal reasoning depends on earlier abilities such as anomaly recognition and temporal localization.

These taxonomies collectively define TSQA as more than “read a chart and answer a fact.” The core operations include detecting temporal structure, selecting relevant intervals, integrating dispersed evidence, and mapping signal-level findings into natural-language answers. PATRA’s emphasis on trend and seasonality, and T3LLM’s emphasis on verifiable stepwise reasoning over numeric sequences, fit this same view of TSQA as layered temporal analysis rather than generic language inference (Lu et al., 26 Feb 2026, Su et al., 27 Dec 2025).

3. Benchmark and dataset ecosystem

The recent TSQA ecosystem is defined by a rapid expansion of benchmarks, each emphasizing a different slice of the problem: large-scale multi-task supervision, controlled skill diagnosis, multimodal reasoning, domain-specific realism, or continual temporal updating.

Resource	Main scope	Reported scale
Time-MQA TSQA dataset	Five task types across twelve domains	192,843 QA pairs
QuAnTS	Human-motion TSQA over skeleton trajectories	large-scale benchmark
MMTS-BENCH	Hierarchical multimodal TSQA benchmark	2,424 QA pairs
ARFBench	Incident-response anomaly reasoning	750 questions
CLTSQA-Data	Continual Time-Sensitive QA	50,000 questions, 5,000 contexts

Time-MQA positions TSQA as a multi-task corpus spanning forecasting, imputation, anomaly detection, classification, and open-ended reasoning QA across twelve domains, including healthcare, finance, energy, traffic, environment, IoT, nature, transport, human activities, machine sensors, AIOps, and the web. It reports 192,843 question-answer pairs, with per-task counts of 42,557 forecasting, 38,657 imputation, 37,000 anomaly detection, 37,000 classification, and 37,629 open-ended reasoning instances (Kong et al., 26 Feb 2025).

QuAnTS makes TSQA concrete in the human-motion setting. Its time series are synthetic tracked skeleton trajectories with 24 joints, 3 spatial coefficients each, over 16 seconds at 20 frames per second, giving sequences of shape $320\times24\times3$ , often flattened conceptually to $320\times72$ . The benchmark is designed around compositional reasoning over multiple consecutive actions rather than single-event recognition (Divo et al., 7 Nov 2025).

MMTS-BENCH contributes a hierarchical evaluation framework with 2,424 TSQA pairs across Base, InWild, Match, and Align. Base targets structural awareness and feature analysis in synthetic controlled settings; InWild targets feature analysis plus temporal reasoning on real-world data; Match evaluates sequence matching; Align evaluates bidirectional cross-modal alignment between time series and text (Yin et al., 9 Feb 2026).

TS-Skill is not organized primarily by downstream task type, but by skill diagnosis. It introduces timestamp-aware questions, broad domain coverage, and human-validated QA quality, and constructs the benchmark through SKEvol, a skill-guided agentic framework combining domain-aware seed generation, skill-controlled question generation, metadata- and code-assisted answer construction, multi-phase signal-grounded verification, and human-in-the-loop curation (Han et al., 23 May 2026).

ARFBench grounds TSQA in production telemetry. It contains 750 questions across 142 time series and 5.38M data points from 63 production incidents at Datadog, and evaluates anomaly reasoning over metrics such as latency, CPU, error counts, queue depth, replica counts, and Kafka lag (Xie et al., 23 Apr 2026).

Beyond time-series-specific corpora, temporal QA research supplies adjacent benchmark lineages. CLTSQA-Data extends Time-Sensitive Question Answering into a continual-learning setting with 50,000 questions and 5,000 contexts partitioned into five temporal subsets from 190–1939 through 2010–now (Yang et al., 2024). TDBench, also in the time-sensitive lineage, systematically generates factual TSQA pairs from temporal databases using temporal functional dependencies, temporal SQL, and the 13 Allen interval relations (Kim et al., 4 Aug 2025).

4. Modeling strategies

Early and current TSQA methods differ primarily in how they represent time series and how they couple signal perception with language reasoning. One strategy is textualization with contextual enhancement. Time-MQA formats raw sequences and auxiliary text into instruction-style prompts and fine-tunes open LLM backbones such as Mistral 7B, Llama-3 8B, and Qwen-2.5 7B with LoRA, treating context $\mathbf{C}$ as background information, feature descriptions, domain-specific knowledge, and task descriptions (Kong et al., 26 Feb 2025).

A second strategy is LLM adaptation through specialized tokenization. ChatTime models time series as a “foreign language”: it normalizes values using the history window, discretizes them into 10K bins, wraps each quantized value as a single vocabulary item such as ###0.2835###, expands the tokenizer vocabulary, and uses a decoder-only LLaMA-2-7B backbone for multimodal generation. This allows mappings from time series to time series, text plus time series to time series, and time series to text (Wang et al., 2024).

A third strategy is time-series-specific alignment. PATRA criticizes two common baselines—serializing numeric values as text tokens and simply projecting temporal patches for multimodal fusion—and introduces a pattern-aware mechanism that extracts trend and seasonality patterns for deep alignment. It then adds a task-aware balanced reward intended to harmonize learning across tasks of different difficulty while incentivizing coherent chains of thought (Lu et al., 26 Feb 2026).

A fourth strategy is review-and-correct reasoning supervision. T3LLM uses three LLM roles: a worker that generates step-wise chains of thought, a reviewer that checks reasoning against the original sequence and the gold answer, identifies the first erroneous or unsupported step, inserts a reflection, and truncates the rest, and a student that is fine-tuned on the corrected traces. Its training objective is the sum of an autoregressive chain-of-thought loss and an answer loss, $\mathbf{X}$ 0 (Su et al., 27 Dec 2025).

A fifth strategy is agentic tool augmentation. TimeART casts TSQA as a ReAct-style trajectory over Query, Thought, Action, Observation, and Final Answer, with the reasoning process formalized as $\mathbf{X}$ 1. Its action space consists of 21 out-of-the-box analytical tools, including numerical operators, pattern detectors, correlation analyzers, forecasting through LightGTS, and anomaly detection through DADA (Wu et al., 20 Jan 2026).