TSAQA/TSAQA-Benchmark

TSAQA: Time Series Analysis Question And Answering Benchmark

Figure 1: Domain/task distribution and task illustrations in TSAQA.

---

Overview

TSAQA is a large-scale, unified benchmark for evaluating the temporal analytical capabilities of language models on time series data. It addresses the fragmented landscape of existing time series QA benchmarks by consolidating 6 diverse analytical tasks under a single standardized evaluation framework.

The benchmark spans 210,000 samples across 13 domains, employs 3 question types (True-or-False, Multiple-Choice, and the novel Puzzling format), and provides training, validation, and test splits for controlled model development and evaluation.

Key properties:

• Scale: 210k samples across 13 real-world domains
• Breadth: 6 tasks covering both conventional and advanced temporal analysis
• Evaluation: Closed-ended questions (TF, MC, PZ) enabling objective, reproducible scoring
• Extensibility: Training-ready splits for instruction tuning and model development

---

Tasks

TSAQA integrates six tasks grouped into two complementary categories:

Group	Task	Description	Question Types
Conventional Analysis	Anomaly Detection	Determine whether the input time series contains anomalies.	TF
	Classification	Classify the input time series into a semantic category.	MC
Advanced Analysis	Characterization	Infer fundamental properties such as trend, seasonality, and dispersion.	TF, MC
	Comparison	Analyze relative similarities and differences between two time series.	TF, MC
	Data Transformation	Identify the relationship between raw and transformed data (e.g., Fourier transform, first-order differencing).	TF, MC
	Temporal Relationship	Determine the chronological dependencies among time series patches.	TF, MC, PZ

Question Types

• True-or-False (TF): The model determines whether a claim about the input time series is true or false.
• Multiple-Choice (MC): The model selects the correct answer from four candidates.
• Puzzling (PZ): Given the first patch of a sequence and four shuffled successor patches, the model must reconstruct the correct chronological order. This novel format evaluates global causal reasoning and temporal fidelity.

---

Dataset

Access

The full dataset is hosted on Hugging Face:

https://huggingface.co/datasets/TSAQA/TSAQA-Benchmark

Loading the Data

Using the datasets library (recommended):

from datasets import load_dataset

dataset = load_dataset("TSAQA/TSAQA-Benchmark")
train_df = dataset["train"].to_pandas()
test_df  = dataset["test"].to_pandas()

Using pandas directly with Parquet:

import pandas as pd

train_df = pd.read_parquet("hf://datasets/TSAQA/TSAQA-Benchmark/train.parquet")
val_df   = pd.read_parquet("hf://datasets/TSAQA/TSAQA-Benchmark/val.parquet")
test_df  = pd.read_parquet("hf://datasets/TSAQA/TSAQA-Benchmark/test.parquet")

Data Splits

Each task is partitioned into 70% training / 10% validation / 20% test. The benchmark allocates 30k samples to each task except Temporal Relationship (60k, given the difficulty of the PZ format).

Question Type Distribution

Question Type	Count	Proportion
True-or-False (TF)	84,632	40.3%
Multiple-Choice (MC)	85,479	40.7%
Puzzling (PZ)	39,889	19.0%

Data Statistics

Time series lengths, description lengths, and question lengths all follow long-tail distributions. All samples are z-score normalized and have lengths drawn uniformly from [32, 256].

---

Data Sources

Core Datasets (Advanced Analysis Tasks)

Dataset	Total Data Points	Domain
AustralianElectricityDemand	1,153,584	Energy
BDG-2 Rat	4,728,288	Energy
GEF12	788,280	Energy
ExchangeRate	56,096	Finance
FRED-MD	76,612	Finance
BIDMC32HR	8,000,000	Healthcare
PigArtPressure	624,000	Healthcare
USBirths	7,275	Healthcare
Sunspot	73,924	Nature
Saugeenday	23,711	Nature
SubseasonalPrecip	9,760,426	Nature
HierarchicalSales	212,164	Sales
M5	58,327,370	Sales
PedestrianCounts	3,130,762	Transport
PEMS03	9,382,464	Transport
UberTLCHourly	1,129,444	Transport
WikiDaily100k	274,099,872	Web

Anomaly Detection Datasets

Dataset	# Samples	Domain
ECG	17,862	Healthcare
SMD	58,888	IT Operations
MGAB	376	Mathematical Biology
Genesis	274	Spacecraft Telemetry
GHL	768	Industrial Control
Occupancy	8,178	Environmental Sensing

Classification Datasets (UCR Archive Subset)

37 datasets selected with ≤4 classes and sequence length ≤400. See the paper (Appendix A.3) for the full list.

---

Results

Zero-Shot Evaluation

The table below reports accuracy (%) for each task and question type. Bold = best, underlined = second-best.

Model	A.D. TF	CLS MC	Char. TF	Char. MC	Comp. TF	Comp. MC	DT TF	DT MC	TR TF	TR MC	TR PZ	Overall
GPT-4.1	55.85	50.38	92.97	89.36	83.57	76.99	54.36	51.13	65.90	79.09	45.77	62.82
GPT-4o	54.32	47.20	88.15	84.15	78.61	69.07	60.66	53.24	62.25	75.58	45.61	60.73
Claude-3.5-Sonnet	51.27	41.23	74.39	78.45	66.59	74.14	65.79	57.07	82.05	82.15	54.56	61.19
Gemini-2.5-Flash	52.08	49.07	85.48	81.08	77.79	72.21	63.62	60.17	75.05	84.49	60.84	65.08
Qwen3-8B	50.60	50.52	77.35	66.87	71.04	63.21	52.43	34.46	65.22	67.14	21.93	51.04
LLaMA-3.1-8B	54.92	50.20	68.10	62.26	67.84	49.98	51.90	36.56	54.82	40.95	6.80	44.93
Ministral-8B	53.35	34.08	71.06	63.93	47.54	52.90	50.70	25.28	50.58	33.88	30.77	44.65
Qwen3-0.6B	50.40	35.83	62.00	48.78	58.03	37.51	49.03	23.62	51.99	37.33	13.38	39.06
LLaMA-3.2-1B	49.47	39.48	63.74	52.55	61.02	36.82	48.87	4.20	48.97	5.44	6.76	35.70
Gemma3-1B	49.15	49.83	63.74	47.71	61.19	43.37	49.37	24.88	49.42	25.84	23.97	43.03

After Instruction Tuning (LoRA, Open-Source Models)

Model	A.D. TF	CLS MC	Char. TF	Char. MC	Comp. TF	Comp. MC	DT TF	DT MC	TR TF	TR MC	TR PZ	Overall
LLaMA-3.1-8B	91.02	91.27	92.44	83.68	86.72	79.31	90.17	86.62	96.94	97.41	67.68	85.26
Qwen3-8B	87.70	90.05	92.37	85.42	86.55	79.08	89.84	84.99	96.84	97.56	66.21	84.29
Ministral-8B	71.56	74.28	91.31	80.78	84.14	74.63	75.15	71.61	94.07	94.15	56.82	74.74
Qwen3-0.6B	83.68	85.78	89.38	74.87	80.65	64.84	80.51	73.28	93.92	93.79	63.34	78.32
LLaMA-3.2-1B	83.08	83.83	87.71	74.37	78.61	60.88	68.09	51.67	91.39	88.81	57.53	73.48
Gemma3-1B	83.10	84.05	87.88	72.54	78.61	59.31	64.06	45.23	91.00	88.05	42.92	69.70

A.D. = Anomaly Detection; CLS = Classification; Char. = Characterization; Comp. = Comparison; DT = Data Transformation; TR = Temporal Relationship. SFT = supervised fine-tuning with LoRA (rank 16, lr 1e-5, cosine schedule, 2 epochs on a single A100).

Key findings:

1. Commercial LLMs consistently outperform open-source models in zero-shot settings. Gemini-2.5-Flash achieves the highest overall score (65.08%).
2. Instruction tuning markedly improves all open-source models; LLaMA-3.1-8B (SFT) achieves 85.26%, surpassing all commercial models evaluated.
3. Puzzling (PZ) questions remain the most challenging format across all settings; the best PZ score is 67.68%.
4. Performance on Fourier Transform and Wavelet Transform questions is substantially lower than on First-Order Differencing, reflecting the difficulty of global frequency reasoning.

---

Benchmark Construction

Samples for the advanced analysis tasks (Characterization, Comparison, Data Transformation, Temporal Relationship) are drawn from the Core Datasets using Hierarchical Uniform Sampling to ensure balanced coverage across domains, datasets, and individual sequences.

• Characterization & Comparison questions are generated via a multi-LLM pipeline (GPT-4o for question synthesis, GPT-4.1 + Gemini-2.5-Flash + Claude-3.5-Sonnet for consensus labeling with weighted majority voting).
• Data Transformation uses NumPy/SciPy to compute exact Fourier, wavelet, and first-order difference transforms; distractors are transformations of unrelated sequences.
• Temporal Relationship uses template-based generation; PZ questions shuffle four consecutive patches of a sequence.
• Anomaly Detection downsamples normal segments to achieve a 1:1 class balance.
• Classification converts original numeric class labels to descriptive textual choices.

Human evaluation on 600 QA pairs (300 characterization, 300 comparison) by six Ph.D.-level experts yields 91.2% agreement for characterization and 87.4% for comparison.

---

Human Evaluation

Figure 2: Distribution of human annotator disagreement categories for characterization (single-series) and comparison (multi-series) tasks.

---

Citation

If you find TSAQA useful in your research, please cite:

@misc{jing2026tsaqatimeseriesanalysis,
      title={TSAQA: Time Series Analysis Question And Answering Benchmark}, 
      author={Baoyu Jing and Sanhorn Chen and Lecheng Zheng and Boyu Liu and Zihao Li and Jiaru Zou and Tianxin Wei and Zhining Liu and Zhichen Zeng and Ruizhong Qiu and Xiao Lin and Yuchen Yan and Dongqi Fu and Jingchao Ni and Jingrui He and Hanghang Tong},
      year={2026},
      eprint={2601.23204},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.23204}, 
}

---

License

This project is licensed under the MIT License.

All datasets used in TSAQA are publicly available. The benchmark was constructed following ethical guidelines to minimize biases and ensure data quality. See the paper for full attribution of each underlying dataset.