README.md

metadata

license: mit
task_categories:
  - question-answering
  - time-series-forecasting
language:
  - en
tags:
  - Time Series
  - Time Series QA
  - Time Series Analysis
  - Time Series Reasoning
  - Time Series Question Answering
  - Unified Time Series QA
  - TSQA
size_categories:
  - 100K<n<1M

Time Series Analysis Question Answering Benchmark (TSAQA)

View our paper at: https://arxiv.org/abs/2601.23204

Introduction

TSAQA is a novel unified benchmark designed to broaden task coverage and evaluate diverse temporal analysis capabilities. TSAQA integrates 6 diverse tasks under a single framework ranging from Conventional Analysis, including anomaly detection and classification, to Advanced Analysis, such as characterization, comparison, data transformation, and temporal relationship analysis. Spanning 210k samples across 13 domains, the dataset employs diverse formats, including true-or-false (TF), multiple-choice (MC), and a novel puzzling (PZ), to comprehensively assess time series analysis.

This benchmark allows development of Large Language Models (LLMs) and Time Series Foundation Models (TSFM) specifically for time series analysis and time series reasoning.

Figure: Data distribution and tasks of TSAQA.

🧩 Tasks of TSAQA. TF, MC, and PZ denote true-or-false, multiple-choice, and puzzling.

Group	Task	Description	Question Type
Conventional Tasks	Anomaly Detection	Determine whether the input contains anomalies.	TF
	Classification	Classify the input time series.	MC
Advanced Tasks	Characterization	Determine the characteristics of the time series.	TF & MC
	Comparison	Compare the characteristics of two time series.	TF & MC
	Data Transformation	Identify the relationship between raw and transformed data.	TF & MC
	Temporal Relationship	Determine the temporal relationship of patches.	TF & MC & PZ

Data Statistics of TSAQA

🧠 Task Groups in TSAQA

TSAQA benchmark encompasses two groups of tasks with six diverse tasks designed to evaluate a model's ability of understanding the fundamental properties of time series data. The TSAQA benchmark includes two major groups of tasks designed to evaluate different reasoning abilities in time series analysis.

🔹 Conventional Analysis Tasks

These are classic tasks widely explored in traditional time series analysis:

1. 🩸 Anomaly Detection – Identify irregular or unexpected patterns in a time series.
2. 🏷️ Classification – Reason about the relationship between a time series and its underlying conceptual category.

🔸 Advanced Analysis Tasks

These novel analytical tasks focus on deeper, intrinsic properties of time series:

3. 📊 Characterization – Infer fundamental properties such as trend, seasonality, and dispersion.
4. ⚖️ Comparison – Reason about relative similarities and differences between two time series.
5. 🔄 Data Transformation – Understand relationships between original and transformed time series (e.g., via Fourier transform).
6. ⏱️ Temporal Relationship – Capture chronological dependencies among time series patches.

---

🧩 Insight:
These advanced analysis tasks push the boundaries of conventional time series modeling—encouraging the development of models that can grasp cognitive concepts of time series and reason over human-posed questions.

📊 Data Collection

In this section, we detail the data sources, including core datasets, anomaly detection datasets, and classification datasets.

---

🧩 Core Datasets

We extract data from multiple time-series datasets, including:

Australian Electricity Demand — Half-hourly electricity demand for Victoria, Australia (2014).
BDG-2 Rat — Building-level electricity data from the Building Data Genome Project 2.
GEF12 — Load forecasting benchmark from the Global Energy Forecasting Competition 2012.
ExchangeRate — Daily exchange rates for currencies of eight countries (1990–2016).
FRED-MD — Monthly macroeconomic indicators from the Federal Reserve Bank.
BIDMC32HR — ICU PPG and ECG recordings from 53 adult patients.
PigArtPressure — Vital sign data from 52 pigs pre/post induced injury.
US Births — Daily number of U.S. births (1969–1988).
Sunspot — Daily sunspot numbers from 1818–2020.
Saugeen — Daily mean river flow data for the Saugeen River (1915–1979).
Subseasonal Precipitation — Daily precipitation (1948–1978).
Hierarchical Sales — SKU-level daily pasta brand sales (2014–2018).
M5 — Walmart hierarchical sales forecasting dataset.
Pedestrian Counts — Hourly pedestrian counts from 66 sensors in Melbourne (2009–2020).
PEMS03 — Traffic flow data collected by Caltrans PeMS.
Uber TLC Daily — Uber pickup counts in NYC (Jan–Jun 2015).
WikiDaily100k — Daily traffic data for 100,000 Wikipedia pages.

📈 Summary of Core Datasets

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

We extract data from multiple anomaly detection benchmarks, including:

• MGAB – Mackey–Glass time series exhibiting chaotic behavior and synthetic anomalies.
• ECG – Electrocardiogram recordings with anomalies corresponding to ventricular premature contractions.
• Genesis – Spacecraft telemetry data from a pick-and-place demonstrator.
• GHL – Gasoil Heating Loop data with simulated cyber-attacks.
• Occupancy – Room occupancy monitoring using temperature, humidity, light, and CO₂ data.
• SMD – Server Machine Dataset from a large Internet company, labeled for anomaly detection.

🧾 Summary of Anomaly Detection Datasets

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

---

🧠 Classification Datasets

We extract data from the UCR Archive using the following criteria:

• Datasets with ≤4 classes
• Time series length ≤400 time points

A total of 37 benchmarks were selected, spanning domains such as robotics, energy, healthcare, synthetic, manufacturing, nature, and transport.

🗂️ Summary of Classification Datasets

Name	# Samples	# Classes	Domain
SonyAIBORobotSurface1	486	2	Robotics
SonyAIBORobotSurface2	771	2	Robotics
FreezerRegularTrain	2,404	2	Energy
FreezerSmallTrain	2,353	2	Energy
ToeSegmentation1	210	2	Healthcare
ToeSegmentation2	129	2	Healthcare
TwoPatterns	3,999	4	Synthetic
CBF	757	3	Synthetic
Wafer	5,744	2	Manufacturing
ECG200	159	2	Healthcare
TwoLeadECG	923	2	Healthcare
ECGFiveDays	704	2	Healthcare
DistalPhalanxOutlineCorrect	690	2	Healthcare
MiddlePhalanxOutlineCorrect	731	2	Healthcare
ProximalPhalanxOutlineCorrect	688	2	Healthcare
DistalPhalanxOutlineAgeGroup	423	3	Healthcare
MiddlePhalanxOutlineAgeGroup	435	3	Healthcare
ProximalPhalanxOutlineAgeGroup	485	3	Healthcare
PhalangesOutlinesCorrect	2,076	2	Healthcare
MoteStrain	1,012	2	Nature
GunPointMaleVersusFemale	362	2	Healthcare
GunPointOldVersusYoung	356	2	Healthcare
GunPointAgeSpan	368	2	Healthcare
GunPoint	169	2	Healthcare
Strawberry	786	2	Nature
ItalyPowerDemand	890	2	Energy
Chinatown	293	2	Transport
BME	137	3	Synthetic
PowerCons	294	2	Energy
DodgersLoopWeekend	111	2	Transport
DodgersLoopGame	115	2	Transport
DiatomSizeReduction	248	4	Nature
SmoothSubspace	236	3	Synthetic
UMD	148	3	Synthetic
Wine	85	2	Nature
Coffee	48	2	Nature
ArrowHead	175	3	Nature