NL2TS-675: Natural Language to Time Series Dataset

Paper: Zero-to-Forecast: Natural Language to Time Series Prediction via Cross-Modal Ensembles

📊 Dataset Overview

NL2TS-675 is a comprehensive benchmark dataset for natural language to time series prediction, containing 675 description–series pairs across diverse domains and temporal patterns.

🎯 Dataset Statistics

• Total Pairs: 675
• Domains: 6 (finance, healthcare, weather, IoT, retail, technology)
• Temporal Horizons: 3 (12, 24, 48 time steps)
• Pattern Types: 5 (trend, seasonal, spike, plateau, irregular)
• Data Split: 70% train (472), 15% dev (101), 15% test (102)

🏢 Domain Coverage

Domain	Train	Dev	Test	Value Range	Description
Finance	121	27	32	30-360	Stock prices, revenue, trading volume
Healthcare	70	9	11	45-325	Vital signs, patient metrics
Weather	53	20	17	5-65	Temperature, humidity, atmospheric data
IoT	66	13	11	20-160	Sensor readings, device metrics
Technology	66	11	13	15-220	System performance, network metrics
Retail	96	21	18	35-270	Sales data, inventory levels

📈 Temporal Patterns

Pattern Types

1. Trend: Monotonic increase/decrease patterns
2. Seasonal: Cyclical and periodic variations
3. Spike: Sudden peaks and impulse responses
4. Plateau: Stable periods with minimal variation
5. Irregular: Complex, non-linear patterns

Temporal Horizons

• Short-term: 12 time steps
• Medium-term: 24 time steps
• Long-term: 48 time steps

📁 Data Format

Each item in the dataset follows this JSON structure:

{
  "uid": "domain_pattern_freq_length_xxx",
  "domain": "finance",
  "text": "Stock price shows gradual increase over time",
  "series": [120.5, 125.2, 128.7, 132.1, 135.6, 138.9, 142.3, 145.1, 148.4, 151.2, 154.8, 157.3],
  "freq": "hour",
  "length": 12,
  "split": "train"
}

Field Descriptions

• uid: Unique identifier with domain, pattern, frequency, length, and sequence number
• domain: One of 6 domains (finance, healthcare, weather, iot, technology, retail)
• text: Natural language description of the time series pattern
• series: Numerical time series as a list of floats
• freq: Temporal frequency (hour, day, week, month)
• length: Number of time steps (12, 24, or 48)
• split: Data split (train, dev, test)

🔬 Quality Controls

Generation Protocol

• Text Descriptions: Human-authored using procedural templates
• Time Series: Deterministic procedural algorithms with domain-specific constraints
• Leakage Prevention: Strict train/dev/test splits with no overlap
• Domain Balance: Representative distribution across all domains

Validation

• Domain-appropriate value ranges
• Realistic pattern generation
• No exact duplicates across splits
• Balanced representation across patterns and lengths

📊 Example Items

Finance Domain

{
  "uid": "finance_trend_up_gradual_12_001",
  "domain": "finance",
  "text": "Stock price shows gradual increase over time",
  "series": [120.5, 125.2, 128.7, 132.1, 135.6, 138.9, 142.3, 145.1, 148.4, 151.2, 154.8, 157.3],
  "freq": "hour",
  "length": 12,
  "split": "train"
}

Healthcare Domain

{
  "uid": "healthcare_spike_recovery_24_001",
  "domain": "healthcare",
  "text": "Heart rate increases during exercise then recovers",
  "series": [72.1, 75.3, 78.9, 82.4, 85.7, 88.2, 90.5, 87.3, 84.1, 80.8, 77.5, 74.2, 71.8, 70.2, 69.1, 68.5, 67.9, 67.2, 66.8, 66.5, 66.1, 65.8, 65.5, 65.2],
  "freq": "minute",
  "length": 24,
  "split": "train"
}

Weather Domain

{
  "uid": "weather_seasonal_daily_48_001",
  "domain": "weather",
  "text": "Temperature shows seasonal variation",
  "series": [15.2, 18.7, 22.1, 25.8, 28.4, 30.1, 29.8, 27.3, 24.1, 20.5, 17.2, 14.8, 12.5, 11.2, 10.8, 11.5, 13.2, 15.8, 18.9, 22.4, 25.9, 28.7, 30.5, 29.9, 27.1, 23.8, 20.1, 16.7, 14.2, 12.8, 12.1, 12.9, 14.7, 17.3, 20.8, 24.2, 27.6, 29.8, 30.9, 30.2, 28.1, 25.3, 21.9, 18.4, 15.6, 13.5, 12.7, 13.4],
  "freq": "day",
  "length": 48,
  "split": "train"
}

🚀 Usage

Loading the Dataset

import json

def load_nl2ts_675(filepath="nl2ts_675.jsonl"):
    """Load NL2TS-675 dataset."""
    items = []
    with open(filepath, 'r') as f:
        for line in f:
            items.append(json.loads(line.strip()))
    return items

# Load dataset
dataset = load_nl2ts_675()

# Filter by split
train_items = [item for item in dataset if item['split'] == 'train']
dev_items = [item for item in dataset if item['split'] == 'dev']
test_items = [item for item in dataset if item['split'] == 'test']

# Filter by domain
finance_items = [item for item in dataset if item['domain'] == 'finance']

# Filter by pattern length
short_series = [item for item in dataset if item['length'] == 12]

Basic Statistics

# Domain distribution
from collections import Counter
domains = Counter([item['domain'] for item in dataset])
print("Domain distribution:", dict(domains))

# Length distribution
lengths = Counter([item['length'] for item in dataset])
print("Length distribution:", dict(lengths))

# Split distribution
splits = Counter([item['split'] for item in dataset])
print("Split distribution:", dict(splits))

📈 Evaluation Metrics

The dataset is designed for evaluation using multiple metrics:

• MAE: Mean Absolute Error
• MSE: Mean Squared Error
• Pearson r: Linear correlation coefficient
• Spearman ρ: Rank correlation coefficient
• DTW: Dynamic Time Warping distance
• Trend F1: F1 score for trend direction prediction

📄 Citation

If you use this dataset, please cite:

@article{zero-to-forecast-2024,
  title={Zero-to-Forecast: Natural Language to Time Series Prediction via Cross-Modal Ensembles},
  author={Zero-to-Forecast Team},
  journal={NeurIPS BERT2S Workshop},
  year={2024}
}

🔗 Related Resources

• Main Repository: https://github.com/gokulsrinaths/BERT2S
• Paper: Zero-to-Forecast: Natural Language to Time Series Prediction via Cross-Modal Ensembles
• Interactive Demo: Available in the main repository

📞 Contact

For questions about this dataset, please open an issue on the main repository: https://github.com/gokulsrinaths/BERT2S

---

Dataset License: MIT License
Paper: NeurIPS BERT2S Workshop 2024