TSFM Python Client

A Python client library for the TSFM (Time Series Foundation Model) Inference Platform. Supports both univariate and multivariate time series forecasting with confidence intervals.

Installation

pip install tsfm-client

Quick Start

import os
import numpy as np
from tsfm_client import TSFMClient

# Set your API key
os.environ['TSFM_API_KEY'] = 'your_api_key_here'

# Create client
client = TSFMClient(api_key=os.getenv('TSFM_API_KEY'))

# Make prediction with confidence intervals
data = np.array([10, 12, 13, 15, 17, 16, 18, 20, 22, 25])
response = client.predict(
    data=data,
    forecast_horizon=5,
    num_samples=100,
    confidence_intervals=[0.8, 0.95]
)

print(f"Forecast: {response.forecast}")
print(f"80% CI: {response.confidence_intervals['80%']}")
print(f"95% CI: {response.confidence_intervals['95%']}")

client.close()

Supported Models

chronos-t5-small (Amazon)

• Type: Univariate time series forecasting
• Framework: Chronos pipeline with T5 transformer architecture
• Max forecast horizon: 64 steps (recommended)
• Optimal use: Fast predictions for single time series
• Default confidence intervals: Uses 10 samples for CI calculation
• Strengths: Quick inference, good for short-term forecasting

toto-open-base-1.0 (Datadog)

• Type: Multivariate time series forecasting
• Framework: Zero-shot transformer model
• Max forecast horizon: 336 steps (recommended)
• Optimal use: Complex multivariate relationships, longer horizons
• Default confidence intervals: Uses 256 samples for CI calculation
• Strengths: Handles multiple correlated variables, robust uncertainty estimation

Supported Input Formats

The client accepts multiple data formats for maximum flexibility:

• Numpy arrays: np.array([1, 2, 3]) (most efficient)
• Python lists: [1, 2, 3] or [[1, 10], [2, 11]] (multivariate)
• Pandas Series: pd.Series([1, 2, 3])
• Pandas DataFrame: For multivariate data

Features

• ✅ Multiple confidence intervals: Get 80%, 90%, 95% intervals in single request
• ✅ Multivariate forecasting: Predict with multiple related time series
• ✅ Flexible sampling: Control uncertainty estimation with num_samples

Advanced Usage

Multivariate Prediction

# 2D numpy array: time steps × variables
multivariate_data = np.array([[20, 65], [21, 63], [22, 61], [19, 67]])
response = client.predict(
    model_name='toto-open-base-1.0',
    data=multivariate_data,
    forecast_horizon=10,
    confidence_intervals=[0.8, 0.9, 0.95],
    num_samples=100
)

Context Manager

with TSFMClient() as client:
    response = client.predict(data=np.array([1, 2, 3, 4, 5]))
    print(response.forecast)

Examples

For comprehensive examples including visualization and model comparison, see the demo notebook.

Requirements

• Python >= 3.10
• Valid TSFM API key
• Dependencies: numpy, pandas, httpx, pydantic

API Reference

TSFMClient.predict()

predict(
    model_name: str = "chronos-t5-small",
    data: Union[np.ndarray, pd.Series, List[float], List[List[float]]],
    forecast_horizon: int = 12,
    confidence_intervals: Optional[List[float]] = None,
    num_samples: Optional[int] = None,
    time_interval_seconds: Optional[int] = None
) -> PredictionResponse

Parameters:

• model_name: Model to use ('chronos-t5-small' or 'toto-open-base-1.0')
• data: Time series data (1D for univariate, 2D for multivariate)
• forecast_horizon: Number of steps to predict
• confidence_intervals: List of confidence levels (e.g., [0.8, 0.95])
• num_samples: Number of samples for uncertainty estimation
• time_interval_seconds: Time between data points in seconds

Returns:

• PredictionResponse with forecast, confidence intervals, and metadata

Roadmap

• 🔄 Batch processing: Process multiple time series in a single request
• 🎯 More models: Additional foundation models coming soon
• ⚙️ Fine-tuning: Support for domain-specific model adaptation

License

MIT License