EpochsFM-TF

A decoder-only foundation model for time series forecasting

Release Date: September 2025

Overview

EpochsFM-TF is a pretrained decoder-only foundation model designed specifically for time series forecasting. It delivers state-of-the-art performance on diverse forecasting tasks while maintaining computational efficiency.

Key Features

• 200M Parameters: Efficient architecture optimized for forecasting
• Patch-based Processing: Handles variable-length time series contexts
• Point and Quantile Forecasts: Provides both mean predictions and uncertainty estimates
• Decoder-only Architecture: Self-attention based stack for sequence modeling
• Multi-horizon Forecasting: Predicts multiple steps ahead

Installation

pip install torch transformers

Quick Start

import torch
from transformers import TimesFm2_5ModelForPrediction

model = TimesFm2_5ModelForPrediction.from_pretrained("comethrusws/epochsFM-tf")
model = model.to(torch.float32).eval()

# Example time series data
past_values = [
    torch.linspace(0, 1, 100),
    torch.sin(torch.linspace(0, 20, 67)),
]

with torch.no_grad():
    outputs = model(past_values=past_values, forecast_context_len=1024)

# Mean predictions
print(outputs.mean_predictions.shape)

# Full predictions (including quantiles)
print(outputs.full_predictions.shape)

Model Specifications

• Architecture: Decoder-only transformer
• Parameters: 200M
• Input: Patch-based time series contexts
• Output: Point forecasts and quantile predictions
• Context Length: Up to 1024 time steps

Use Cases

• Demand forecasting
• Financial time series prediction
• Energy consumption forecasting
• Traffic and resource planning
• Anomaly detection preprocessing

License

This project is available under a custom license.

• Non-commercial use: Free for personal projects, research, and educational purposes
• Commercial use: Requires explicit permission. Contact [email protected] for licensing inquiries

See LICENSE file for full terms.