Samay Dashboard

An interactive dashboard for time-series forecasting using state-of-the-art foundational models.

Overview

Samay Dashboard is a web-based tool that enables users to interact with and evaluate several foundational time-series forecasting models, including:

• Large Pre-trained Time-Series Model (LPTM) - Our primary model, as described in this research paper
• TimesFM - For benchmarking and comparison
• Chronos - For benchmarking and comparison
• TimeMOE - For benchmarking and comparison

The dashboard provides a comprehensive interface to:

• Upload custom datasets
• Load pre-trained forecasting models
• Fine-tune models on user data
• Generate forecasts and visualize results
• Explore summarized research findings and methodologies

Architecture

The system consists of two main components:

1. Frontend Dashboard: A user-friendly interface for interacting with the models
2. Backend API: A Flask-based RESTful API that handles model operations on a private NVIDIA DGX server

Backend API

The Flask API provides four main endpoints:

1. /load_model

Loads a specified forecasting model into memory.

# Example usage
POST /load_model
{
  "model_name": "LPTM"  # or "TimesFM"
}

2. /upload_dataset

Allows users to upload custom time-series datasets.

# Example usage
POST /upload_dataset
Form data:
- dataset: [CSV file]
- model_name: "LPTM"  # or "TimesFM"

3. /finetune

Fine-tunes a loaded model on the uploaded dataset.

# Example usage
POST /finetune
{
  "model_name": "LPTM"  # or "TimesFM"
}

4. /run_inference

Runs inference on the loaded model and returns forecast visualizations.

# Example usage
POST /run_inference
{
  "model_name": "LPTM"  # or "TimesFM"
}

Features

• Model Selection: Choose from several state-of-the-art time-series forecasting models
• Dataset Management: Upload and manage custom time-series datasets
• Fine-tuning Capabilities: Adapt pre-trained models to specific domains
• Visualization: Interactive plots of historical data and forecasts
• Research Integration: Access to research methodologies and comparative performance metrics
• Benchmarking: Compare LPTM performance against other foundational models

Implementation Details

The backend is implemented as a Flask API with CORS support, specifically configured to work with the frontend hosted at samay-dashboard.vercel.app. The API communicates with the models and datasets stored on an NVIDIA DGX server.

Model Support

Currently, the system supports:

• LPTM: Large Pre-trained Time-Series Model with configurable forecasting horizons
• TimesFM: Google's TimesFM model (1.0-200m-pytorch) for comparative analysis

Forecasting Visualization

The system generates visualizations that show:

• 512 timesteps of historical data
• 192 timesteps of ground truth
• 192 timesteps of model forecast

Research Background

This project implements and uses the models and methods described in Large Pre-trained Time-Series Models. The LPTM model forms the core of our forecasting capabilities, with other models included primarily for benchmarking and performance comparison.