Natural Language Time Series Forecasting with Google’s TimesFM Model
NEO built ChronoSight, a full-stack forecasting application that lets you query time series data in plain language using Google’s TimesFM-2.5-200M zero-shot forecasting model.
Problem Statement
We asked NEO to: Build a time series forecasting system that removes the usability gap—no data science background, no manual data shaping. Users ask questions in plain language and receive forecasts with confidence bands and trend analysis. The system should use a zero-shot foundation model so no fine-tuning is required.
Solution Overview
NEO built ChronoSight — a natural language time series forecasting platform:
- TimesFM-2.5-200M — 200M-parameter foundation model from Hugging Face; zero-shot forecasting with no training on your data
- Natural Language Query Routing — Parse queries (e.g. “Apple stock forecast next 30 days”), resolve data source, fetch history, run inference, return forecast with visualizations
- Multiple Data Sources — Live market data (Yahoo Finance), curated economic datasets (CO2, airline passengers, etc.)
- Full-Stack App — FastAPI backend, React 19 + TypeScript frontend with Recharts; confidence bands and trend summary
Workflow / Pipeline
| Step | Description |
|---|---|
| 1. Natural Language Query | User types a question (e.g. “Forecast airline passengers through Q4”) |
| 2. Query Parsing & Routing | Identify data source, fetch historical series; suggest reformulation if no match |
| 3. TimesFM Inference | Zero-shot forecasting; point estimate and confidence bands |
| 4. Visualization & Metrics | Historical + forecast chart; trend direction, magnitude, natural-language summary |
Technical Details
- Backend: FastAPI, Python 3.10+; model loaded once at startup
- Frontend: React 19, TypeScript, Vite, Tailwind CSS, Recharts
- Requirements: ~2 GB disk for model weights, 4 GB RAM minimum; backend port 8000, frontend 5173
- Disclaimer: Forecasts for educational/exploratory use; not for production financial decisions without validation and monitoring
Repository & Artifacts
Generated Artifacts:
- FastAPI forecasting API with async inference
- Natural language query router and data source integration
- React frontend with charts and confidence bands
- Support for Yahoo Finance and statsmodels-style datasets
When to Use This Approach
- Quick exploratory analysis on multiple series without training per-series models
- Limited historical data where fine-tuning would overfit
- Baseline to compare against specialized forecasting approaches
- Speed of iteration matters more than marginal accuracy gains