Climate CO2 Forecasting ML

Overview

A forecasting workflow spanning validation, backtesting, interval monitoring, experiment evidence, serving, and visualization.

Problem

Time-series models can appear strong when evaluation leaks future information or compares models on inconsistent windows.

Solution

Built a reproducible forecasting workflow with leakage-safe rolling backtesting, interval evaluation, experiment tracking, registry-style metadata, API contracts, and a frontend evidence view.

Technical Decisions

• Chronological evaluation is shared across model families.
• The local deterministic demo remains separate from optional hosted infrastructure.

Outcome

The project keeps model comparison reproducible and reports limitations instead of treating model complexity as evidence of quality.

Key Features

• Rolling-origin evaluation prevents future leakage.
• Interval monitoring complements point forecasts.
• Experiment and registry metadata connect training evidence to serving.

Architecture

1. 01

   CO2 dataset

2. 02

   Validation

3. 03

   Rolling backtest

4. 04

   Model comparison

5. 05

   Interval evaluation

6. 06

   Registry metadata

7. 07

   API and frontend

Tech Stack

• Python
• Forecasting
• Backtesting
• Intervals
• FastAPI
• React
• Experiment tracking

Verification

• Leakage-safe rolling backtesting
• Interval evaluation
• Offline deterministic demo and API checks

Security & Privacy

• The workflow uses public climate data and requires no private user data.
• Environment-specific credentials are not part of the default review path.

Limitations

• Forecast results are an educational engineering demonstration, not climate-policy guidance.
• Public repository evidence is provided without claiming a live full-stack deployment.

Future Improvements

• Extend probabilistic forecast calibration.
• Compare additional rolling windows and drift signals.