AI-Driven ML: Automating Time-Series Forecasting with Anton

Use AI to do ML - vibe forecasting is coming

Anton is an open-source AI agent designed to execute full-stack data science tasks from database connection to model deployment via natural language. In a live demonstration, the tool processed $183.9M in aerospace purchase data to generate a 6-month demand forecast in under two minutes.

Why This Matters

Traditional forecasting involves a week-long manual ‘tax’ of cleaning Postgres data, engineering lag features, and validating XGBoost baselines. Anton addresses this by automating the boring half of data work, enabling what is termed ‘vibe ML’—where analysts describe requirements and the agent handles the reliable implementation of well-known statistical methodologies.

Key Insights

• Anton executes Python and SQL within a local sandbox on Mac, Linux, or Windows, ensuring credentials stay in a local vault.
• The agent automates feature engineering including lags at 1, 2, and 4 weeks, 3-month rolling averages, and cyclical month encoding.
• Forecast transparency is provided through a ‘scratchpad’ that logs every step as a Jupyter notebook for human inspection.
• A production test on aerospace electronics data yielded a 14.6% MAPE on quantity and a 49.9% MAPE on spend due to high-cost variance.
• Interactive HTML dashboards are generated and hosted instantly without requiring platform engineering for deployment.

Practical Applications

• Supply Chain Management: Forecasting parts demand for programs like F-35 sustainment to shift from week-long manual analysis to 2-minute automated cycles.
• Automated EDA: Using natural language to map Postgres schemas and identify trends like 84% on-time delivery rates without manual SQL querying.
• Executive Reporting: Generating live-linked dashboards for $183.9M spend overviews that combine actuals, fitted lines, and confidence bands.
• Pitfall: Blindly trusting spend forecasts with high cost-variance components which can dominate error metrics and require manual caveats.

References:

• https://dev.to/mindsdbteam/use-ai-to-do-ml-vibe-forecasting-is-coming-4jkg
• https://github.com/mindsdb/anton
• https://4nton.ai/view/90ccae06f/386c3121