anomaly-detection-engine

Financial metrics are noisy on purpose: Mondays don’t look like Fridays, 9 AM doesn’t look like midnight, and a classic z-score panics at every payday. This engine watches real-time financial streams without crying wolf.

How it works

Three layers: ingest, statistics, insight. The statistics layer uses a Modified Z-Score built on MAD (Median Absolute Deviation) instead of standard deviation, so a single whale transaction can’t drag the baseline. Day-of-week normalized baselines and same-hour comparison handle weekly seasonality. Result: 60% fewer false positives than the naive approach it replaced.

The insight layer

When something does fire, GPT-4 (via OpenRouter) gets the alert plus cross-metric context — net flow vs. cash out, funnel conversion ratios — and writes a human-readable root cause analysis. The on-call human reads a paragraph, not a wall of numbers.

Open source

The statistical core of this engine now lives on PyPI as madwatch: pip install madwatch gets you the same MAD-based Modified Z-Score, streaming detection and day-of-week baselines.