Anomaly Detection Engine

[rohitdash08]

Feature

Detect unusual spending spikes by category or merchant.

Why This Matters

Key insight feature for user retention.

Acceptance Criteria

• Flags anomalies vs baseline
• Includes explanation metadata

Test Plan

• Baseline comparison tests

[aliraza556]

@rohitdash08 Please assign me?

[aliraza556]

Hi @rohitdash08,

I'd like to take on the Anomaly Detection Engine. I've already gone through the codebase and understand how expenses, categories, and the existing insights layer work -- so I can hit the ground running.

Here's my plan:

1. Baseline Calculation
For each user, I'll compute rolling averages (mean + standard deviation) per category and per merchant over the last 30/60/90 days. These get recalculated whenever new expenses are added -- same cache invalidation pattern we already use for budget suggestions.

2. Anomaly Detection (keeping it practical)
No heavy ML libraries needed. I'll use straightforward statistical methods:

• Z-score for categories with enough data -- flag anything beyond 2 standard deviations from the user's baseline
• IQR method as fallback for categories with skewed or sparse data (e.g., medical, travel)
• Frequency detection -- catch unusual patterns like multiple transactions at the same merchant in a short window when that's not normal for the user

Each method runs in order and we take the strongest signal. Simple, testable, no black boxes.

3. Explanation Metadata
Every flagged anomaly will include:

• anomaly_type -- spike, frequency, or new_merchant
• severity -- low / medium / high based on how far it deviates
• baseline_value vs actual_value for transparency
• explanation -- plain English, e.g. "Your dining spend this week is 3x higher than your usual weekly average"

4. API

• GET /anomalies -- list anomalies with filters (date range, category, severity)
• GET /anomalies/summary -- quick stats for a dashboard widget

5. Tests

• Baseline comparison tests (as specified in acceptance criteria)
• Unit tests for each detection method with controlled data
• Edge cases: new users with little history, categories with only 1-2 transactions, no anomalies scenario

Everything stays within the existing stack -- Python, Flask, SQLAlchemy, numpy for stats. No unnecessary dependencies.

Happy to discuss further if needed.

[alexchenai]

/attempt #72

Hi, I would like to implement the Anomaly Detection Engine. I have already implemented the savings opportunity detection engine (#119, PR #532) using the same SQLAlchemy patterns and Flask Blueprint architecture. I plan to build a statistically robust z-score based anomaly detector for both category-level and merchant-level spending, with explanation metadata. ETA: 24 hours.

[pabloDarkmoon24]

Hi! I've worked on a fix for this issue.

PR: #630

The implementation addresses the requirements described here. Happy to iterate based on your feedback!