Archipelago
Cross-sectional market analysis tool for comparing 2000+ stocks across multiple metrics simultaneously.
Full-Stack Developer
2024–Present
Next.js 15 · TypeScript · Visx · Python/FastAPI · TimescaleDB
Problem
Momentum screening workflows broke above ~150 tickers when done in spreadsheets. Needed to compare stocks across multiple metrics at once while preserving distribution shape—something rank tables can't do.
- —Render 2000+ data points across 4 metrics without lag
- —Support multiple time scales per metric (1W, 1M, 3M, 6M, 1Y) in the same view
- —Support time-series queries and relational joins without splitting systems
Key Decisions
Four chart types instead of tables
Voronoi scatter, treemap, parallel coordinates, and deviation decomposition
Each chart answers a different analytical question—clustering, relative weight, multi-metric comparison, and contribution analysis. Target user: analyst running weekly screens.
Composite metric IDs
Encode time scale in metric identifier (e.g., "change_pct:1m")
Needed the same base metric on multiple axes with different lookback periods. Simplifies state management and allows saved views to persist cleanly.
Pre-calculated time scale metrics
Dedicated time_scale_metrics table with LEFT JOIN, computed during daily collection
Avoided runtime calculations for 2000 tickers × 6 time scales on every request. Query returns all metrics instantly.
Outlier detection
Extreme values (microcaps with 400%+ momentum) compressed the useful range into a corner. Log scales broke intuition; fixed-percentile clipping hid real anomalies.
Built IQR-based detection with three modes: hide (clean view), highlight (flag for review), or isolate (show only outliers). Sensitivity is per-metric, since volatility and value outliers follow different distributions. Configuration persists in saved views.
Outcomes
- —2000 stocks render across 4 chart types with no visible lag
- —Interaction latency: 2–3s → sub-50ms after moving filtering client-side
- —Saved views eliminated ~15 minutes of session setup per analysis
- —Dashboard codebase reduced 64% (4,587 → 1,643 lines) through component extraction