Tech Stack
Under the hood, for the curious.
Prediction Pipeline
Daily prices go through 42 causal features with fractional differencing, then an XGBoost direction model trained on triple-barrier labels with purged, embargoed walk-forward cross-validation. Probabilities are isotonic-calibrated, and GARCH-shaped conformal bands give 80% and 90% price ranges that cover what they claim.
An LLM verifier reads recent headlines and can lower or veto each call. Every forecast is stored and graded when it matures, so the track record on the Advisor page is real.
The Layers
Built and maintained by one developer, on free tiers.
Routers for market data, predictions, trading, payments and the AI advisor. APScheduler runs ingestion and the nightly prediction job.
Users, wallets, trades, price history and every stored prediction, with versioned migrations that run at boot.
A static site on Cloudflare Pages, the API on Render, GitHub Actions CI and a Cloudflare Worker that keeps the API warm.