870K
Headlines Processed
9
Live API Endpoints
6
FRED Macro Series
9
Build Phases
Overview
Volarix is a production-grade GenAI financial intelligence system built from the ground up. The system ingests 870,000+ deduplicated financial headlines sourced from Kaggle and the NewsAPI, combined with six macroeconomic time series from FRED (Federal Reserve Economic Data), to model, predict, and explain market stress regimes.
The entire pipeline — from raw data ingestion through to a live React dashboard — is deployed on HuggingFace Spaces via Docker with a FastAPI backend exposing 9 working REST endpoints. The frontend is deployed separately on Vercel.
Architecture
PyTorch Temporal Fusion Transformer
From-scratch TFT implementation for multi-horizon time series forecasting. Attention mechanisms over both static and time-varying features with interpretable variable importance.
XGBoost Regime Classifier + Platt Calibration
Gradient boosted regime classification with Platt scaling for calibrated probability outputs. Trained on engineered macro + sentiment features.
Split Conformal Prediction
Distribution-free uncertainty quantification producing guaranteed coverage intervals. Applied over forecast outputs for statistically valid prediction bands.
HMM Regime Baseline + Historical Analogue Engine
Hidden Markov Model baseline for unsupervised regime detection, paired with a historical analogue engine that retrieves the most similar past market environments.
Live API Endpoints
GET/healthSystem health check
GET/regime/currentCurrent market regime classification
GET/forecast/stressMulti-horizon stress forecast with conformal intervals
GET/analoguesHistorical market analogues for current environment
GET/macro/latestLatest FRED macro series snapshot
GET/sentiment/latestLatest news sentiment aggregation
GET/vix/historyHistorical VIX series with regime labels
GET/model/featuresTFT feature importance scores
GET/dashboard/summaryFull dashboard data bundle
Tech Stack
PyTorchXGBoostFastAPIDocker
HuggingFace SpacesReactVercel
FRED APINewsAPIKaggle
Conformal PredictionHMMPlatt Scaling
PandasNumPySeaborn