StockXpert

What it is

A deep-learning equity prediction and stock recommendation platform covering all 100 Nifty NSE equities. The model predicts cross-horizon (1D–10D) returns with calibrated confidence, served behind a sub-second API and a full-stack trading dashboard.

Key points

• Hierarchical model architecture — combines ResNet, GRU, Bi-LSTM, and multi-head attention to capture cross-horizon price dependencies, outperforming single-architecture baselines.
• Leakage-free feature pipeline — RSI, MACD, ADX, Bollinger Bands, SMA/EMA distance, and volatility indicators computed from raw OHLCV across strict train/validation/test splits, so no future information bleeds into training data.
• Multi-objective Huber loss — directional regularisation and tail-event oversampling improved short-horizon return correlation by 12% over baseline and reduced directional error on high-volatility equities.
• Snapshot-first serving layer — Cloudflare R2 storage with market-aware TTL caching delivers sub-second API responses for all 100 Nifty stocks during market hours and eliminates redundant inference cost.
• Full-stack dashboard — Next.js + FastAPI; live screener with real-time alerts, OHLCV candlestick charts, technical-indicator overlays, and forecast tables (1D–10D) with calibrated confidence scores.

Result

End-to-end production system: data pipeline, model training, snapshot serving, and trading dashboard — all with reproducible eval harnesses.