Data Scientist focused on time-series forecasting, financial analytics, and production-oriented ML.
I build forecasting systems end to end—from reliable data collection and time-aware preprocessing to model training, experiment tracking, APIs, and containers. My main interests are financial time series, demand forecasting, and reproducible MLOps.
🚀 Featured Project — Moex_predict
Moex_predict is a modular, five-step closing-price forecasting system for Moscow Exchange equities such as SBER, GAZP, and ROSN.
Best published SBER result: 0.90% MAPE, improved from 1.20%—a 25% relative reduction in average percentage error in the v1.5 experiment.
Technical highlights
| Area | Implementation |
|---|---|
| Data pipeline | MOEX ISS pagination, retries, local caching, market-index data, and USD/RUB context |
| Feature engineering | Returns, volatility, lag features, RSI, MACD, Bollinger Bands, and ATR |
| Models | Attention-LSTM, TCN, and Transformer encoder models for multi-step forecasting |
| Training | Hydra experiment configs, Optuna tuning, Huber loss, AdamW, early stopping, pruning, and gradient clipping |
| Experiment tracking | MLflow runs plus versioned model, scaler, and metadata artifacts |
| Model lifecycle | Performance monitoring and staleness-based retraining logic |
| Serving | FastAPI endpoints for training, prediction, model inventory, and health checks |
| Delivery | Docker Compose plus unit and integration tests |
Why it matters: this is not only a forecasting notebook. It covers the full ML lifecycle—from raw market data to a reproducible, API-served prediction system.
End-to-end retail demand forecasting with XGBoost, Optuna, and time-aware validation.
- Achieved RMSLE 0.75094 on the Kaggle test set
- Improved over moving-average and linear-regression baselines by 15%
- Used lag, rolling, seasonal, and holiday features with
TimeSeriesSplit - Added Pytest coverage, Docker packaging, and CI/CD automation
Probabilistic forecasting, uncertainty estimation, and reinforcement learning for data-driven decision systems.
- Email: nasdorm.ml@inbox.ru
- Telegram: @Nasdorm