I’m a Junior Data Scientist passionate about time series forecasting and financial market analytics. I build end‑to‑end ML pipelines—from data ingestion and feature engineering to model deployment—for real‑world problems on MOEX.

Moex_predict is my flagship project, forecasting Moscow Exchange blue‑chip stocks (SBER, GAZP, LKOH) with cutting‑edge Transformer architectures and rigorous hyperparameter tuning.

• Performance: MAPE 0.90%, MaxErr 24.7 RUB on hold‑out SBER data
• Features engineered: RSI, MACD, Bollinger Bands, ATR, log‑returns, volatility, SMA
• Modeling: Encoder‑only Transformer with positional embeddings, ensemble strategies
• Optimization: Optuna study targeting 95th percentile error
• Data handling: Robust MOEX API pagination, caching, and preprocessing
• Deployment-ready: FastAPI endpoint serving live predictions

Forecasting MOEX blue-chip stocks (SBER, GAZP, ROSN) using Transformer models with advanced feature engineering and hyperparameter tuning.

• Performance: MAPE 0.90%, MaxErr 24.7 RUB (SBER)
• Features: RSI, MACD, Bollinger Bands, ATR, log-returns, volatility, SMA
• Model & Optimization: Encoder-only Transformer, Optuna tuning on 95th percentile error, ensembling
• Deployment: FastAPI service with Docker

Retail sales forecasting for a major retailer using XGBoost and Optuna.

• Performance: RMSLE 0.7509 (Top 450 on Kaggle)
• Tech: Python, XGBoost, Optuna, TimeSeriesSplit
• Impact: Improved demand planning and inventory management

1. Data Ingestion & Cleaning: MOEX API pagination, missing‑value handling, caching raw data
2. Feature Engineering: Technical indicators and statistical metrics for robust signal extraction
3. Model Development: Transformer encoder with attention, custom positional encoding, fine‑tuned via Optuna
4. Ensembling & Calibration: Combine multiple model seeds and architectures, quantile‑based aggregation
5. Evaluation & Monitoring: Time‑series cross‑validation, custom loss for tail‑error minimization
6. Deployment: FastAPI microservice for real‑time inference, Dockerized for scalability

• Dive into Temporal Fusion Transformers (TFT) for multi‑horizon forecasting
• Experiment with Informer and Reformer for long sequence efficiency
• Integrate real‑time market data streams and sentiment features
• Expand ensemble with Gaussian Process and Bayesian Neural Nets for uncertainty quantification

• ✉️ Email: nasdorm.ml@inbox.ru
• 💬 Telegram: @Nasdorm
• 🐙 GitHub: github.com/NasdormML

“Perfection is not attainable, but if we chase perfection we can catch excellence.” – Vince Lombardi