Прогнозирование временных рядов

Time series forecasting критично для capacity planning, revenue forecasting и operational decisions.

Use cases

Traffic forecasting: капацити планинг infrastructure. Revenue projection: 30/90/365-day forecasts. Peak load prediction: sport events, promotions. Churn rate trends. Fraud volume forecasts. FX rate hedging. Customer service staffing.

Модели

Classical: ARIMA, SARIMA — good baseline. Prophet (Facebook): friendly с seasonality, holidays. Deep learning: LSTM, Transformer, TFT (Temporal Fusion Transformer), NBEATS. Ensembles: комбинация классических и neural. AutoTS, AutoML tools для быстрого prototyping.

Features engineering

Lag features: value at t-1, t-7, t-30. Rolling statistics: moving averages, std. Holiday effects: soccer matches, holidays. External regressors: marketing spend, weather. Seasonality: day-of-week, month. Trend decomposition. Каждая feature — potential improvement.

Evaluation

MAPE, sMAPE для percentage errors. MAE, RMSE. Backtesting: rolling window validation. Compare vs naive baselines (yesterday's value). Business metrics: forecast accuracy → cost savings. Prediction intervals обязательны для uncertainty quantification.