Time Series Forecasting

Time series forecasting predicts future values based on historically ordered observations. From stock prices to server load to disease spread, time series data is everywhere. This skill covers the full spectrum from classical statistical methods (ARIMA, exponential smoothing) through modern approaches (Prophet, neural networks with LSTMs and transformers), including seasonal decomposition, trend analysis, and anomaly detection.

## Key Points

1. Always decompose the series first (trend, seasonal, residual) to understand what you are forecasting
2. Test for stationarity before applying ARIMA; differencing is the standard remedy for non-stationarity
3. Exponential smoothing (Holt-Winters) is a strong baseline that is often competitive with complex methods
4. Prophet excels at business forecasting with multiple seasonalities and interpretable components
5. LSTMs and transformers require more data but can capture nonlinear patterns that statistical methods miss
6. Anomaly detection should run alongside forecasting to flag when the generating process may have changed
7. Model selection depends on data size, complexity, and whether interpretability is required
8. Always evaluate with proper time-series cross-validation (expanding or sliding window), never random splits