## TL;DR
Time series forecasting predicts future values from historical sequences. Traditional statistical methods (ARIMA, ETS) compete with deep learning approaches (LSTM, Transformer variants) depending on data volume and pattern complexity.

## Core Explanation
Classical decomposition: trend (long-term direction), seasonality (periodic patterns), and residuals (noise). ARIMA models capture auto-regressive and moving-average dynamics. Prophet (Facebook) decomposes time series into trend + seasonality + holiday effects with interpretable parameters.

## Detailed Analysis
Deep learning for time series: LSTMs handle long-term dependencies; Transformers capture global patterns but require positional encoding modifications. PatchTST segments series into patches, treating each patch as a token — significantly improving transformer efficiency for long sequences.

## Further Reading
- Nixtla: StatsForecast & NeuralForecast Libraries
- Kaggle: Time Series Competitions
- Hyndman: Forecasting: Principles and Practice