The guide
Time series break the assumption that rows are independent. Today's value depends on yesterday's, so you cannot shuffle the data, and a random train and test split leaks the future into training.
The classical tools are still strong. Moving averages and exponential smoothing separate the trend from the noise. Autocorrelation shows how far back the series remembers, and that is how you choose the order of an autoregressive model. An AR model is linear regression on the series' own past.
Sliding windows turn a series into a normal supervised dataset for any model, including neural networks. Dynamic time warping answers a different question: how alike are two sequences that have the same shape but run at different speeds?