Loading
How to describe and forecast data that arrives in order. The course covers smoothing, autocorrelation and stationarity, autoregressive models, windows and honest backtests, dynamic time warping, and neural forecasters built from LSTMs and causal convolutions.
01
A moving average replaces each point with the mean of a window. Exponential smoothing gives recent points more weight, and Holt's method adds a trend term.
02
Autocorrelation measures how much a series resembles a shifted copy of itself. Most classical models assume the series is stationary, so its mean and spread do not drift.
Practise it
03
An AR(p) model predicts the next value as a weighted sum of the last p values. Fitting it is a linear regression on lagged copies of the series.
Practise it
04
Sliding windows turn one series into many supervised rows. The test data must always come after the training data, and scaling must use training statistics only.
05
DTW aligns two series that move at different speeds and then measures their distance. Dynamic programming finds the cheapest alignment path.
Practise it
06
An LSTM reads a window step by step and keeps a memory. A stack of dilated causal convolutions sees a long history and never reads the future.