Overview

Filtering refers to a process of transforming a time series signal and producing a new signal. Many times, filters are designed to try to remove noise from a signal.

In the context of deep learning, a time series filter is roughly equivalent to the convolution portion of a convolutional neural network.

Filters

The following are common filters used in time series analysis.

• Moving Average : The moving average is a standard workhorse of time series analysis. It is particularly prominent in finance. Moving averages are often used to smooth a signal, as well as used in trend identification. Moving averages include the Exponentially Weighted Moving Averages, which calculates a moving average but weights the points differently.
• Hodrick Prescott : The Hodrick Prescott filter is also used to smooth a signal. The filter works by considering a trade off between having the filter values close to the signal values and the amount of curvature in the resulting signal. That is, the filter tries to reduce the amount of curvature (or twisting) in the signal.
• State Space Model / Kalman Filter :