Practical Neural Network Recipies in C++

60

Chapter

4

Differencing

One very popular technique for preprocessing time-series data is to

compute the difference between successive points. Thus, rather than

working with the raw data, we work with its changes. This can have

many advantages, but a few dangers lurk about also.

First, note that differencing a time series completely eliminates

a constant trend by converting it to a constant offset. Slowly changing

trends are usually eliminated enough that they can be ignored. Even

seasonal trends, if they have a long enough period, are often eliminat-

ed for all practical purposes. All the complexities of the preceding

sections on trend ca