Sequential quality

This metric is aimed at time series data with a regular frequency.

The sequential metric assess the similarity in the size and fluctuation in numerical values that vary over the sequence. The similarity is compared between the source and synthesised versions of the data. In order to assess similarity, each sequence in the source and synthesised data is considered separately and a range of metrics are calculated for each sequence. For example, the maximum and minimum values for the sequence are calculated, the number of times the sequence crosses 0 is calculated and so on. For a full list see the temporal and statistical sections of the TSFEL (Time Series Feature Extraction Library) documentation. Each metric is calculated for each sequence and then averaged over all the sequences and the resulting number compared between the source and synthesised data to produce a similarity statistic. The final similarity is the average of the individual similarities.

The synthesiser automatically detects whether the time series data has a regular frequency. If the metric is required in other situations then it is possible to force it to be generated by setting assume_fixed_frequency to True.

This can also be used for sequences with nearly regular frequencies (for example, a heart monitor that occasionally misses a reading). However, it does not completely assess the similarity in time gap between widely varying frequencies, such as bank transactions, or assess the correlation between records.