Comparing Seasonal Time Series Requires Validation Beyond Correlation Distance

Summary

The document considers a method for comparing calendar months across historical data: calculate correlations across eight daily variables, then use Euclidean distance to identify similar months. The replies point toward established time-series methods for periodic data, including seasonal modeling, and mention clustering as another possible way to group observations.

The responses do not evaluate the proposed correlation-and-distance procedure, provide an empirical comparison, or establish that it predicts market behavior. Instead, they caution against analyzing seasonal patterns informally when dedicated time-series methods are available, and emphasize that any candidate model should be judged out of sample. The discussion is brief and offers references rather than implementation details, so it leaves questions about feature scaling, dependence between observations, and the design of a valid forecasting test unanswered.

Key ideas

• The proposed comparison uses correlations across multiple daily variables and Euclidean distance to find similar historical months.
• Seasonal time-series methods provide a more systematic framework for studying periodic patterns.
• Clustering can be considered when grouping similar observations fits the research question.
• Out-of-sample performance is essential for assessing whether a pattern has predictive value.

Tags

# Does the correlation of matrices have explanatory power when building a pattern recognition model?


# Does the correlation of matrices have explanatory power when building a pattern recognition model?












I'm using 8 different variables (with daily observations) with the purpose to compare different months across the historical data. For that purpose I calculate the correlation between each month and the historical months in the data and then calculate the Euclidean distance in order to find the closer month.

Does it make sense? Is there any literature regarding such experiments?

## Answer by Richi Wa (score 2)

https://quant.stackexchange.com/a/15517

There is a vast literature on modelling time-series with periodcities. Rob Hyndman is one of the leading reseaerchers in this area. He has published the R package `forecast` and a free online text book on this subject (with another package and R code in the book). Your task is covered starting here.

## Answer by James (score 0)

https://quant.stackexchange.com/a/14800

Analyzing seasonal time series "by hand" is not a good idea because there is a lot of time series machinery developed just for that. A simple example in R can be found here.

You can apply clustering if it feels more natural but the main question is whether your model works out of sample.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.