Add an anomaly detection for standalone indicators

[larsyencken]

Overview

Using ChatGPT or another method, create an online method for detecting notable outliers or potential data errors.

Prior art

See the "Find interesting views" tab of the ETL Insighter, which truncates an indicator to the last 20 years and feeds the data plus a prompt to ChatGPT 4o.

Challenges

• The 128k context window is not enough to put the entire data series in
  • You could pass the data file for use with code interpreter (likely worse results)
  • You could truncate the data (e.g. last 20 years)
  • You could split the data into multiple chunks, and do detection on each chunk
  • You could thumbnail subsets of the data, and do detection based on thumbnails