FlexMeasures / flexmeasures / 29761771462

82%

main: 82%

push

github

Filter annotation regressors by their own belief time

Annotations carry a belief time of their own, which may be NULL. Read it
instead of stamping every annotation with the forecast vantage point, and
treat NULL as "always known" -- the case for the holiday calendars written
by `flexmeasures add holidays`, which record no belief time.

This removes the row-duplication workaround in split_data_all_beliefs, which
fabricated a second copy of every annotation row at
`belief_time = event_start + resolution` so it would satisfy the realized-only
filter. That filter tests `belief_time > event_start` to pick measurements over
forecasts, which is the wrong question for a calendar: a holiday is never
realized after the fact, so a truthfully stamped annotation was invisible to
the training window. Annotation regressors are now exempt from the
realized-versus-forecast split and are governed solely by their own belief
time, so every belief_time in the frame means what it says.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E6t3NeRRhFtLc4KvCCxFBt
Signed-off-by: F.N. Claessen <felix@seita.nl>

38 of 43 new or added lines in 1 file covered. (88.37%)

14895 of 18252 relevant lines covered (81.61%)

0.82 hits per line