Observational data is collected under naturally occurring conditions rather than through deliberately assigned interventions. Its collection process shapes what can be inferred: the observed distribution may differ from the natural distribution, and valid analysis requires methods suited to that motivation. section 4.5.1
Treat observational data as a delivery and modeling risk to be recorded in project artifacts:
- Create a collection-motivation ticket describing why the data exists, which variables were selected, and which common cases or related entities may be missing. section 4.5.1
- Separate observational data from controlled experimental data before choosing an inference approach; selection protocols can control bias for one variable while introducing it for others. section 4.5.1 section 5.4.1
- Add a time-period checkpoint before fusion: operational snapshots may combine different ranges, calendar effects, or trading patterns and distort production models. section 5.4.1
Collection choices create blind spots
A butterfly collection that favors rare specimens, omits other insects, or records only selected seasons does not represent the surrounding population. Similarly, combining retail snapshots taken on different weekdays can shuffle prime trading days across periods and mislead a model. section 4.5.1 section 5.4.1
