Each one starts from a question I could not answer with the data I had, and ends with something someone else can rerun.
The same imputation reconstruction can score as excellent or as poor depending only on how the error is normalised. A sensitivity analysis of reporting choices in imputation benchmarks.
Four questions about Portuguese public procurement, the valid SQL that answers them, and why every answer is wrong. 192,036 contracts, 15.5 bn euros.
A nearest-neighbour imputation method for mixed-type data, benchmarked against six tuned competitors over twelve datasets — with an honest map of where it fails.
More in the pipeline — emergency-department demand forecasting and a co-absence network study, both on public data. They go live here once there is something measurable to show.