lguayacan

Where do cost outliers come from?

SQL analysis in BigQuery

Laura Camila Guayacan Ardila

Question: which projects look unusually expensive per square foot, and why? Same 6,123-project dataset (Kaggle source) as the regression case study.

Outliers by area

Equipment share

Result in four lines

Not project type. Share flagged: Residential 7.4%, Commercial 6.8%, Industrial 6.5% (chi-square p about 0.56).
Size. 84% of projects under 2,000 sq ft are flagged; above 5,000 sq ft, none.
Fixed cost. Equipment cost is about 100,000 USD in every size band, so it is 32% of cost on small projects and 7% on large ones. Removing it cuts the outliers by half (423 to 205).
Not weather or risk. Correlations with cost per sq ft are near zero ( r below 0.05).

Files: All SQL queries (10 queries, commented)

Cross-check: the IQR fences and the 423 flagged projects match the Google Sheets workbook exactly (upper fence 194.85 USD per sq ft).

Takeaways

- The IQR rule applied to a ratio such as cost per sq ft mostly detects a scale effect, not anomalies. A fairer screen compares each project with others of similar size, or uses regression residuals. - About half of the effect is explained by equipment cost. The rest is unexplained. A plausible but untested hypothesis: other fixed costs the dataset does not record (mobilization, permits, design).

Limitations

Synthetic data; no actual-versus-planned costs; outliers are flagged and documented, not removed. The lack of weather and risk effects says nothing about real projects.