Drive failure prediction — Q1 2024 S.M.A.R.T. telemetry

Gradient boosting on engineered S.M.A.R.T. window features. All numbers in docs/metrics.json.

Test ROC-AUC

1.000

Test PR-AUC

0.999

Precision

99.4%

Recall

97.0%

False alarm rate

0.1%

Fleet drives

60,000

Failures in fleet

978

Features

307

ROC curve (held-out test drives)

Precision-recall curve (held-out test drives)

Detection lead time — recall vs. days before failure

Top features by importance

Lead-time analysis

Window ends this many days before failure	Drives	Flagged	Recall	Mean risk
1	928	923	99.5%	0.990
3	902	895	99.2%	0.989
7	853	849	99.5%	0.987
14	778	768	98.7%	0.973
30	590	576	97.6%	0.965

S.M.A.R.T. data © Backblaze (CC BY-SA 4.0), Q1 2024, via backblaze.com/cloud-storage/resources/hard-drive-test-data. Model, metrics and reproducibility details are in the project README. Raw telemetry is not included in the repository.