Time-series drift detection and adaptive retraining — benchmark datasets and baselines
Description
Eight time-series datasets (5 real + 3 synthetic with known drift points) for studying concept/data drift detection and adaptive model retraining in forecasting, plus no-adaptation baseline SMAPE values computed from the experiment logs.
Files
Steps to reproduce
Code (thesis Appendix A): https://github.com/marinaleb81/vkr_code 1. Install: pip install -r req.txt (Python 3.10+). 2. Reproduce experiments H1–H4: PYTHONPATH=src python scripts/run_all_experiments.py 3. baseline_metrics.csv holds the no-adaptation baseline SMAPE per dataset and model, averaged over 5 seeds, computed from the experiment logs (records with detector="no_drift", strategy="none"). Real datasets are in data/ as <name>_{train,val,test}.csv. Synthetic series have known drift points (concept@500, data@500, gradual 400–700). See DATA_DESCRIPTOR.md for columns, units, and provenance.