How We Covered Concept Drifts In Time Series Forecasting For Public Transport

The Coronavirus lockdowns altered public transport occupation data in very unpredictable ways. Ultimately, these changes in occupation data are perfect examples of sudden concept drifts that can be blockers in most machine learning deployments. We managed to overcome the obstacles by developing methods and engineering features that allowed us to adjust forecasts based on unforeseen changes in the occupation data. In this talk, we give insights into our journey from idea development to the ways how we overcame the challenges and share our learnings.