Event Details

Predicting partially observed survival curves via factor analysis with application to demand forecasting in short-term rental markets

We construct prediction models for population-level survival curves sampled from a heterogeneous mix of populations. The models consider a paradigm where a curve for a certain population is partially observed and a forecast of the curve’s remainder is useful for downstream decision making. The problem has a discrete-time perspective and the aim is to recast across-population hetereogeneity as a multivariate sampling model under a Lehmann-type survival formulation. We propose two PCA-based forecasting models that take the partially observed survival curve, as well as fully observed historical survival curves into account. The work is motivated by models of market occupancy for the short-term vacation rental market, where forecasts of market occupancy for a given date in the future are subsequently fed into a dynamic pricing algorithm. The performance of the models is investigated on a recently released Wheelhouse dataset of short-term vacation rental market occupancy. We compare the performance of the models using the Integrated Quadratic Distance (IQD) and find that the PCA-based forecasting models outperform Holt's linear trend model on these data.

This talk is based on joint work with Alex Lenksoski (Norwegian Computing Center) and Thordis Thorarinsdottir (University of Oslo).