Previous domain adaptation methods are designed to work for a single task, either classification or regression. In this paper, the task of the learner is to produce both an estimation and an ordinal classification of instances that are consistent in that the classification of instances into quantiles is derived from the estimated values. We propose an extension of the boosting for transfer method (TrAdaBoost), Joint Quantile Loss Boosting Domain Adaptation (TrAdaBoost.JQL) for regression transfer learning, that aims to jointly minimize regression and ordinal classification errors. Motivated by the real-world problem of poverty targeting using the Proxy Means Test, we empirically show that TrAdaBoost.JQL can consistently reduce RMSE and inclusion and exclusion errors for estimating per capita household expenditure, across a wide variety of districts in Indonesia, compared to other reweighting-based and invariant feature representation-based domain adaptation methods. We design TrAdaBoost.JQL to be flexible as to the chosen eligibility (poor) threshold used in poverty targeting practice and as to whether estimation or ordinal classification accuracy is prioritized.