Predicting Multi-Party Vote Choice from Issue Batteries: A Calibrated Gradient-Boosting Approach Using German VAA Data

Vote Advice Applications (VAAs) generate vast traces of political attitude data but only a subset of users self-annotate their intended vote, limiting downstream analysis. Using a large German VAA dataset with 31 issue items, I evaluate whether attitudes alone can predict party choice and produce calibrated probabilities suitable for auto-labeling unannotated VAA records. I compare multiple modelling techniques including a model evaluation and a gated approach (left-wing/rightwing prediction before party prediction) to maximize their predictive power under a stratified 60/10/30 split. Models are tuned on macro-F1 and calibrated on the validation set (isotonic). LightGBM attains the best test performance and excellent probability calibration, enabling aggregation of probabilities to user-base vote shares. Error structure aligns with known cleavages and wrong predictions mostly deviate towards politically close parties. Furthermore, the gated approach has a slight advantage in its predictive power yet yields worse predcited probabilities. Methodologically, I show that calibrated gradient boosting over attitude items alone yields reliable multiclass vote probabilities.