Search papers, labs, and topics across Lattice.
This study investigates the predictability of childbirth within three years using a comprehensive dataset from Dutch residents aged 18-45, involving 147 researchers employing various predictive methods. Despite utilizing advanced models like large language models and transformers, the best performance was achieved with traditional logistic regression, revealing that even with rich data, the accuracy of predictions was moderate (best F1 scores of 0.59 for registers and 0.76 for surveys). The research highlights a predictive ceiling influenced by stochastic biological factors, suggesting that inherent randomness in reproduction significantly limits our ability to forecast individual life events accurately.
Even with extensive data and advanced algorithms, predicting childbirth remains frustratingly imprecise, with chance playing a larger role than expected.
Major life events have proven difficult to predict. Does this reflect limits of theory, data, and algorithms, or the large role of chance? We examine one outcome - having a child within three years - through a near-ideal setting for prediction: a data challenge where 147 researchers predicted births for Dutch residents aged 18-45, using survey data and full-population registers. Methods ranged from logistic regression to a large language model and transformers. Predictions were moderately accurate (best F1: register 0.59, survey 0.76); advanced models did not outperform classical ones; and the larger registers did not beat the survey. Simulating the stochastic biology of conception and pregnancy, we estimated a predictive ceiling (survey F1 ~ 0.86-0.94, register 0.88-0.96). Observed performance falls short of this ceiling, implicating imperfect data, methods, and unmodelled chance, while the ceiling itself shows that chance in reproduction alone sets a non-trivial limit on predicting individual lives.
