A large multilingual model is pre-trained on a massive corpus of text from over 100 different languages. Importantly, the pre-training process only uses monolingual documents; the model never sees parallel sentences (e.g., an English sentence and its direct French translation) during this phase. After pre-training, the model is fine-tuned for a sentiment analysis task using only English-language data. Surprisingly, when this fine-tuned model is tested on German-language reviews, it performs significantly better than random chance. Analyze and explain the key mechanisms and properties of the multilingual pre-training process that enable this successful cross-lingual transfer, despite the absence of explicit cross-lingual training data.