Leveraging Log Probabilities in Language Models to Forecast Future Events