Online discussions often derail into toxic exchanges between participants.\nRecent efforts mostly focused on detecting antisocial behavior after the fact,\nby analyzing single comments in isolation. To provide more timely notice to\nhuman moderators, a system needs to preemptively detect that a conversation is\nheading towards derailment before it actually turns toxic. This means modeling\nderailment as an emerging property of a conversation rather than as an isolated\nutterance-level event.\n Forecasting emerging conversational properties, however, poses several\ninherent modeling challenges. First, since conversations are dynamic, a\nforecasting model needs to capture the flow of the discussion, rather than\nproperties of individual comments. Second, real conversations have an unknown\nhorizon: they can end or derail at any time; thus a practical forecasting model\nneeds to assess the risk in an online fashion, as the conversation develops. In\nthis work we introduce a conversational forecasting model that learns an\nunsupervised representation of conversational dynamics and exploits it to\npredict future derailment as the conversation develops. By applying this model\nto two new diverse datasets of online conversations with labels for antisocial\nevents, we show that it outperforms state-of-the-art systems at forecasting\nderailment.\n