That's a really misleading graph, for a couple of reasons:
1. It conflates the forecast with the now-cast. The former is a prediction of what will happen on Election Day (whether that day is six months away or one day away). The latter is a prediction of what would happen if the election were held that day. In theory, those two would converge to the same value on the day of the election, though in practice, that wouldn't actually happen due to computational differences.
2. Silver never says that he "should be judged" by [just] the final result. In fact, he's gone out of his way to say that. The problem is, that's the only point on the graph where we can compare a model (predicted value) to the actual (observed value). There isn't an election on any of the other days, so even if both agreed to look at the now-cast and use that as grounds to evaluate the model, we still would only have one datapoint which is nonzero in both dimensions. In other words, we have a blue line, yes, but we only have a red dot. Taleb wants to extrapolate the red dot into a horizontal blue line, and then use that to judge Silver's model, which is ridiculous.
Your second point is kind of confused. You don't need the prediction to coincide with the event, to judge the quality of the prediction. If I predict the elections for 2 years time then you can still judge whether I got it right or wrong. If I do that for multiple elections for multiple cycles, you can decide whether I'm a good forecaster.
If I then make another prediction that's always 1 year out, you can decide how good that prediction is and start to make a curve of how good my predictions are by how far out they are.
Taleb is saying that the blue line is so incorrect it's meaningless and deceptive and Nate should product a red line which is a real probability that he can actually be judged on.
1. It conflates the forecast with the now-cast. The former is a prediction of what will happen on Election Day (whether that day is six months away or one day away). The latter is a prediction of what would happen if the election were held that day. In theory, those two would converge to the same value on the day of the election, though in practice, that wouldn't actually happen due to computational differences.
2. Silver never says that he "should be judged" by [just] the final result. In fact, he's gone out of his way to say that. The problem is, that's the only point on the graph where we can compare a model (predicted value) to the actual (observed value). There isn't an election on any of the other days, so even if both agreed to look at the now-cast and use that as grounds to evaluate the model, we still would only have one datapoint which is nonzero in both dimensions. In other words, we have a blue line, yes, but we only have a red dot. Taleb wants to extrapolate the red dot into a horizontal blue line, and then use that to judge Silver's model, which is ridiculous.