petm, I fail to see the possibility for a different interpretation. The model predicts point 50. When 5 days later new data comes in, the model makes another prediction for 50 days ahead, let's call that point 55.
After this new data on day 5, the model does not predict the value for day 50 anymore, only the value for day 55. Day 50 value is shown on the graph not as a prediction of the path to day 55, but as a recording of the day 50 prediction made 5 days ago.
Let's say the new data between day 0.and day 5 hints at a massive slowdown of the melting season. The model would generate a day 55 prediction that is much higher than it would have been otherwise. This by necessity requires the whole path to day 55 to be higher. The day 50 prediction is not valid anymore, and if somehow the model would be tweaked to produce predictions for 45 days ahead using the same methodology, the day 50 prediction made on day 5 would have been much different, much higher, than the day 50 prediction made on day 0. Because of the new data.
In other words, when day 55 prediction is made, all previous predictions are no longer valid, as they are based on old data. Therefore the curve of all past predictions is not the predicted curve the ice will take to get to the last prediction 50 days ahead of today. No matter how you slice it and dice it, the model does not predict the path. Not because it's a good or bad model, but because it simply does not predict the path, regardless of model skill.