Haven't had time to process it fully yet, but my initial thoughts are:

1. There should be a commit and reveal process. You are already complicating the actions required of curators so that is not for regular people anymore. Also the time periods involved are short. So it is not too much to ask of them to first make their commitment to the blockchain, and then when the first stage is over make the reveal in an appropriate time. If they do not reveal in time, their prediction is automatically considered wrong. Using commit-and-reveal means less information is available to other curators and I think this closeness penalty thing becomes less necessary.

2. The closeness penalty seems flawed and even if it was fixed I dislike it compared to something else, like commit-and-reveal. It seems like an attempt to prevent copy cat predictors, something commit-and-reveal would handle better. The formula provided CP = (1 - abs(EP-P) / EP)^2 means there is no penalty if someone votes exactly the same as someone else. Even if you meant CP = 1 - max(0, (1 - abs(EP-P) / EP))^2, it still seems like sacrificial sockpuppet bots could still easily ruin the prediction market system for others at virtually no cost.

3. The time is so short. I understand that votes, which determine the outcomes, are coming in early and so you want the curators to make their predictions before that. I think instead that each curator's reward should be based on what additional information they provide to the blockchain beyond what is already known from voters. So as of the time a curator makes their hidden prediction, the amount of abs_rshares voted for the post up to that point is recorded. Then, after the reveal stage is over, that number is divided by the total abs_rshares that ended up voting for the post, which gives the fraction that is used in modulating the curator's share of the curation rewards. Someone who makes a very accurate prediction after nearly everyone has voted should get nearly 0 rewards because such a prediction is easy to do and adds no value. Someone who makes a close prediction well before most votes arrived gets rewarded far more. This mechanism would still mean that curators would need to vote early to get the most value out of their predictions. But the time scale would no longer be hard-coded but instead dependent on how quickly the people who would eventually vote on the post actually do vote on the post relative to its creation time. So a post may be submitted at a time when the people who would most appreciate the post are sleeping. And perhaps those awake are not aware of the fact that people who will appreciate and upvote the post soon are still sleeping. Then when one of these people wakes up, they can see the post (which may have been submitted hours earlier), recognize the value, notice that there are currently both few votes on it and few predictions, and make the decision to make a prediction based on their knowledge that their peers in their timezone waking up will very likely be upvoting it soon. Or some post that is days old with few votes may suddenly become relevant because of some world event. Those who are quick to make predictions on that post before the votes start coming in can still make a nice profit if their prediction turns out to be true.

4. None of this addresses the issue that without a reward for voting (not curating, but voting), there is no financial incentive for voters to upvote other people's post (that they believe will be valued by the community) rather than just upvoting their own posts. The superlinear payout for voting rewards of a post means that voters can earn more money upvoting the successful posts (which ideally would be the good posts) rather than just voting their own posts to try to get a share of the rewards stream.

I'll add more thoughts and possible other recommendations after I think about the proposal more deeply.