It will affect the trading and worse it will affect the trading inconsistently so we can’t even use mathematics to subtract it
Nothing ever affects the trading consistently! It’s never the case that in an important market you can just use math to decide what to bet.
The resulting percentage cannot be used as a measure of trustworthiness nor as a measure of the underlying probability of event X.
Sure it can. If you ever see a prediction market which you don’t think is measuring the underlying probability of its event, you can make money from it (note that this is about manipulating whether the event will happen. Obviously if the market might be misresolved all bets are off). It’s provable that, no matter what manipulation or other meta-dependencies exist, there’s always some calibrated probability the market can settle at. If a manipulator has complete control whether an event will happen or not and will manipulate it to maximize its potential profit, the market settle at or fluctuate slightly around 50%, and in fact the event thus has a 50% chance of happening. If you give me any other manipulation scenarios, I can similarly show how the market will settle at a calibrated probability. Manipulation is bad because it creates bad incentives outside the market, and because it usually increases the entropy of events (bringing probabilities closer to 50%; this will always be the case if manipulation in either direction is equally cheap), but I don’t think it can threaten the calibration of a market.
But you can manipulate prediction markets in much more easy, mundane and legal ways.
I think my point generalizes. There’s a bunch of ways to manipulate stock prices. I assume they cause some problems, but use laws and norms to prevent the worst behavior and it ends up working pretty well. Prediction markets may face more of a problem, since I’d expect them to be easier to manipulate, but I don’t think there’s a qualitative difference.
Suffice to say, if you combine the fact that 1) humans can’t instantly know all the new information with 2) the fact we can’t know whether the market has updated because of information that we already know, new information, or people updating on the assumption that there is new information, with 3) recursive mindgames and 4) these constant ‘ripples’ of shifting uncertainties; you’ll get so much asynchronous noise that the prediction market becomes an unreliable source.
Sometimes reality is deeply unpredictable, and in those cases prediction markets won’t help. But if you think that a prediction market will be unreliable in cases where any other method is reliable, you can use that to get rich.
I think the core of what I’m trying to get across is that (modulo transaction costs), a prediction market is as reliable as any other method, and if it’s not you can correct it and/or get rich. Manipulation is bad because it changes the probability that the event happens, not because it makes prediction markets unreliable. Manipulation can make all methods of prediction work less well, it cannot make prediction markets work less well than another method.
There’s an important distinction here between prediction the next token in a piece of text and predicting the next action in a causal chain. If you have a computation that is represented by a causal graph, and you train a predictor to predict nodes conditional on previous nodes, then it’s true that the predictor won’t end up being able to do better than the original computational process. But text is not ordered that way! Texts often describe outcomes before describing the details of the events which generated them. If you train on texts like those, you get something more powerful than an imitator. If you train a good enough next-token predictor on chess games where the winner is mentioned before the list of moves, you can get superhuman play by prepending “This is a game which white/black wins:”. If you train a good enough next-token predictor on texts that have the outputs of circuits listed before the inputs, you get an NP-oracle. You’re almost certainly not going to get an NP-oracle from GPT-9, but that’s because of the limitations of the training processes and architectures of that this universe can support, it’s not a limitation of the loss function.