One way to implement this would be to
-> Find pair of markets that present very high abs correlation (e.g. "Trump will be president in 2024", "Biden ... 2024")
-> If abs(p_yes_trump - p_no_Biden) > threshold, do action

We should then define 1. threshold and 2. action.
-> For threshold, we should probably investigate a few examples.
-> For action, could you comment on the optimal ways to act in such markets?

My suggestion for implementation:
-> Fetch markets from polymarket using their new API (note that gnosis/prediction-market-agent-tooling#459 is a larger task which will be partially but not fully tackled here)
-> The markets above contain only title, description (all are open since we want current prices), volume, outcomes and outcomePrices.
-> Use Langchain's InMemoryVectorStore to find markets that more closely match a given question. Pick top-K.
-> We pass the top-K markets (title, outcomePrices, outcomes) to an LLM and ask for it for an estimate. We ask that it considers markets with higher volume as having a higher likelihood to be correct.
-> We then place a bet based on this probability.
-> We should also test out a few markets, e.g. US election, Elon, etc.

My suggestion for implementation:

Isn't that exactly what Think Thoroughly is doing already? Except TT is fetching markets from Omen itself. It seems like if we add a new flag to the TT agent to only use the correlated markets, this task is done.

Or maybe TT's get_correlated_markets could be upgraded to also fetch from Polymarket (and when we are at it, why not Manifold as well?)

It seems like if we add a new flag to the TT agent to only use the correlated markets, this task is done.

Indeed TT can be used - so this agent could inherit from that and cast a new prediction based only on markets and not scenarios.
@evangriffiths any thoughts?

From my understanding of arbitrage (I didn't really know what it was before entering the crypto world 😅), what you're describing here isn't it.

Arbitrage is the practice of buying and selling the same or similar asset in different markets to profit from price differences... Arbitrage can be beneficial because it can lead to riskless profits...

What you're proposing seems to be "using similar markets to inform the prediction for another market, so we can make a more accurate prediction", which I think is different.

o1 is smarter than me so it can explain it with a simple example:

So the task would just be finding such cases of these markets. To start, this could be simply on the same platform (i.e. both A and B are on Omen/Polymarket. But in the future we could look for arbitrage opportunities across platforms (A is on Omen, B is on Polymarket, and the agent has to bridge tokens/have 2 wallets)