Everything looked fine to me except for the inclusion of Polymarket data. I place a large bet on something I don't think is going to happen so the suckers come in to buy in on the other side. For example, I buy YES on an event I don't even think is going to happen. This sudden increase in demand raises the price of YES shares and signals to other traders that the event might be more likely to happen. As others see the price increase and fear missing out, they buy more YES shares. The price continues to rise. I sell off my YES shares gradually at the inflated price, offloading them to new buyers who now believe YES will win. Once my position is fully exited, I’ve locked in a profit, despite believing the event won’t occur.
You can check the Halawi paper, which does that using historical data. You’d need to pull out a split of questions based on their expected log probabilities. Not trivial . Probably using a betting market as ground truth could make sense.
@@TrelisResearch If we had, say, 15,000 hardware repair/replacement records for a certain niche, would you be able to fine tune the llm for forecasting these issues or is it best to utilize a forecasting model?
@@user-jk9zr3sc5h if you have that many records you're almost guaranteed better results using a for-purpose forecasting model. Something like AWS Sagemaker could help you train that up by submitting a spreadsheet.
Everything looked fine to me except for the inclusion of Polymarket data. I place a large bet on something I don't think is going to happen so the suckers come in to buy in on the other side. For example, I buy YES on an event I don't even think is going to happen. This sudden increase in demand raises the price of YES shares and signals to other traders that the event might be more likely to happen. As others see the price increase and fear missing out, they buy more YES shares. The price continues to rise. I sell off my YES shares gradually at the inflated price, offloading them to new buyers who now believe YES will win. Once my position is fully exited, I’ve locked in a profit, despite believing the event won’t occur.
Fascinating! Thank you!
Thank you for sharing!
very good! thanks
Would there be a fine tuning method to improve forecasting?
You can check the Halawi paper, which does that using historical data.
You’d need to pull out a split of questions based on their expected log probabilities. Not trivial . Probably using a betting market as ground truth could make sense.
@@TrelisResearch If we had, say, 15,000 hardware repair/replacement records for a certain niche, would you be able to fine tune the llm for forecasting these issues or is it best to utilize a forecasting model?
@@user-jk9zr3sc5h i'd be happy to help with that
@@user-jk9zr3sc5h if you have that many records you're almost guaranteed better results using a for-purpose forecasting model. Something like AWS Sagemaker could help you train that up by submitting a spreadsheet.
Best.
Thanks