@sarahshine007 I concurred that.
Strongest model should aim to beat baseline by 5-10% but not "predict it right every time". Still you might be able to gain profit in place like polymarket and kalshi. e.g. you buying a market with implied possibility of 20% but your model says 35%.
Predict-Raven is an open-source forecasting agent framework.
We just shipped a sprint update for the World Cup: champion, group stage results and knockouts
Yes — it predicts 🇪🇸 Spain to win it all, with reasoning traces and sources fully public. DM me for free trial access 🎆
Victor, I agree with you in some sense. There are indeed many classical ML method to predict sport outcome. In predict-raven, we use elo rating with bayesian updates to predict. https://t.co/4U3ygOUYnE
In a sense, this website give us a base elo rating. while LLM use news (player status or team style) and other factors to update elo.
I also wanna highlight that coding agent can run any ML with sufficient data. Future forecasting harness will be a hierarchy that LLM call ML model as methods.
There's a few instances of scientific heroics I've seen from seniors on Gemini where, given some issue in the stack (say, model behavior or training dynamics), they find and prove out a fix that seems quite distant from the original bug but works well and is elegant
I think that's a hallmark of research skill at the very top: the space of solutions for some modeling issue is often very vast, and making 🚀 progress relies on some ability -- hard-fought intuition over many years; System 2 transformed to System 1, I imagine -- to both know the full vastness of the search space, and then bisect with speed and high signal at each step to arrive at an answer
@patpcj like ur work. I will try to get inspiration and integrate feature from this fantastic work into my opensourced forecasting agent project, to provide better search