About

I study physics at METU. For the past year most of my time has gone into prediction markets: the probability models, the infrastructure that feeds them, and the execution layer that acts on them.

I work in a fixed order. First I write down how I think the market works, in words, before any code. Then the Markov decision process that falls out of that description. Then I fill the chains with probability calculations from data I trust. An MDP is to me what a free-body diagram is in mechanics.

Two habits do most of the work. The first is deriving before fitting: when I needed a fair value for a binary contract on a Brownian price, the probit form is what the assumption produces, not something I looked up. Every parameter you fit instead of derive is somewhere a bad assumption can hide. The second is looking for the window where my own model has to be wrong, and running that test before the one that would confirm it.

There is no machine learning in any of this. The statistics are hand-rolled on numpy. A Kalman filter I can debug is worth more to me than a library I cannot.

I build with Claude Code, usually several agents at once: one writing an estimator, one trying to break it, one keeping the research log honest. It is how one person runs a five-machine stack and still has time to think about the market. The split is fixed. The market philosophy, the hypotheses, the axioms I am willing to assume, and every call about what to trade or retire are mine. What I hand off is implementation, cross-checking and documentation. It also makes falsification cheaper, and asking for the test that would kill your own result is easier when you are not the one who has to write it.

The results that made me look worse are on this site too. I retired a weather strategy after a scan of 32 elimination events put every entry band at negative expected value. A market-making study I spent two months on concluded that the market is efficient at the level I can reach. A football model bought contracts on teams that were not playing, and I had been ranking opportunities in a way that made that inevitable.

What I am missing is scale, and people who have been doing this longer to argue with.