@martincasais.eth

every trader claims the news moves the price. i spent my final year at university testing whether that's actually true. my thesis: an ML pipeline that takes media sentiment + market indicators and predicts whether solana closes up or down the next day. random forest vs a performer transformer, each trained with and without sentiment. defended it last week. the honest results after months of work: sentiment carries some signal. it ranked among the most important features. but it didn't consistently improve predictions. its value depends entirely on the model you feed it to. my favorite finding was a trap: the model with the best F1 looked like a winner on paper. the confusion matrix told the real story. it was predicting UP almost every single day. great metric, useless model. daily crypto direction stays close to a coin flip. anyone selling you certainty is selling the narrative, not the data. so today this cast is three things: 1. the end of my degree 2. my full thesis, linked below (fair warning: it's in spanish) 3. the starting line. smart contract security + applied ML on crypto data is next full pipeline in the repo. https://github.com/usctrabajo/TFG-sentimiento-solana/blob/main/tfg.pdf https://github.com/usctrabajo/TFG-sentimiento-solana

0 reply

0 recast

0 reaction