Inside Astock, the Trading Framework That Turns Financial Language into Returns

Can News Really Beat the Market?

Most people agree that news moves markets. The harder question is whether machines can read that news well enough to trade on it before the edge disappears. For years, that promise has hovered over quantitative finance like a permanent headline. Countless papers have tried to turn headlines, tweets, and sentiment into alpha, yet many of them stop at a shallow benchmark: prediction accuracy. They may tell you whether a model guessed the next move correctly, but they rarely tell you whether that prediction can survive the brutality of real trading.

That is what makes the Astock study interesting. Instead of treating stock prediction as a clean NLP classroom exercise, the paper pushes the problem closer to market reality. The authors build a China A-shares dataset with stock-specific news, minute-level prices at the time of publication, and 24 stock factors per company. More importantly, they evaluate the final system not only by classification metrics such as accuracy and F1 score, but also by annualized return, maximum drawdown, and Sharpe ratio in simulated trading.