I built a system that reads local sports news in 40+ languages to catch weakened squads before the line moves. AMA about the pipeline
For the last few months I've been building a pipeline that monitors local sports media in 100+ countries, ties every "X players out" fact to a specific upcoming fixture, and issues a signal with a concrete market before kickoff. The idea is simple: bookmakers are slow exactly where nobody watches, second and third divisions, women's and youth leagues. When a club in Peru doesn't pay wages and four starters refuse to play, the local press knows hours before the odds move.
Some honest numbers from the public archive (small samples, losses included): 61% winning bets over 36 settled, positive CLV, and 57% of bets beat the closing line. Every signal is archived with its delivery time and the final score, nothing gets edited after the fact.
The engineering bits that turned out hardest: multilingual extraction with anti-hallucination guards (the model can only pick from real quoted markets, it physically can't invent a number), entity resolution across 40 languages (homonym teams, women's and reserve squads are a minefield), and closing-line snapshots for CLV.
Biggest lesson so far: a strong news signal is not the same thing as a good bet. We only learned where the actual edge lives by replaying our own archive.
Happy to answer questions about the approach. I won't go into specific sources or the betting logic itself since that's the product, but everything else is fair game. Mods, if a link to the public archive is okay under rule 1, I'll drop it in a comment.