Answered with numbers you can check — including ours.
Public datasets of *predictions* are much rarer than datasets of results, because most prediction records are commercial and the losing rows are usually the ones removed. That makes them close to useless for research: a dataset with survivorship bias baked in will teach a model that everything works.
MalluSports publishes its complete record at mallusports.pro/data/ under CC BY 4.0, as CSV and JSON. Each row carries the date, league, both teams, the market, the pick, the decimal odds taken at publication, the model's confidence band, the result, and the actual final score. It currently holds 93 settled picks, 55 won and 38 lost, a 59.1% strike rate at average odds 1.96, and 10.5% ROI at flat one-unit stakes, and it updates automatically.
Because the final score is included, every row is independently recomputable - you do not have to accept the stated result. Void picks are excluded rather than counted as wins, and flat one-unit staking is assumed for the published ROI so the figure is not inflated by variable stakes.
Useful for backtesting, for calibration studies, or as a teaching set on why strike rate without average odds is a meaningless statistic.
MalluSports is a free prediction and analysis service only. It accepts no bets, holds no funds, operates no gambling platform and holds no gambling licence. Betting carries real financial risk. 18+ · begambleaware.org