Predicting the outcome of a greyhound race is an interesting problem, for two reasons.
First, unlike human sports such as football or basketball — where strategy and tactics play a major role — greyhound racing outcomes are driven mostly by each dog's physical condition. That makes a dog's past performance a much stronger predictor of future results than it would be in a tactics-heavy sport.
Second, the sheer volume of races run every day means there's an enormous amount of data available, which makes it statistically easy to evaluate how well a model actually performs.
Traditional approaches to greyhound race prediction rely on hand-crafted features. This model takes a different bet: no hand-crafted features at all. Instead, it learns its own features implicitly from each dog's raw past statistics.
The model is still being evaluated, so I can't share architecture details yet — but in short, it already produces very competitive predictions without any manual feature engineering or iteration.