Implementation of probability generated by scikit-learn on predicting the result of horse racing

I have built a simple training model using scikit-learn(a python package) to predict the champion in a horse race with over 15000 horse racing data entries. And turned out the best fit model is decision tree model. After I runned the predict_proba function on the horses of a coming race, it gave out the probabilities. I don't think it would simply be the chance of that particular horse winning the match. Should I do simple scaling like dividing each of the proba by the sum of them or there is a better way to deal with it?