In principle, NN can do well predicting future outcomes of events given history of previous similar events. If it weren't for conditionals, which they can approximate with sharp Tanh or other sigmoid functions, NN are Turning complete; meaning, I think they can compute any computable question. However, you might find better solutions with other algorithms, as in algorithms that run faster. I suggest you read some of the articles on fivethirtyeight.com about their methodology: https://fivethirtyeight.com/tag/methodology/. You might also consider the wikipedia article on the Elo rating system. You might find better algorithms.
More directly to your question at hand if you stick with the NN. If you have a universe of 2 players and you always put the winner of matches in the first input, I don't think it will learn that the first input is the winner if you train it right. If you train it on two matches with opposite results, it should return 50% odds for either player. If you feel paranoid about the question, you can include who won with the input and then randomize the order. You will probably also need to include the score. You could try {Away, Home, AwayScore, HomeScore, <other match data>} as your data input and W/L as your output.
I think it can be done. But I won't recommend it for starting. I would start with the tutorial on LeNet: https://reference.wolfram.com/language/tutorial/NeuralNetworksOverview.html Then from there, you might consider the numeric regression problem, "Predicting from Numeric Data." Then from there, maybe turn back to tennis matches.
