I’ve tuned my Bayesian model a bit. Previously, it used the cumulative sum of historical results point spread as its data input, now it uses each individual game spread. Perhaps an example can make this clearer:
Consider two teams, A and B, who have met 5 times in the past, with following results:
The above series gives a cumulative spread of +2, which normalized by number of games becomes 2/5 = 0.4 for team A, indicating that team A is more likely to win. However, that result is highly influence by the 10-0 result of the first game, while not putting enough weight on the fact that the four last games team A actually lost.
To cope with that, I changed the model, instead of considering the cumulative spread in its calculations, to consider each historical outcome individually, which in the above example would result in team B being regarded as the winner.
Below the new predicted outcomes and corresponding odds.