We present a statistical model which uses data on National Football Le
ague games and betting lines to study how agents learn from past outco
mes and to test market efficiency. Using Kalman Filter estimation, we
show that teams' abilities exhibit substantial week-to-week variation
during the season. This provides an ideal environment in which to stud
y how agents learn from past information. While we do not find strong
evidence of market inefficiency, we are able to make several observati
ons on market learning. In particular, agents have more difficulty lea
rning from ''noisy'' observations and appear to weight recent observat
ions less than our statistical model suggests is optimal.