Dynamic Learning and Market Making in Spread Betting Markets with Informed Bettors

The spread betting market is a prevalent form of prediction market. In the spread betting market, participants bet on the outcome of a certain future event. The market maker quotes cutoff lines as "prices," and bettors take sides on whether the event outcome exceeds the quoted spread lines. We study how the market maker should move the spread lines to maximize profit. In our model, anonymous bettors with heterogeneous strategic behavior and information levels participate in the market. The market maker has limited information on the event outcome distribution. She aims to extract information from the market's responses to her spread lines (i.e., "learning") while guarding against an informed bettor's strategic manipulation (i.e., "bluff-proofing"). In terms of effective policies to adjust the market maker's spread lines, we show that Bayesian policies (BPs) that ignore bluffing are typically vulnerable to the informed bettor's strategic manipulation. To be more precise, the regret for the market maker is linear in the number of bets, and we identify certain strategies of the informed bettor that are profitable. We also show that the poor performance of BPs in our setting is not due to incomplete learning: when the informed bettor is absent in our setting, many simple policies eventually learn the event outcome distribution and achieve a bounded regret. Full Paper: https://ssrn.com/abstract=3283392