This paper examines the under/overreaction effect driven by sentiment belief in an artificial limit order market when agents are risk averse and arrive in the market with different time horizons. We employ agent-based modeling to build up an artificial stock market with order book and model a type of sentiment belief display over/underreaction by following a Bayesian learning scheme with a Markov regime switching between conservative bias and representative bias. Simulations show that when compared with classic noise belief without learning, sentiment belief gives rise to short-term intraday return predictability. In particular, under/overreaction trading strategies are profitable under sentiment beliefs, but not under noise belief. Moreover, we find that sentiment belief leads to significantly lower volatility, lower bid-ask spread, and larger order book depth near the best quotes but lower trading volume when compared with noise belief.
Temporal logistic neural Bag-of-Features for financial time series forecasting leveraging limit order book data