State Space Approach to Adaptive Fuzzy Modeling: Application to Financial Investment

This paper proposes a new state space approach to adaptive fuzzy modeling under the dynamic environment, where Bayesian filtering sequentially learns the model parameters including model structures themselves as state variables. In particular, our approach specifies the state transitions as meanreversion processes, which intends to incorporate and extend the established state-of-art learning techniques as follows: First, the mean-reversion levels of model parameters are determined by applying some existing learning method to a training period. Next, filtering implementation over test data enables on-line estimation of the parameters, where the estimates are adaptively tuned for each new data arrival based on the obtained reliable learning result. In this work, we concretely design a Takagi-Sugeno- Kang fuzzy model for financial investment, whose parameters follow autoregressive processes with the mean-reversion levels decided by particle swarm optimization. Since there exist Monte Carlo simulation-based algorithms called particle filtering, our methodology is applicable to a quite general setting including non-linearity, which actually arises in our investment problem. Then, an out-of-sample numerical experiment with security price data successfully demonstrates its effectiveness.