RLDDE: A novel reinforcement learning-based dimension and delay estimator for neural networks in time series prediction

Time series prediction is traditionally handled by linear models such as autoregressive and moving-average. However they are unable to adequately deal with the non-linearity in the data. Neural networks are non-linear models that are suitable to handle the non-linearity in time series. When designing a neural network for prediction, two critical factors that affect the performance of the neural network predictor should be considered; they are namely: (1) the input dimension, and (2) the time delay. The former is the number of delayed values for prediction, while the latter is the time interval between two data. Prediction accuracy can be improved using suitable input dimension and time delay. A novel method, called reinforcement learning-based dimension and delay estimator (RLDDE), is proposed in this paper to simultaneously determine the input dimension and time delay. RLDDE is a meta-learner that tries to learn the selection policy of the dimension and delay under different distribution of the data. Two benchmarked datasets with different noise levels and one stock price are used to show the effectiveness of the proposed RLDDE together with the benchmarking against other methods.