Forecasting on Trading: A Parameter Adaptive Framework Based on Q-Iearning

TL;DR: FTQL method proposes a new design of learning mechanism by considering the performance of trading strategy based on the forecasting results and asset allocation strategy, which proves its effectiveness and universality on parameter adaption in the volatile stock market.

Abstract: How to build and improve forecasting models for stock prices are hot issues in both finance and computer fields in recent years Under the framework of reinforcement learning, the design of learning mechanism is important This paper presents a novel forecasting on trading Q-Iearning parameter adaptive forecasting framework (FTQL) Instead of learning from the forecasting errors directly, FTQL method proposes a new design of learning mechanism by considering the performance of trading strategy based on the forecasting results and asset allocation strategy FTQL firstly extracts the features from closing prices of both the underlying stock and its associated asset and set up the logistic regression (LR) models to provide directional forecasting Then, generate the investment signals according to the predictive results and apply an asset allocation strategy to execute the trading signals, for example, a variant of Uniform Constant Rebalanced Portfolios (UCRP) strategy Finally, Q-Iearning, the key point of FTQL, selects the best parameter combination of the signal generation models every day by taking the best actions according to the trained Q matrix FTQL proves its effectiveness and universality on parameter adaption in the volatile stock market