Box Jenkins Neural Network Keynecian Reinforcement Learning Based Financial Big Data Analysis for Optimal Prediction

Abstract

In the digital era, Economic and Finance information is huge and attainable. The financial industry is a key to stimulating the evolution of the national economy, and with the immense availability of huge data makes the process of financial data analysis both a time consuming and tedious decision making in the financial market. By analyzing this Economic and Finance information significantly and easily, valuable perceptions can be provided. For rational economic growth, information pertaining to economic affairs and likelihoods is critical to decision-makers like governmental bodies, financial institutions and banks. However, precise predictions have been demand owing to the intricacy and unpredictability of financial and economic systems in the midst of persistent changes in economic environments. This work provides to approaches for better economic prediction and decision-making using a novel method called, Multivariate Box Jenkins Neural Network-based Keynecian Reinforcement Learning (MBJNN-KRL). In this work, Financial Data Analysis Deep Economic Prediction model that contains a Multivariate Box Jenkins Long Short Term Memory Forecasting model, for better economic prediction and convergence speed even in case of large data is first designed. The deep reinforcement learning technique is then adapted to retrain neural networks and rebalance the aggregate expenditure periodically. We combined the multivariate Multivariate Box Jenkins Long Short Term Memory Forecasting and the Keynecian Reinforcement Learning models into a single integrated reinforcement learning model named Multivariate Box Jenkins Neural Network-based Keynecian Reinforcement Learning (MBJNN-KRL). Finally, distinct sets of experiments are carried out on real large-scale data set, and the results entirely prove the efficiency and robustness of the proposed method in arbitrary economic prediction based on financial data. Our empirical data revealed that the MBJNN-KRL method could attain combative financial performance compared to traditional methods in terms of predictive accuracy, convergence speed and prediction error.

Recent advances in reinforcement learning in finance

“…However, it introduces extra overheads and leads to suboptimal regret. More recently, Wei and Luo (2021) propose a general approach that is applicable to various RL settings (including both episodic MDPs and infinite-horizon MDPs) and achieves optimal dynamic regret without any prior knowledge of the degree of nonstationarity.…”

Non-stationary Risk-Sensitive Reinforcement Learning: Near-Optimal Dynamic Regret, Adaptive Detection, and Separation Design

“…To achieve a nearly-optimal regret without the prior knowledge of the variation budget, (Auer, Gajane, and Ortner 2019) and (Chen et al 2019) maintain a distribution over bandit arms with properly controlled variance for all reward estimators. For RL problems, the seminar work (Wei and Luo 2021) proposes a black-box reduction approach that turns a certain RL algorithm with optimal regret in a (near-)stationary environment into another algorithm with optimal dynamic regret in a non-stationary environment. However, the above works only consider risk-neutral RL and may not apply to the more general risk-sensitive RL problems.…”

“…To overcome this limitation, we propose a meta-algorithm that adaptively detects the non-stationarity without the knowledge of B, while still achieving the similar dynamic regret as in Theorems 3.1 and 3.2. In particular, we generalize the black-box approach (Wei and Luo 2021) to the risk-sensitive RL setting and design a non-stationarity detection based on the exponential Bellman equations (3).…”

“…Non-stationarity detection on the exponential value functions. Different from non-stationarity detection for riskneutral RL (Wei and Luo 2021), we design non-stationarity detection mechanism for the exponential value functions (3) instead of the value functions (1) in Algorithm 3. This is because the non-linearity of the risk-sensitive value function makes it difficult to obtain its unbiased estimation, which is needed in the design of non-stationary detection mechanism.…”

“…(3) Furthermore, the non-linearity of the objective function (see Equation ( 1)) makes it difficult to obtain an unbiased estimation of the value function, which is needed in the design of a non-stationary detection mechanism in risk-neutral nonstationary RL (Wei and Luo 2021). (4) It is unclear whether the risk control and the handling of the non-stationarity can be separately designed when achieving the optimal dynamic regret.…”