Data Augmentation for Deep Candlestick Learner

How to augment limited stock price data is an open problem in stock trend prediction. Most innovative data augmentation schemes adopted in image processing community cannot be used directly in time-series data. Although the traditional financial data simulation method can generate time-series data, there are some defects when considering the real-world market. For example, Monte Carlo simulation is one of the primary traditional tools applied extensively in financial engineering research, economics, and a wide array of other fields during the past four decades [1]. However, Monte Carlo simulation is ultimately a statistical model, which means it requires several assumptions. Those assumptions may be unrealistic and depends on the individual circumstances. There are three primary disadvantages as follows: 1. Monte Carlo simulations need distribution assumptions to built around a specific type of statistical distribution. If we use the right distribution assumption, the results are valid. However, if we use the wrong one then the results will be meaningless [2].