Imputation Missing Stock Prices Using Generative Adversarial Networks and Attention Mechanism

Missing value in time series data can disrupt the analysis of historical data, especially stock market data, which requires the sequence of data points. This research proposes a method for imputation of missing stock prices in multivariate time series. The Generative Adversarial Networks (GAN) method, such as Generative Adversarial Imputation Networks (GAIN), and attention mechanism fill the missing value. The simulated missing pattern as dataset training is produced using the missing completely at random (MCAR) method. The GAN method learns the missing value pattern using generator and discriminator parts and attention mechanism, which are self-attention and multi-head attention. The proposed method was tested using various lookbacks, missing rates, hint rates, and evaluation methods on 100 stock issuers. The GAIN multi-head attention significantly outperforms baseline GAIN and GAIN self-attention methods.