Forecasting mixed frequencies time series with missingness and its application to the domain of stock price prediction

Abstract

Handling missing values is crucial for accurately forecasting time series with different sampling rates. In stock price prediction, for example, the daily stock prices and quarterly valuation figures are sampled at a different rate, and both are useful in estimating the daily stock price’s future. This research proposes combining imputation methods and an additional feature that characterizes the missingness, specifically combining imputation methods that provide different information and a feature that specifies the time since the last observation. It proposes a methodology for producing forecasts based on long short-term memory network model on time series with missing values. This study demonstrates how applying this methodology can overcome challenges faced by classical imputation methods. Methods face challenges, such as underestimating or overestimating the actual signal when the forecasting task requires different imputed information. The research also showcases the influence on forecasting of using imputation methods with different behaviour. Through extensive empirical evaluations on several stock datasets with different features, it is shown that this methodology is more robust and performs better than classical approaches with various types of data.