Deep Learning Based Approaches for Imputation of Time Series Models

Abstract

Market price forecasting models for Fresh Produce (FP) are crucial to protect retailers and consumers from highly priced FP. However, utilizing the data for forecasting is obstructed by the occurrence of missing values. Therefore, it is imperative to develop models to determine the value for those missing instances thereby enabling effective forecasting. Usually this problem is tackled with conventional methods that introduce bias into the system which in turn results in unreliable forecasting results. Therefore, in this thesis, numerous imputation models are developed alongside a framework enabling the user to impute any time series data with the optimal models. This thesis also develops novel forecasting models which are used as a gauging mechanism for each tested imputation mode. However, those forecasting models can also be used as standalone models.

The growth and success of deep learning has largely been attributed to the availability of big data and high end computational power along with the theoretical advancement . In this thesis, multiple deep learning models are built for imputing the missing values and also for forecasting. The data used in building these deep learning models comprise California weather data, California strawberry yield, California strawberry farm-gate prices, USA corn yield data, Brent oil type daily prices and a synthetic time series dataset. For imputation, mean squared error is used as an metric to gauge the performance of imputation whereas for forecasting a new aggregated error measure (AGM) is proposed in this thesis which combines mean absolute error, mean squared error and R2 which is the coefficient of determination.

Different models are found to be optimal for different time series. These models are illustrated in the recommendation framework developed in the thesis. Different stacking ensemble techniques such as voting regressor and stacking ML ensemble are then utilized to have better imputation results. The experiments show that the voting regressor yields the best imputation results. To gauge the robustness of the imputation framework, different time series are assessed. The imputed data is used for forecasting and the forecasting results are compared with market deep and non-deep learning models. The results show the best imputation models recommended based on work with the synthesized datasets are in fact the best for the tested incomplete real datasets with Mean Absolute Scaled Error (MASE) <1 i.e. better than the naive forecasting model. Also, it is found that the best imputation models have higher impact on reducing the forecasting errors compared to other deep or non-deep imputation models found in literature and market.