A novel imputation methodology for time series based on pattern sequence forecasting.

Area of Science:

• Data Science
• Statistical Modeling
• Time Series Analysis

Background:

• Conventional methods for handling missing time series data, such as mean imputation or last observation carried forward, often lack precision.
• Existing algorithms like Pattern Sequence Forecasting (PSF) excel at forecasting but do not inherently address missing value imputation.

Purpose of the Study:

• To introduce and evaluate a novel univariate time series imputation method, imputePSF.
• To enhance the Pattern Sequence Forecasting (PSF) algorithm for simultaneous forecasting and backcasting to impute missing values.

Main Methods:

• The imputePSF method modifies the PSF algorithm to identify and utilize repeating patterns in observed data.
• It characterizes periodic characteristics of existing observations to estimate missing values more precisely.
• Imputation accuracy was assessed by simulating missing data across three univariate datasets and comparing imputePSF against established imputation techniques.

Main Results:

• The imputePSF algorithm demonstrated a reduction in error estimates compared to conventional imputation methods.
• Simulations showed that imputePSF provides more precise estimates for missing values, particularly in datasets exhibiting periodic and repeating patterns.
• The method effectively balances forecasting and backcasting for accurate data imputation.

Conclusions:

• The imputePSF algorithm offers a more precise imputation method for univariate time series data.
• It is particularly advantageous for datasets with discernible periodic and repeating patterns.
• This enhanced PSF approach improves the handling of missing data in time series analysis.