The Impact of Data Imputation on the Forecast Quality of SARIMAX Models in Studying the Ground Temperature of the Permafrost

Continuous monitoring of permafrost ground temperature underpins geocryological and climate assessments, yet in practice it is hampered by data gaps: long borehole observation series are regularly interrupted by equipment failures, scheduled maintenance, and changes in measurement methodology. The autonomy of logger systems is a particularly acute problem – a limited battery resource, memory faults, and data loss at extremely low temperatures produce both isolated and multi-year gaps that reduce the suitability of the series for analysis and forecasting. To address these issues, this study compares imputation methods for monthly permafrost ground-temperature series and evaluates how imputation affects SARIMAX forecast accuracy. The source data are observations from research boreholes at the Chabyda site (Central Yakutia) for 1982–2022. The methodology consists of two parts. First, eleven imputation methods are compared by their ability to reconstruct artificially masked observations under a bootstrap-masking procedure, with accuracy assessed using the MAE and RMSE metrics. Second, it analyzes how filling an artificially imposed fraction of missing values affects the SARIMAX temperature forecast. For permafrost temperature monitoring series it is shown for the first time that the imputation method best at reconstructing missing values does not necessarily provide the best subsequent forecast. SARIMAX-based imputation yields the lowest reconstruction error for masked values; however, the lowest imputation error does not always lead to the lowest forecast error: over the forecast horizon, simpler methods that reconstruct exogenous series by interpolating residuals prove preferable. The final forecast accuracy depends nonlinearly on the choice of imputation method and is associated with the fraction of missing data and the SARIMAX specification procedure. The main conclusion is that the imputation method should be chosen according to the ultimate goal – reconstruction quality or forecast quality – rather than by the reconstruction error alone.