uniTS-MissRecoPred: framework for degradation and reconstruction of univariate time series with missing values to improve forecasting effectiveness

Missing data significantly hampers the analysis of time series in operational monitoring, environmental sensing, and many other domains. We present uniTS-MissRecoPred, an open-source Python library that standardizes reconstruction (imputation of missing values inside the observed horizon) and links it to forecasting (prediction beyond the training segment). For controlled experiments, the software introduces synthetic missingness under Missing Completely At Random (MCAR), Missing At Random (MAR), and Missing Not At Random (MNAR) mechanisms with configurable rates, including the possibility of adding new mechanisms; researchers may also supply their own series that already contain gaps so that the missingness pattern is dictated by the data rather than the built-in generators. The distribution integrates more than twenty reconstruction methods spanning simple imputation, interpolation, statistical modeling, and Stable Diffusion inpainting with time-series-to-image encodings, reports reconstruction error selectable in configuration, and shows the results in a dashboard. A downstream stage trains and evaluates eleven forecasting families (twelve selectable identifiers), reports prediction error on held-out horizons, and exposes paired dashboards; dashboards report both reconstruction fidelity and downstream error metrics under multiple missingness settings. Resource usage (time, CPU, RAM, GPU memory) is logged; reconstruction dashboards rank models with a weighted combined score that balances normalized error against normalized compute cost. Reproducibility is supported through versioned YAML configuration, a documented programmatic API, optional runtime or entry-point registration of new models without forking the repository, and worked examples on open datasets (industrial boiler outlet temperatures and vibration traces).