How do you handle missing data in time series forecasting pipelines at scale?

Missing data in time series is handled using interpolation, model-based imputation, or learned representations depending on the pattern and scale.

Missing values in time series can break temporal continuity and bias forecasting models. At scale, strategies include simple interpolation (linear, spline), statistical imputation (forward/backward fill), and model-based methods (Kalman smoothing, MICE, or deep generative models). Advanced systems explicitly model missingness using masking in RNNs or Transformers so the model learns to distinguish missing vs observed patterns.