What is the best way to handle missing data in a time-series forecasting model?

All answers to this question.

Never use mean imputation for time-series; it destroys the variance and ignores the temporal order. For short gaps (1-2 hours), "Forward Fill" or "Linear Interpolation" is usually sufficient. However, for larger gaps in seasonal data, you should use "Seasonal Decomposition of Time Series" (STL) or "Spline Interpolation." These methods account for the underlying trend and cycles. In 2024, I worked on a smart-grid project where we used Kalman Filters for imputation. It was remarkably effective because it predicts the missing value based on the state of the system, which kept our downstream forecasting model's error rate (MAPE) much lower than simple interpolation would have.

   Answered 2025-09-22 by Helen Adams

If the data is missing at random (MAR), can we just drop those rows, or will that break the "lag" calculations needed for models like ARIMA?

   Answered 2025-10-05 by Brian Scott

• Brian, you cannot drop rows in time-series! Most models like ARIMA or Prophet require a continuous, equidistant time index. Dropping a row creates a "jump" in time that makes the lag terms ($y_{t-1}$) mathematically incorrect. You must impute the values to keep the timeline intact. If the gap is too large to impute reliably, you might have to treat the data as two separate series or use a model like XGBoost that can handle sparse inputs better than traditional econometric models.

     Commented 2025-10-12 by George Baker

I’ve had great luck with the imputeTS package in R. It has a function called na_kalman that is incredibly accurate for sensor data gaps.

   Answered 2025-10-20 by Mary Nelson

• I second that, Mary! The imputeTS library is a lifesaver. It’s one of the few tools that handles both the trend and the seasonality during the imputation process effectively.

     Commented 2025-10-25 by Kevin Wright