Handling data gaps in time series using imputation

Time series forecasts depend on sensors or measurements made in the real, messy world. The sensors flake out, get turned off, disconnect, and otherwise conspire to cause missing signals. Signals that may tell you what tomorrow's temperature will be or what your blood glucose levels are before bed. Alfred Whitehead and Clare Jeon explore methods for handling data gaps and when to consider which.

Time series forecasting is everywhere. It tells you what tomorrow’s temperature will be, your company’s stock price on Friday, and your blood glucose levels before bed. Often these forecasts depend on sensors or measurements made out in the real, messy world. Those sensors flake out, get turned off, disconnect, and otherwise conspire to cause missing data in your signals. Alfred Whitehead and Clare Jeon explore a number of methods for handling data gaps and advise you on which to consider and when. You’ll see how to perform tests to determine which method suits your problem the best. And all of this is illustrated with real data from a continuous blood glucose monitor. The methods they handle include random assignment, average-based imputation, last observed carried forward, linear interpolation, spline interpolation, moving average, Kalman smoothing with structural model, Kalman smoothing with auto-ARIMA model, Stineman interpolation, k-nearest neighbors, and seasonality with Prophet.