Explain the concept and challenges of time series imputation.
Instruction: Discuss the strategies for imputing missing values in time series data, including the challenges faced and how to overcome them.
Updated
Example Answer
Time-series imputation estimates missing observations while respecting their timing and the purpose of the analysis. I would first distinguish a retrospective reconstruction from a forecast or online decision. Linear interpolation between two known endpoints can be reasonable for reconstructing an old gap, but using the later endpoint would leak future information at an earlier prediction time.
For online use, I would consider past-only methods, such as a bounded last-observation carry-forward or a filtering model, and record the missingness and age of the last observation. A missing sensor reading is not automatically zero. I would fit imputation settings within each training fold and validate them using artificial gaps that resemble real outage lengths and patterns. Longer gaps and informative missingness need special care. More elaborate imputation is not automatically more accurate, and filled values should not be treated as equally certain as measurements.
Worked example
At 10:05, a value observed at 10:10 is unavailable. Interpolating a 10:05 feature from 10:00 and 10:10 is valid only for a retrospective task that permits both endpoints.
Why this works
It makes the information boundary explicit and evaluates methods against realistic gaps rather than convenience.