It depends on the type of the missing data: MAR, MNAR, or MCAR.

https://en.wikipedia.org/wiki/Missing_data

If it is MCAR, you can use out-of-sample MAE or AvgRelMAE to measure the accuracy of imputation techniques, more details about measures for accuracy evaluation can be found here: Chapter Forecast Error Measures: Critical Review and Practical Recommendations

But here more details about your data are needed, preferably, with examples. There are many alternative measures and procedures that depend on the following features: 1) number of time series, 2) number of observation within each time series, 3) reasons for missing data, 4) the distribution of the number of adjacent missing values, 5) domain of actuals.

Generally, you can compare alternative methods for imputation using the same measures as those used for accuracy comparisons. Here you can use this framework: Conference Paper Data Formats and Visual Tools for Forecast Evaluation in Cyb...

It depends on the type of the missing data: MAR, MNAR, or MCAR.

https://en.wikipedia.org/wiki/Missing_data

If it is MCAR, you can use out-of-sample MAE or AvgRelMAE to measure the accuracy of imputation techniques, more details about measures for accuracy evaluation can be found here: Chapter Forecast Error Measures: Critical Review and Practical Recommendations

But here more details about your data are needed, preferably, with examples. There are many alternative measures and procedures that depend on the following features: 1) number of time series, 2) number of observation within each time series, 3) reasons for missing data, 4) the distribution of the number of adjacent missing values, 5) domain of actuals.

Generally, you can compare alternative methods for imputation using the same measures as those used for accuracy comparisons. Here you can use this framework: Conference Paper Data Formats and Visual Tools for Forecast Evaluation in Cyb...

Dear Andrey Davydenko

Thank you for your respond.

Dear Hasrul,

I completely agree with Andrey Davydenko , with a few supporting points. In respect of multiple imputation models, the usual assumption is that the data are missing at random (MAR) anchored on the assumptions of ignorability/exogeneity (a feature of an experiment design whereby a data collection method does not depend on missing data, nonconfounding, or absence of omitted variable bias which is standard in most analysis models. MAR is much safer than the more restrictive missing completely at random (MCAR) assumption which is required for listwise deletion.

Best method is artificial neural networks

Hi dear Hasrul, I think it depends on the type of data collected: if the time-series is linear or non-linear. For the first one, there are many methods, the main one based on the Box-Jenkins methodology. In the nonlinear case, one must be careful with haotic type data. Please, do not confuse with random data...

Dear Hasrul Hazman Hasan

You can select the method based on the shape of time series curve ( before and after missing data).