From: Nikhil Gupta <mywork.ng@gmail.com>

I conveyed this feedback to @Max Mergenthaler during our preliminary conversation a few weeks ago and would like to discuss it with the broader team.

I stumbled upon the requirement that historical data contain no missing values when using TimeGPT. However, it's common to encounter some gaps in practical datasets. This poses a challenge, especially for users unfamiliar with time series forecasting techniques (which is likely a large target audience for Nixtla), as they might struggle to handle these missing values appropriately.

For instance, employing a simple forward-fill method to address missing values often yields subpar forecasting results. This is particularly evident when recent data points, which TimeGPT tends to prioritize, are missing. Consequently, the forecast tends to be understated and fails to capture seasonal peaks effectively (see naive imputation below).

To address this issue, I experimented with an iterative imputation approach. This involved first generating an in-sample forecast using TimeGPT, followed by using this forecast to impute missing values and finally leveraging this imputed dataset for forecasting. The outcome demonstrated significant improvement compared to the initial attempt.

Incorporating this functionality as a built-in option in the API would benefit both users and Nixtla. Firstly, it would alleviate concerns for less experienced users, who would no longer need to select appropriate imputation methods. Secondly, Nixtla could offer this enhanced capability at a premium, potentially doubling the token charge.

I'm open to discussing this further if you'd like.