Question: Temporal Continuity for Missing Targets at the end of a Series

[ryanon4]

In a real-world deployment of TabPFN-Time-Series for inference, it is likely that not all target values are available, such as when using a future "look-ahead" target definition — for example, where the target at timestep t is based on the percentage change between t and t+6 hours (e.g. energy consumption forecasting).

In such cases, if I prepare training data only up to timestamp t - 6 hours and then predict on a real-time test sample at timestamp t, there will be a 6-hour gap (missing target data) between the end of the training data and the test sample. Thus, to use TabPFN, we need to drop these rows with missing values from the end of the dataset.

Given this, I would like to confirm how this gap affects the model’s assumptions or prediction validity. Specifically:

• Does TabPFN-Time-Series require strict temporal continuity between training and test timestamps for valid predictions?
• During inference, does the model only use features from the current test timestamp’s row (in addition to the previously fitted training data), or does it rely on sequential patterns or features derived from prior timestamps, similar to how we would expect an RNN to behave?
• Would missing recent training timestamps (e.g., the last 6 hours) impact prediction quality solely due to less representative training data, but not cause functional issues during inference?

[liam-sbhoo]

Hi @ryanon4

Here's my two-cents:

Does TabPFN-Time-Series require strict temporal continuity between training and test timestamps for valid predictions?

Technically, tabpfn-ts doesn't assume temporal continuity in the data at all; the data doesn't even have to be regular. However, in the current implementation of RunningIndexFeature and AutoSeasonalFeature, it assumes the data to be regular and continuous. This can be easily fixed, I think.

During inference, does the model only use features from the current test timestamp’s row (in addition to the previously fitted training data), or does it rely on sequential patterns or features derived from prior timestamps, similar to how we would expect an RNN to behave?

It's the former case, i.e. the model only use features from the current test timestamp’s row (in addition to the previously fitted training data). Therefore, having NaN-values for the features in the test timestamps might hurt the performance.

Would missing recent training timestamps (e.g., the last 6 hours) impact prediction quality solely due to less representative training data, but not cause functional issues during inference?

Yup, exactly!