Target encoder and leakage

[RostislavStoyanov]

Hello,

I've been reading the Chronos-2 paper and also went through the code. I have a question regarding the TargetEncoder. From my understanding you are calculating encoded values based on both historic and future values(if such are present) for the respective category. Wouldn't this lead to target leakage, which would be especially prevalent when there are multiple instances of the same category in the same patch?

[abdulfatir]

@RostislavStoyanov I presume your concern is related to fine-tuning and not zero-shot inference. Could you point to specific portions of the code?