Thanks for writing this, it clarifies a lot. I'm especially fascinated by Chronos-2's ability to leverage similar series via in-context learning. Does the model infer similarity implicitly from the time series data itself, or does it require some pre-defind feature space for comparin series?
Thank you for reading my work! :) No, you just feed the data, and Chronos-2 does the job automatically! The only thing you have to do is enable in-context learning and use the multivariate setting.
If the time series are similar or have causal relationships (e.g., items in a store or financial time series with similar properties), in-context learning will give you a performance boost!
Thanks for writing this, it clarifies a lot. I'm especially fascinated by Chronos-2's ability to leverage similar series via in-context learning. Does the model infer similarity implicitly from the time series data itself, or does it require some pre-defind feature space for comparin series?
Thank you for reading my work! :) No, you just feed the data, and Chronos-2 does the job automatically! The only thing you have to do is enable in-context learning and use the multivariate setting.
If the time series are similar or have causal relationships (e.g., items in a store or financial time series with similar properties), in-context learning will give you a performance boost!