forecast_with_covariates() may have the problem of test data leakage

[zl123411]

Hi TimesFM team, when I was conducting tests using the method you provided, forecast_with_covariates(), I found that when there was a sequence relationship among the time series in the test set, xreg would leak the context information of the subsequent time series to the previous one，thereby making the prediction results of the previous time series exceptionally good

[zl123411]

please check "master/v1/notebooks/covariates.ipynb"

[siriuz42]

That's a good point. As of now, I can only think of batch_size=1 inference as a quick fix. In hindsight we may need to timestamp thresholding what can go into the XReg.