Sagemaker Canvas Time Series Forecasting

The start date of a SageMaker Canvas forecast is not configurable because it's inferred based on the end date of your input dataset. Specifically, the global end date - meaning that if you have item A whose data finishes at 2024-07 and item B whose data finishes at 2025-12, your forecasts should always start at 2026-01.

I haven't encountered the different behaviour you mentioned between single-prediction and batch/deployed prediction to be honest: Wonder if your single-prediction was using a Quick Build model, or if it was definitely the same input dataset & predictor version?

My most likely guess would be that there's some entry in your dataset which is making Canvas think the target data extends to 2025-12. This could happen if, for example, you're passing in additional time-varying features (like price, in-stock, etc) for the forward-looking 2024-08 to 2025-12 range... BUT some record somewhere has an erroneous value filled in the target column, instead of leaving it blank?

SM Canvas' forecasting models use SageMaker AutoML under the hood, so you could try checking out the docs on time-series forecasting in SageMaker AutoML for more hints on data preparation. You might also find this sample Python notebook a useful illustration of data preparation for forecasting in Canvas - but it's a little out of date as was designed to run on SageMaker Studio Classic, so the library versions might be a bit stale.