For time-series forecast, I did 2 prong approaches. Here is what I think about pros/cons
use AWS Forecast
Managed service partially/fully.
Quite slow speed of training as it might try many algo to find the best prediction.
When AutoML is on, you would not know which Algo you’re using. It’s all hidden.
NO model/endpoint is exposed. I can’t query myself via REST api, nor I can’t integrate Forecast to QuickSight easily.
Forecast can export the results to csv file(usually multiple) . I need to merge all to ONE file myself (via Glue → Athena)
But, it does support BackTesting (only via SDK / JupyterLab , not console) , I can export and charting.
SageMaker studio jupyter lab - DeepAR
DeepAR seems to be the pick for time-series but I had to manipulate DataFrame quite a lot to produce the exact input format.
I successfully created a model and endpoint.
A lot faster speed of training.
I don’t know how to integrate this model in QS side as I can’t easily find the example.
**Question : How do I integrate DeepAR model to QuickSight dataset ? **
**As you may know, DeepAR takes only 2 mandatory columns from the training. ( timestamp, series of values) (I don’t use cat, etc) ** I don’t know what to do with schema file here.
When integrating with sageMaker, Quicksight offers you to upload a schema file to map inputs from the dataset to the input params to be used by the model for prediction.
Using this mapping when the data is ingested into SPICE, model is inferred and the outputs are returned to QuickSight and are made available for you to be able to visualize.
An example of this integration can be found in this blog.
Additionally, just wanted to let you know that QuickSight offers out of the box forecasting with RCF with backwards cast for accuracy tests. have you tried that yet. link
I am afraid that you overlooked the ‘timeseries model’ part.
My timeseries model only takes 3 columns which are timestamp and uniqueid, demand column.
So, in schema file, I only defined 2 as timestamp is not supported.
My goal is not to create a new column of future perediction. I’d like to use the existing column , demand. But, it doesn’t let me.
Anyway, I just added a new column in the schema file, selected that column. I saved the ‘Augment with SageMaker’. But, it doesn’t look like the model is making a prediction automatically.
The example link is to show classification problem. I’d like to know how to integrate TimeSeries model(DeepAR) into QS side. I do have model and endpoint. I just can’t figure out how to use this Model and connect/predict.
Hello @tbdori, are you still facing issues integrating the SageMaker DeepAR time series model with QuickSight dashboards? If you found a solution, do you mind posting the steps you took to complete this? Otherwise, I did find this post about someone else trying to use time series data. The AWS expert that responded linked this documentation so if you haven’t resolved the issue it may help you as well!
