Hi @mfenton, in Amazon QuickSight, when you create a machine learning (ML) forecast, the model used for forecasting is based on the entire dataset available, not just the subset of data displayed in the visualization. This means that even if you are only displaying the last 60 days of sales data in your line chart, the ML forecast will still consider all 365 days of data provided in your dataset to build the model and make predictions.
Here’s a breakdown of how this works:
Data Utilization in ML Forecasting
- Training Data: The ML forecast model in QuickSight uses the full extent of the dataset available to train the model. This includes learning patterns, trends, seasonality, and other relevant characteristics from the entire dataset, which in your case, is the last 365 days of daily sales data.
- Model Application: After the model is trained on the full dataset, it is capable of making forecasts for future periods. This is where your 30-day forecast comes into play.
Visualization Filter
- Filter Impact: The filter you apply to only display the last 60 days of data affects only the visualization aspect and not the model training process. The ML model does not “see” this filter; it only affects what portion of the data (and the forecast) you are displaying in your visual.
- Forecasting Window: The 30-day forecast is generated based on the model trained on the entire 365 days of data and is appended to the end of the dataset or the last known data point.
Best Practices
- Data Refresh and Re-training: If your dataset updates regularly (e.g., new sales data being added daily), consider re-training your ML model periodically to incorporate the most recent data, which can help in maintaining the accuracy of the forecasts.
- Model Evaluation: Evaluate the performance of your ML forecast by checking how well past forecasts have matched up to actual observed values. This can give you insight into whether your model’s training on the full dataset effectively captures the trends and seasonality in your sales data.
By understanding these aspects, you can better utilize QuickSight’s ML forecasting capabilities to make informed decisions based on comprehensive data analysis rather than just the visible subset in the visualization.
More information can be found here.
Did this solution work for you? I am marking this reply as, “Solution,” but let us know if this is not resolved. Thanks for posting your questions on the QuickSight Community!