Create your first forecast

Have these details ready:

• An experiment to store this run in, or access to create one.
• Past data in a place Forecast can read. This can be a database, object storage or a direct public file link. You can't upload a data file from your computer.
• The date or time column and the value you want to forecast.
• For more than one series, the columns that identify each series. For example, each store and product pair might form one series.
• The past date range, forecast end date and frequency, such as daily.
• Access details from your data owner. Keep passwords, connection strings and cloud keys private.

For this example, use one row per day. Include a date column and a sales_units column with numbers. Your data must cover the full training period.

If you use planned inputs such as prices or promotions, you'll also need their values for each forecast date. Ask your data owner or Forecast administrator if you need help with access or column names. See data requirements for more detail.

1. On Home, select New run.
2. Choose or create an Experiment for this run.
3. Under What do you want to do?, choose Train and predict. This trains models from past data and makes forecasts in one run.
4. In Historical data, choose the source type your data owner gave you. Enter its location and access details.
   • For a public file, choose Public URL and enter a direct link ending in .csv or .xlsx. A file-sharing page that needs sign-in won't work.
   • For an Excel workbook, enter its sheet name too.
   • For a database or object storage, use the details from your data owner.
5. Enter the column names. Match their spelling and capital letters exactly.
   • Set Timestamp column to date and Target columns to sales_units.
   • Leave Identifier columns blank for this single-series example. For more than one series, add the columns that identify each one.
6. In Forecast period, set the dates and frequency.
   • Set Training data duration to cover the past data the model should learn from.
   • Under Prediction data duration, set End date to the last date you want to forecast.
   • For daily data, choose a daily Forecast level. Set Forecast frequency multiple to 1.
   • Check that the forecast horizon shown matches the period you want.
7. In Models and metrics, select Apply AutoML defaults to get a starting set of models. Choose Mean absolute error as the comparison metric if it's available. This score helps compare models. Lower error is generally better.
8. Select Continue through the required steps. At Customize, leave optional areas unselected for this example. If you use known columns, follow the prompt to set up Future data.
9. On Review, check the experiment, source, columns, dates, frequency, models and metric. Enter a clear Run name, such as January daily sales. Then select Submit run.

If a field shows an error, fix it before you continue. Use the step list or Edit on Review to go back and change a choice. For data rules and settings, see Run reference.

After you submit, select View run details. The run keeps working in the background. A successful submission doesn't mean the run has finished.

Open Monitoring to track progress and check errors. When the run is completed, open Results. Use Plot to see the forecast pattern or Table to read the values. You can also review scores that compare the models.

Check the dates and values against what you know about your business. Review the data and assumptions as well as the chart and scores. For the next steps, see monitor a run and review results.

Use what you learn to improve your next run. If results look wrong, check the data, columns, dates and frequency with your data owner. You can start a new run with different settings. The existing run won't change.

To compare changes to planned inputs, start with a completed forecast that uses those input columns. Then follow compare scenarios. You can also share a run with people who need access.