Replacement Dataset

A replacement dataset is a modified version of the baseline related time series that contains only the values that you want to change in a what-if forecast. The replacement dataset must contain the forecast dimensions, item identifiers, and timestamps in the baseline related time series, as well as at least 1 changed time series. This dataset is merged with the baseline related time series to create a transformed dataset that is used for the what-if forecast. The replacement dataset must be in CSV format.

This dataset should not contain duplicate timestamps for the same time series.

What follows are several examples of how you can specify a replacement time series and how those specifications are interpreted. Consider the case where you are forecasting daily and the forecast horizon is 2022-08-01 through 2022-08-03. The baseline related time series for all examples is given in the following table.

item_id	timestamp	price	stock_count
item_1	2022-08-01	100	50
item_1	2022-08-02	100	50
item_1	2022-08-03	100	50
item_2	2022-08-01	75	500
item_2	2022-08-02	75	500
item_2	2022-08-03	75	500

Unchanged values

To apply a 10% discount on item_1 for 2022-08-02 and 2022-08-03, it is sufficient to specify the following for the replacement dataset:

Replacement dataset
item_id	timestamp	price
item_1	2022-08-02	90
item_1	2022-08-03	90

However, it's also valid to specify unchanged values in the replacement dataset. When used as replacement datasets, each of the following three tables will yield the same results as the previously provided table.

Replacement dataset with an unchanged column
item_id	timestamp	price	stock_count
item_1	2022-08-02	90	50
item_1	2022-08-03	90	50

Replacement dataset with unchanged rows
item_id	timestamp	price
item_1	2022-08-01	100
item_1	2022-08-02	90
item_1	2022-08-03	90
item_2	2022-08-01	75
item_2	2022-08-02	75
item_2	2022-08-03	75

Replacement dataset with unchanged rows and columns
item_id	timestamp	price	stock_count
item_1	2022-08-01	100	50
item_1	2022-08-02	90	50
item_1	2022-08-03	90	50
item_2	2022-08-01	75	500
item_2	2022-08-02	75	500
item_2	2022-08-03	75	500

Missing values

Missing values in the replacement time series are replaced with values from the baseline related time series. Consider the scenario where you apply a 10% discount on item_1 for 2022-08-02 and 2022-08-03 and increase the stock of item_2 on 2022-08-01. This replacement dataset is sufficient:

Replacement dataset with missing values
item_id	timestamp	price	stock_count
item_1	2022-08-02	90
item_1	2022-08-03	90
item_2	2022-08-01		5000

The values missing from this table are imputed from the baseline related time series.

Extraneous values

Extraneous values in the replacement time series are ignored when creating a what-if forecast. That is, values in the replacement dataset that do not correspond to values in the baseline related time series are not modeled. Consider this replacement dataset:

Replacement dataset with extraneous values
item_id	timestamp	price	stock_count
item_1	2022-08-01	100	50
item_1	2022-08-02	100	50
item_1	2022-08-03	100	50
item_2	2022-08-01	75	500
item_2	2022-08-02	75	500
item_2	2022-08-03	75	500
item_3	2022-08-01	50	125
item_3	2022-08-02	50	125
item_3	2022-08-03	50	125

The rows containing item_3 are ignored and are not part of the what-if analysis.

Historical changes

Changes in the replacement dataset that are outside of the forecast horizon are ignored. Consider this replacement dataset:

Replacement dataset with values outside the forecast horizon
item_id	timestamp	price	stock_count
item_1	2022-07-31	100	50
item_1	2022-08-01	100	50
item_1	2022-08-02	100	50
item_1	2022-08-03	100	50
item_1	2022-08-04	100	50
item_2	2022-07-31	75	500
item_2	2022-08-01	75	500
item_2	2022-08-02	75	500
item_2	2022-08-03	75	500
item_3	2022-08-04	75	500

The rows containing 2022-07-31 and 2022-08-04 are ignored and are not part of the what-if analysis.

Forecast dimensions

If you include forecast dimensions in your dataset, then you must include them in the replacement dataset. Consider this baseline related time series:

item_id	store_id	timestamp	price	stock_count
item_1	store_1	2022-08-01	100	50
item_1	store_1	2022-08-02	100	50
item_1	store_1	2022-08-03	100	50
item_1	store_2	2022-08-01	75	500
item_1	store_2	2022-08-02	75	500
item_1	store_2	2022-08-03	75	500

Therefore, the replacement dataset for a 10% discount in all stores on 2022-08-02 would be the following:

item_id	store_id	timestamp	price
item_1	store_1	2022-08-02	90
item_1	store_2	2022-08-02	67.5