Handling records with no actual data in the target time series data

Question

We are using AWS Forecast to predict the price trends of a certain product. When preparing the target time series data, we are considering using past price records for each product. However, some products are only available for certain periods, resulting in days without price data. When making predictions in this situation, it feels like the forecasts are slightly lower than the actual prices. Is this because the missing data days were predicted as zeros? Additionally, how can we interpret the days with missing data not as zeros but as "days when the product is not available" for better prediction?

ceriddle · Answer

If you don't specify a filling logic then by default Amazon Forecast will fill any missing values with zeros.  To change this default behavior you need to perform missing value filling, by specifying the types of filling to implement when you call the `CreatePredictor` operation.  Filling logic is specified in `FeaturizationMethod` objects using JSON like the example below:
`[
    {
        "AttributeName": "target_value",
        "FeaturizationPipeline": [
            {
                "FeaturizationMethodName": "filling",
                "FeaturizationMethodParameters": {
                    "aggregation": "sum",
                    "middlefill": "zero",
                    "backfill": "nan"
                }
            }
        ]
    },
    {
        "AttributeName": "price",
        "FeaturizationPipeline": [
            {
                "FeaturizationMethodName": "filling",
                "FeaturizationMethodParameters": {
                    "middlefill": "median",
                    "backfill": "value",
                    "backfill_value": "2",
                    "futurefill": "max"               
                    }
            }
        ]
    }
]`

Your target time series data supports the following filling methods:
* "zero" - 0 filling.
* "value" - an integer or float number.
* "nan" - not a number.
* "mean" - the mean value from the data series.
* "median" - the median value from the data series.
* "min" - the minimum value from the data series.
* "max" - the maximum value from the data series.

You can also reference the documentation here for more information - https://docs.aws.amazon.com/forecast/latest/dg/howitworks-missing-values.html

Handling records with no actual data in the target time series data

Conteúdo relevante