Forecast horizon calculations

Historical data helps PlanIQ work out what time intervals apply to the forecast horizon. PlanIQ calculates if forecast predictions are by the week, month, or year.  

How far users can predict the forecast horizon, comes mostly from the data provided, but also from the algorithm. A few algorithms predict for the same length as the historical data timeline. Others only predict for a third of the historical data timeline. For the former, even if the algorithm can predict a forecast horizon of 12 months, this is still capped based upon historical data. 

Example:

Based on historical data and the selected algorithm, PlanIQ can generate predictions weekly for a total period of 6 months. The total period is the forecast horizon. 

Even though you can have a long-term forecast, we recommend you set your horizon for a shorter time period. This maximizes accuracy and data quality. We recommend you set your forecast horizon to cover half the time frame included in your historical data. 

Example: For 24 months of history, we recommend you forecast 12 months ahead.

The Anaplan AutoML algorithm has properties and behaviors similar to neural network algorithms.

Amazon Ensemble has a forecast timeframe that's different from the other algorithms. It's one-fourth of all historical data, unless that dataset is greater than:

• 52 weeks, for weekly data
• 36 months, for monthly data (ARIMA, ETS, and Prophet algorithms)

Currently, Amazon Ensemble doesn't support daily data.

The Anaplan Prophet and MVLR algorithmss can forecast for up to 50% of the provided historical data. Example: With 24 months of history, you can forecast 12 months ahead.

The ARIMA and ETS algorithms let you forecast for an extended horizon. They span nearly the entire timeline of your historical data. The forecast horizon is equal to the length of historical data less one period. 

Examples:

• If you have 24 months of history, you can have a forecast of up to 23 months. 
• If you have 52 weeks of data, you can have a forecast of up to 51 weeks.

The neural network algorithm CNN-QR has specific requirements:

• You can only forecast up to one-fourth of your historical timeline.
• Your historical dataset must have at least 300 data points.

The ultimate maximum supported horizon depends on your historical data's time period and the number of data points it contains. 

Example:

• If you have 24 months of history, CNN-QR enables you to forecast up to 8 months, provided you have enough data points (a minimum of 300)
• If the dataset has a low number of items (below 300 data points), or it's too sparse, then the supported forecast horizons are shorter, say 4–6 months 

DeepAR+ is another neural network algorithm with specific requirements. This algorithm:

• Enables you to forecast up to a fourth of your historical timeline.
• Requires that the initial three quarters your historical timeline must have at least 300 data points.
  • Example: With a historical timeline of 36 months, your forecast horizon is up to 9 months. The first 24 months must include 300 data points.

There's another data limitation to keep in mind:

• It excludes items with very short history (shorter than one half of the historical timeline) from the Deep AR+ learning process. It then examines the remaining items (typically IDs or SKUs). The maximum forecast horizon will be one-third of the average timeline of the remaining items.
  • Examples: With a historical timeline of 36 months, the process excludes all items with a history shorter than 18 months. 
    • If the remaining items' average timeline is 36 months, the maximum supported horizon will be 9 months.
    • If the remaining items' average timeline is 21 months, the maximum supported horizon will be 7 months.