Target Depth & Granularity Considerations
Customers often ask critical questions about SensibleAI Forecast related to choosing the optimal target depth, time granularity, and whether to forecast at the unit or dollar level. Another common inquiry is whether SensibleAI Forecast outputs should serve as root inputs for driver-based forecasts or whether forecasts should be generated directly at the final driver-based output level. Here are some key considerations to help guide these important decisions:
Data Sparsity Considerations
Data sparsity often occurs when forecasting targets become too granular in either dimensionality (e.g., SKU level) or frequency (daily, weekly, monthly). For instance:
-
At a Product Category level, data may consistently appear across all periods (daily, weekly, monthly), providing robust coverage for accurate modeling.
-
At the SKU level, however, data might be significantly sparser, with sales, shipments, or demand recorded only in approximately 25% of the periods (e.g., an SKU that sells roughly every fourth day, week, or month).
While there's no universal sparsity threshold defining when data becomes too sparse for accurate forecasting, increased sparsity is generally associated with lower forecasting accuracy. It's critical to recognize that data sparsity alone doesn't fully determine forecasting viability—sparse datasets can still capture meaningful trends or cycles—but it remains an essential factor in deciding your forecasting granularity.
Accuracy of Financial and Operational Drivers
When integrating unit-based forecasts into driver-based planning, it's crucial to identify the appropriate granularity that allows for accurate driver application. Consider the following:
-
Forecasting at a deeper granularity (e.g., SKU level) and aggregating upward might yield higher accuracy, particularly if unit prices or costs vary significantly within broader categories.
-
Conversely, forecasting at a higher-level granularity (e.g., Department level) and allocating downward can sometimes suffice if your drivers exhibit minimal variation across lower-level targets.
Example scenario:
A company’s product hierarchy is structured as follows:
Department > Product Category > Product > SKU
When generating a P&L forecast using unit-based forecasts, applying an average price per unit at the Department level might result in inaccuracies due to substantial price variation within departments. Conversely, applying average prices at the SKU level provides greater accuracy since SKU-level prices tend to remain consistent.
In summary, deeper granularity typically enables more precise integration with financial or operational drivers. However, when drivers at lower granularity levels are uncertain or based on rough estimates, it might be more effective to forecast directly for final outputs, such as revenue or required labor hours, rather than using unit-based forecasts as intermediaries.
Managing Target Turnover
Targets that frequently start up or shut down (e.g., annual SKU turnovers or product model changes) can negatively impact the accuracy and stability of your forecasts. These targets often fail to accumulate sufficient data for ML models to produce reliable predictions before becoming obsolete.
A practical guideline is to forecast at a granularity that ensures an average target lifetime of at least five years. This approach helps stabilize your forecasts by reducing the noise created by frequent product or target turnovers.
Summary Recommendations
-
Monitor Data Sparsity: Aim for granularity that balances data completeness and forecast accuracy.
-
Match Granularity to Driver Accuracy: Choose forecasting depth based on how accurately you can apply financial or operational drivers.
-
Minimize Target Turnover: Forecast at a granularity that avoids frequent introduction and retirement of targets, maintaining at least a five-year average lifespan per target.
By carefully evaluating these considerations, you can effectively leverage SensibleAI Forecast to produce accurate, actionable forecasts tailored to your organization's strategic and operational needs.