FAQ: Data Granularity and Metric Definition

In what cases do customers need to define separate forecast metrics (e.g., splitting sales by country, region, department, product category, or even individual SKUs)? What are the criteria for recommending such splits?: Defining separate forecast metrics is necessary when the operational use or underlying demand patterns differ significantly across segments. The decision to split is a trade-off between the stability of aggregated data and the actionability of granular forecasts.
Customers should define separate forecast metrics in cases where the structure of the business data or the required use of the forecast demands a high level of detail.

Scenario
Example of Split
Justification

Operational Use Case Requires Detail
Splitting sales by department or product category.
Forecasting labor needs or inventory at a granular level. For example, a store needs separate forecasts for the fresh produce department versus the electronics department to accurately staff those areas or manage perishable stock.

Inconsistent Data Structures/Schedules
Splitting metrics for different organizational units.
If organizations have different start days of the week for their operational reporting, a separate metric is needed. Since weekly metrics can only have a single start day of week, scheduling organizations with a different start day of week might benefit from a separate metric to align forecasts with their reporting calendar.

The criteria for recommending a split focus on the need for detail and the quality of the resulting data.
Data Availability and Stability
Aggregation for Stability: As a general principle, aggregate data usually gives us better results than splitting the data because combining smaller data sets smooths out noise, outlier errors, and random fluctuations, leading to a more robust and reliable overall forecast.
Granularity Requires Richness: Separate, granular metrics (like splitting store-level metrics to be department-specific) are only helpful if those segments have their own rich data sources to ensure adequate amounts of data are available to make reliable forecasts. If the data set for the granular split is too small or sparse (e.g., a low-volume product category), the forecast will be unstable and potentially unreliable.
Actionability and Business Utility
Targeted Labor/Inventory: Separate metrics are essential when the resulting forecast needs to be directly actionable at a granular level. For example, splitting a store's total sales forecast into separate department forecasts is more helpful to target labor needs in specific departments and optimize staffing where it is needed most.
Alignment with Reporting/Planning: Splits are recommended when segments operate on fundamentally different planning cycles or data structures. For instance, a split is necessary if a segment uses a different day as the start day of week for its business reporting, ensuring the forecast aligns with operational reality.
What are the trade-offs (e.g., accuracy, computational cost, data requirements) of forecasting at a more granular level versus a more aggregated level?: When planning and forecasting, one of the most important decisions is the level of detail you choose—known as the "granularity." Forecasting at a detailed level (like every 15 minutes) provides high specificity but can introduce noise, while forecasting at a broad level (like per-day) offers stability but less direct utility. The right level depends entirely on your business use case.
This is a summary of the key trade-offs between highly granular and more aggregated forecasting approaches:

Feature
Granular Forecast (Example: 15 minute level)
Aggregated Forecast (Example: Daily level)

Actionability
High: Direct input for ordering, stocking, and staffing specific departments/items.
Low: Requires manual spread across the day or is split into even parts.

Forecast Accuracy (Statistical)
Lower Stability : Data is often sparse, volatile, and noisy. More susceptible to large, random errors.
Higher Stability : Noise, outliers, and random errors are smoothed out, leading to a more robust result.

Business Value
High Utility: Better at capturing unique, local, and segment-specific demand drivers.
High Reliability: Better for planning and long-term decisions.

Data Requirements
High: Massive volumes of high-resolution, row-level data to ensure there are enough observations to model the volatility of individual items or locations.
Low: Smaller data volumes since the act of summing data points naturally reduces the number of records, requiring only enough history to identify broad seasonal trends.
Are there minimum data history requirements (e.g., number of data points, time period) for a metric to be forecastable at a specific granularity?: Yes, the ability to generate a reliable forecast is directly linked to the amount and density of historical data provided. We have established requirements to ensure the data is sufficient for our models to detect meaningful patterns like trend and seasonality.
Our general recommendation for reliable forecasting is to have three years of historical data for the metric being analyzed. This period allows our models to effectively capture yearly seasonal cycles and long-term trends.

The number of required data points varies based on the frequency (granularity) of your data:

Granularity
Example Interval
Data Point Requirement
Rationale

High Density (Hourly)
Every hour
Three years is sufficient.
The model accumulates a large volume of data points quickly, making three years a robust dataset for pattern detection.

Low Density (Daily, Weekly)
Once per day or week
Three years is the minimum requirement.
This time period provides enough data points to identify core trends and seasonal variations necessary for generating a stable forecast.

We recommend a minimum of three years of data history to ensure optimal performance from our forecasting models, regardless of the data's granularity.

When introducing forecasting for entirely new contexts (e.g., a new country, a brand new product line, or a completely new business metric), what is the minimum amount of historical data (e.g., in terms of time periods like months/years, or number of data points) generally required before the ML model can generate valuable and reasonably accurate predictions?

When introducing forecasting to a truly new context, our approach relies on connecting the new scenario to existing, trained metrics to rapidly generate valuable predictions.

Note: The new location must be tracking an already trained metric (e.g., a metric for which a model already exists and is performing well in other locations) to use this functionality.

Forecasting using ML Model for New Locations (e.g., a New Country or Store): For metrics trained and forecasted using the ML model, when opening new locations we offer use of our sister organization" or analogous entity approach to jumpstart the forecasting process. By leveraging our internal sister organization feature, our model can make reasonable assumptions about the new location's performance until the point where it has accumulated at least two-weeks of its own data. After that point, the model can rely on the new store's own growing history for increasingly accurate predictions.

Forecasting using Trend-Based Forecasting Model for New Locations (e.g., a New Country or Store): For metrics trained using the Trend-based Forecasting Model, the sister organization data would be used until the new location has gathered a full one year of its own data to enable those specific historical comparisons.

Scenario	Example of Split	Justification
Operational Use Case Requires Detail	Splitting sales by department or product category.	Forecasting labor needs or inventory at a granular level. For example, a store needs separate forecasts for the fresh produce department versus the electronics department to accurately staff those areas or manage perishable stock.
Inconsistent Data Structures/Schedules	Splitting metrics for different organizational units.	If organizations have different start days of the week for their operational reporting, a separate metric is needed. Since weekly metrics can only have a single start day of week, scheduling organizations with a different start day of week might benefit from a separate metric to align forecasts with their reporting calendar.

Feature	Granular Forecast (Example: 15 minute level)	Aggregated Forecast (Example: Daily level)
Actionability	High: Direct input for ordering, stocking, and staffing specific departments/items.	Low: Requires manual spread across the day or is split into even parts.
Forecast Accuracy (Statistical)	Lower Stability : Data is often sparse, volatile, and noisy. More susceptible to large, random errors.	Higher Stability : Noise, outliers, and random errors are smoothed out, leading to a more robust result.
Business Value	High Utility: Better at capturing unique, local, and segment-specific demand drivers.	High Reliability: Better for planning and long-term decisions.
Data Requirements	High: Massive volumes of high-resolution, row-level data to ensure there are enough observations to model the volatility of individual items or locations.	Low: Smaller data volumes since the act of summing data points naturally reduces the number of records, requiring only enough history to identify broad seasonal trends.

Granularity	Example Interval	Data Point Requirement	Rationale
High Density (Hourly)	Every hour	Three years is sufficient.	The model accumulates a large volume of data points quickly, making three years a robust dataset for pattern detection.
Low Density (Daily, Weekly)	Once per day or week	Three years is the minimum requirement.	This time period provides enough data points to identify core trends and seasonal variations necessary for generating a stable forecast.