You are here:
Forecasting
Use forecasting in Data 360 to predict future values from historical data. Forecasting models automatically identify trends and seasonal patterns to generate predictions without requiring data science expertise.
The default algorithm is Holt-Winters and it’s best suited for structured time-series data with clear trends and recurring seasonal patterns. For example, retail sales that peak each quarter or monthly revenue that grows steadily over time.
All forecasting models and algorithms use the new runtime.
Forecasting supports various use cases.
- Revenue forecasting: predict future revenue trends with historical performance data
- Demand planning: estimate product or service demand across time periods
- Sales trend analysis: identify growth or decline patterns in pipeline or bookings
- Capacity planning: forecast resource needs based on historical usage trends
- Seasonal pattern analysis: analyze recurring peaks and dips in business activity
Single-variate forecasting enables these use cases by using one historical variable. For example, forecast next month's revenue by using only past revenue data when a clearly defined metric exists.
How Forecasting Works
Forecasting models learn from historical time-series data—a sequence of values recorded at regular intervals, such as daily sales or monthly revenue. The model identifies patterns, such as trend direction and seasonality, and projects those patterns into the future to generate predictions.
Forecast outputs include prediction intervals that represent uncertainty in the forecast. The upper and lower bounds define the range of expected values, where narrower ranges indicate higher confidence.
Each forecasted record includes these outputs:
- Forecasted value—the predicted value for each future time period
- Forecast upper bound—the high end of the predicted range
- Forecast lower bound—the low end of the predicted range
Review these forecasting concepts.
Partition columns define how Data 360 groups data for forecasting. For example, when forecasting revenue by account, the account field is the partition column. Each model supports up to five partition columns, enabling segmented forecasting across multiple dimensions.
Time-series grouping defines the aggregation interval for the data, such as minute, day, month, or year. The model automatically infers forecasting frequency from this grouping, and you don't configure it separately.
Seasonality captures recurring patterns that repeat over a fixed interval, such as weekly traffic spikes or quarterly sales increases. Define seasonality manually for up to 24 periods.
Null handling imputes missing values in training data automatically with the built-in imputation method. The model uses imputed values only during training and doesn't write them to forecast outputs.
Considerations
Review these considerations.
-
Forecasting inference consumes the inference billing meter.
-
Batch transform forecasting also consumes the Data Transform—Batch meter based on rows processed.
-
If missing values aren't reliably imputed, forecasting inference stops.
- Use Holt Winters Forecasting
Predict future values based on historical patterns in a single variable by using the Holt-Winters algorithm. For example, use a data transform to forecast sales by account based on opportunity data.
