To generate the most accurate forecasts, you'll need to consider a number of factors that can influence your data. Note that the machine learning models that are in place in the Forecast platform address all of these issues.
These influencing factors include:
-
Trends: Often data will follow a trend. In the example below, which shows the total cost of a customer, the orange part is the actual data and the blue part is our forecast forward. This example shows that costs are decreasing, so our forecast also needs to decrease.
- Seasonality: Data will often have seasonality, meaning its behavior may vary during the days of the week, on weekends and so on. If, for example, the data has weekly seasonality, this must also be reflected in the forecast.
- Special events: Special events and external factors may also influence the data. For example, you may see drops in costs due to beginning of month discounts, and increases during large scale experiments. Sales events, marketing campaigns, and deployments of software can all influence the data.
- Adaptivity: Occasionally there are unexpected changes. Cost reduction efforts may change the behaviour of the data, as well as new RI agreements, new customers and more. We can’t forecast this in advance, but we must respond quickly and adapt our forecasts quickly to the new normal.
- Robustness to anomalies: On the other hand, we don’t want to adapt to any transient anomaly that occurs, so the Forecast platform models must be robust to them.