FAQ: Scheduling Events and Seasonality

Do recurring events that are the same day year over year need to be explicitly defined as scheduling events in the system, or can the ML model automatically learn and account for these patterns?: Reoccurring events with a significant impact on performance should be explicitly defined as Scheduling Events. While our machine learning model can learn from patterns in the data, defining these high-impact days provides a clear signal that they are not just another data point. This ensures the model gives them the proper weight and doesn't treat them as part of a general trend.
Defining events like Black Friday, Cyber Monday, or other business-specific holidays as Scheduling Events tells the model to place a higher weight on comparative historical data for that specific date. This allows the model to learn from these unique, year-over-year patterns instead of misinterpreting them as an anomaly. By explicitly identifying these days, you are essentially telling the model, "Pay special attention to this date, as its historical performance is a key indicator for this year's behavior."

When deciding what to include, we recommend focusing only on events that you are confident will have a huge impact on the metric you are tracking. It's generally better to err on the side of caution and include fewer events rather than more. If you are unsure whether an event will have a large enough impact, we recommend holding off on creating it for this year and observing the behavior. This approach helps prevent the model from over-indexing on less significant events and ensures the highest level of predictive accuracy.
How does the ML model differentiate between a true recurring scheduling event and a one-time anomaly that happened to fall on the same day in a previous year?: The machine learning model differentiates between recurring events and one-time anomalies by analyzing multiple data points and historical patterns over time.
For events specifically configured in the scheduling events feature, the model is trained to give specific attention to activity on those days. This explicit configuration acts as a strong signal, telling the model that these are true recurring patterns it should prioritize.

In contrast, while the model can identify anomalies—like a one-time event that happened on the same day last year—it won't treat them differently than any other data point unless they are specifically configured within the scheduling events feature. The model learns from all historical data, but only the pre-configured events receive special consideration in its analysis.
For events that have a variable date but are known in advance (e.g., Easter, Chinese New Year), how should these be handled? Do they need to be defined as scheduling events?: Business-relevant holidays that have a large anticipated impact should be entered as a Scheduling Event, including the holidays that fall on the same date every year. Additionally, we recommend adding up to 7 days before and / or after the event if that holiday has an impact on days leading up to or following it.
Is there a negative impact (Example: degraded accuracy, increased processing time, model confusion) to defining a scheduling event that ultimately has no discernible impact on the forecast metric, or one whose impact is currently unknown? Is it safer to define all potential events, or only those with a clear expected influence?: To ensure the best performance from our machine learning model, we recommend a more selective approach to adding scheduling events.
When you create a scheduling event, you're signaling to the model that the business performance on that specific date is unique and more significant than the normal, day-to-day activity. Including events of unknown or minor impact can cause the model to hyper-fixate on those dates. This reduces its ability to accurately identify broader, long-term patterns and trends in your business data.

To maximize the model's predictive power, only include events that you know will have a significant, out-of-the-ordinary impact on your business. By providing a curated list of high-impact events, you'll enable the model to more reliably detect meaningful correlations and generate a more robust and accurate forecast.
If my business sees significant changes for a season year over year (e.g., between Thanksgiving and Christmas), do I need to define Scheduling Events for that entire period?: The ML model already factors in year-over-year performance for a month, so a season of this duration does not need to be specified. We do recommend creating scheduling events for Thanksgiving and Christmas if those specific dates are particularly influential to your business.
Can I create events that are only considered for specific business metrics?: No, scheduling events will be considered for all metrics in your business.