Alert prediction uses seasonality analysis to identify recurring alert behavior and surface likely future alert patterns. The new policy UI focuses on policy state, seasonality window, and optional filtering controls.
This policy works as an early-warning layer by learning recurring alert behavior over time and highlighting expected future occurrences.
Access and purpose
Use this policy to proactively detect expected recurring alert patterns and reduce reaction time.
Go to Setup > Account > Alert Policies. Select Alert Prediction and click Create New or + Add.
Permissions typically required:
- OpsQ View to view policies.
- OpsQ Manage to create, edit, delete, and change state.
Policy states
Choose one state:
- Enabled: Policy is active and prediction logic runs.
- Disabled: Policy remains saved but inactive.
In this UI, only these two states are available for alert prediction policies.
General details
| Field | Description |
|---|---|
| Policy Name | Required. Use a descriptive name that reflects prediction intent, such as Weekly Backup Pattern Prediction. |
Seasonality timeframe
Choose the historical observation window used for seasonality detection.
| Field | Description |
|---|---|
| Seasonality Timeframe | Select one of the available windows: 7 Days, 10 Days, 30 Days, 60 Days, or 90 Days. In the current UI, 7 Days is preselected by default. |
| Generate prediction alerts for new patterns found | Checkbox control. When selected, prediction alerts are created automatically for newly discovered patterns. |
Guidance:
- Use shorter windows (7-10 days) for daily recurring alerts.
- Use longer windows (30-90 days) for weekly or monthly trends.
Important behavior notes from existing platform behavior:
- Even when selecting 60 or 90 days, seasonality analysis can use a platform-defined effective maximum training horizon.
- Changing timeframe may require retraining to apply the new learning window.
When this setting is disabled, pattern discovery can still occur, but automatic alert creation is suppressed.
If auto-generation is disabled, prediction alerts can still be created manually from predicted alert views where available.
Policy filter
Use filters to scope which alerts are included in prediction training and evaluation.
| Field | Description |
|---|---|
| Resource Filter | Optional selector to limit prediction to specific resources or resource groups. Use the edit icon to open resource filter criteria. |
| Filter Criteria | Optional query conditions to include only selected alert subsets in training and prediction. |
The resource filter criteria dialog uses a query builder with + Query and enables Apply only after query conditions are provided.
Keep filters broad unless you have a clear data segmentation goal. Very narrow filters can reduce usable learning data.
Data and ML readiness considerations
Prediction quality depends on enough historical alert repetition. Policies typically move through ML states such as:
- Awaiting Data
- Training Queued
- Training Initiated/Started
- Training In Progress
- Ready
- Error
Recommended operating pattern:
- Create policy with a realistic timeframe.
- Allow training to reach Ready.
- Review predicted alerts and seasonal groups.
- Adjust filters/timeframe and retrain if results are too broad or too narrow.
Common exclusions and eligibility used in seasonality behavior:
- Maintenance-window alert activity is generally excluded from seasonality learning.
- Inference-type alerts are generally excluded from prediction training.
- Active alert-state filters may apply depending on platform configuration.
Save behavior
- Click Save to create the policy.
- Save remains disabled until required values are entered.
In this UI, policy creation is in a single panel and there are no additional policy-rule sections for prediction beyond seasonality and filters.
After saving, use Train Now from the policy list (when available) to start or refresh model training immediately.
Validation checklist
Before saving, confirm:
- Policy state matches rollout intent.
- Timeframe reflects expected recurrence cycle.
- Automatic prediction alert generation is intentionally enabled or disabled.
- Filter criteria, if used, still leave enough data for reliable pattern detection.