High-severity (14d)
—
Time-Series Intelligence
Continuous anomaly detection and forecasting across all key business metrics. Each metric uses the most accurate model (proven by 12-fold rolling-origin backtest). High-severity findings include LLM narration and a suggested action. The system learns from your feedback — useful patterns stay sensitive, noisy ones get suppressed automatically.
Medium (14d)
—
Forecasts active
—
Metrics tracked
—
Avg accuracy (MAPE)
—
Value caught
—
Loading anomalies…
Loading forecasts…
Loading accuracy…
How accuracy is measured: 12-fold rolling-origin backtest. For each fold the model is trained on a 90-day window and asked to predict the next 7 days; we compute MAPE (Mean Absolute Percentage Error). Lower is better. The chosen model per metric is the one with the lowest MAPE — every metric uses the model proven to be the most accurate for it.
Filter:
Loading dimensional anomalies…
Top catches by ₹ value
No catches recorded yet — mark anomalies "useful" with a saved-₹ value to populate.
Most accurate forecasts
Loading…
Forecast accuracy over time (period: daily)
What the system has learned from feedback
No learning yet — needs ≥ 3 feedback signals per (metric, detector, period). Mark anomalies useful/not useful to start training.
Forecast trend computed from intel_forecast_actuals: every prediction is scored on the day its target date arrives.
Learning thresholds live in intel_learning: detectors with high "not useful" rates have their alert thresholds raised automatically, so the system gets quieter and more on-target over time.
Loading run history…