StatsForecast

This catalog had no forecasting cell. Observe paths (cost, load, demand, case volume, token spend) still need non-LLM predictors before an agent proposes action. StatsForecast is the first pack in Agentic → Forecasting: production-grade statistical forecasts and anomaly flags that can feed PM4Py-style monitoring or sleep/orchestration planners — with humans/gates still owning irreversible moves.

Operators pass a long-format panel (unique ids, timestamps, values). StatsForecast fits one or more models per series, then predict(h=…, level=[…]) returns point forecasts and optional intervals. Cross-validation and anomaly helpers reuse in-sample intervals. Distributed backends parallelize over many series.

1. Shape history as Nixtla long DF (unique_id, ds, y).
2. Choose models (AutoARIMA, SeasonalNaive, …).
3. fit then predict horizon h (optional level for intervals).
4. Push forecasts into dashboards, gates, or agent tools as proposals, not authority.

Related: pm4py · abm-agent-behavior-mining · topics/04-data · topics/14-process-mining · topics/09-self-monitoring

• Use when: you need scalable statistical forecasts (many series, intervals, exogenous vars) as observe/propose inputs for agents or ops.
• Skip when: the signal is unstructured text/docs (use RAG), or you need deep neural forecasters only — look at Nixtla NeuralForecast as a peer outside this pack.
• Prefer instead: PM4Py for process conformance on event logs; LLM judges for language quality — not for numeric forecast baselines.

• Stationarity / regime shifts: Auto* models still fail when the data-generating process jumps; measure CV on your series.
• Not causal: forecasts ≠ interventions; PresPM-style action still needs policy and outcome logging.
• Scope: StatsForecast does not store company truth; bad upstream meters produce confident wrong charts.
• Peer stack: NeuralForecast / HierarchicalForecast are separate Nixtla packages — do not pretend this pack covers them.