What is time-series forecasting?
Forecasting supports inspection planning, demand analysis and sensor monitoring. It predicts a future sequence rather than one static spatial value.
Approaches and starting model
Statistical baselines and neural forecasting models should be compared on historical holdouts. Chronos-2 supports pretrained forecasting; useful results still require representative data and local evaluation.
Inputs and outputs
Time-indexed numerical histories with series identifiers and supported covariates enter the forecasting pipeline. Output is a future trajectory with quantiles for the requested forecast horizon.
Where it helps in GIS and digital twins
A candidate forecasting component for water-level, energy or environmental sensor dashboards. Link predictions to sensor/asset IDs, timestamps and units; retain forecast issue time separately from observation time.
How to start
Start with a zero-shot backtest using the documented pipeline. Compare against persistence and seasonal baselines before considering adaptation. Fine-tuning, if needed, should follow the version-matched repository guidance.
How to evaluate results
Use rolling temporal backtests, never a random time split. Evaluate error by horizon and interval coverage. Retain enough history to represent local seasonality and missing-data patterns.
Limitations
Forecasts cannot guarantee sudden events or account for information absent from the history. Irregular timestamps, changed units, leakage from future covariates and sensor drift invalidate comparisons.
Official starting-model documentation
The category artwork illustrates its starting model; it is not an evaluated prediction.