Demand forecasting for a defined operational decision
A forecast is useful when its horizon, data and error measures fit the decision you need to make. OctOpus can work with time-series data to compare forecasting approaches and retain the code and results. Start with what you will order, staff or plan, rather than choosing a model before defining the forecasting problem.
Supply time, values and the series identity
Provide a date or timestamp and the quantity to forecast. For a panel, identify the product, store or other series. Explain sampling frequency, missing periods, stockouts and unusual events. External inputs such as promotions must be available for the forecast horizon if the model depends on them; future information should not enter training by accident.
Define the horizon and a useful baseline
Forecasting tomorrow, next week and next quarter are different tasks. Pick the horizon and update cadence, then compare candidate models against a simple reference appropriate to the series. OctOpus supports experiments across statistical, tabular, deep and foundation-model approaches where the task and configured runtime allow them. The best approach is determined by observed evaluation, not a universal family ranking.
Preserve time order in evaluation
Inspect the chronological training and evaluation periods and whether the test mimics the information available at forecast time. Compare errors by horizon and series, not only a pooled score. Select measures suited to the business: percentage errors can become misleading around zero demand. Backtests and uncertainty estimates should be inspected when the chosen workflow produces them, rather than assumed from a chart.
Put forecasts into the planning workflow
Review predictions alongside actual observations in a dashboard, or build an app that exposes the supported forecast action. Keep predicted and observed values clearly labeled. Verify the model’s serving inputs and response, data freshness and any external features before operational use. A demand forecast alone does not solve replenishment constraints or establish an optimal purchasing policy.
Key capabilities
- Explicit horizon, frequency and series identifiers.
- A baseline compared with supported candidate models.
- Chronological evaluation with inspectable code and results.
- Forecasts carried into dashboards and usable applications.