Forecast two uncertainties
Demand and lead time are separate probability laws, not single best guesses.
- Expand requirements through bills of materials.
- Keep missing observations distinct from observed zeros.
- Condition remaining lead time on the age of an open order.
Compare models on matched historical cases
Forecast quality is checked separately from purchasing cost. Demand and lead-time analytics compare models on common eligible observations.
- Compare negative log likelihood, ranked probability score and bias; inspect quantile coverage and cumulative errors.
- Open a historical case to inspect its PMF, CDF and uncertainty bands.
- A model restricted to particular groups is assessed within those groups, without discarding other groups from the overall comparison.
Route eligible models to the right segments
Demand profiles and lead-time context cells can use different saved base models, selected from earlier validation evidence.
- Demand routing balances NLL and RPS ranks; bias breaks ties. Lead-time routing selects the lowest supported validation RPS.
- Frozen groups, eligibility masks and selected components remain inspectable. Sparse contexts use supported parent or fallback routes.
- Save coverage changes, then retrain to apply them. Browsing or filtering analytics does not refit groups or change saved routes.
Compare the consequences of a purchase
Candidate purchases are evaluated over possible demand and receipt outcomes, producing cost distributions and component contributions.
- Compare an alternative quantity or mode against saved evidence.
- Separate forecast accuracy from the economic quality of a decision.
- Inspect heuristic termination and Monte Carlo uncertainty.
- Filter and export selected materials; keep exclusions and incomplete plans visible.
Trace data, models and results independently
Prepare source data, train models and calculate decisions as separate versioned operations, or request a combined refresh.
- Train without replacing the displayed calculations, or calculate again from compatible saved models without fitting them again.
- Trace settings, source revisions, model artifacts and linked executions behind a result.
- Configuration and job submission are explicit actions; saving settings alone does not start a calculation.
Inspect execution rather than just a progress bar
The runtime workspace links preparation, training and calculation to their recorded work, statuses and resources.
- Read stage durations, nested model/component calls and cache activity. Inclusive component timings may overlap and must not be summed as extra stage time.
- Inspect sampled worker-tree RSS and recording coverage; missing historical samples do not become reconstructed peaks.
- The developer view is read-only and restricted to administrators. Opening it starts no training or calculation.
Who does what
After configuration, workers prepare forecasts, scenarios and reports. Refresh, activation and job submission are explicit; the app does not place purchase orders.