Reading your forecast

How to interpret point forecasts, selected models, scores, and cold-start results.

The forecast chart

The chart places historical demand and the future point forecast on the same monthly timeline. Each forecast value is y_hat for one DFU and one future ds month.

This delivery produces point forecasts only. It does not calculate or display confidence intervals. Treat the forecast as a model estimate, not as a guaranteed upper or lower bound.

Selected model and score

Each DFU result contains:

FieldMeaning
best_modelThe model selected for that DFU from the configured candidates
metricThe metric configured for model selection
scoreThe selected model's value for that metric; it may be unavailable for a cold-start DFU
cold_startWhether the DFU was forecast without normal historical model selection
cold_start_methodThe fallback method used when cold_start is true

The available configured metrics are:

MetricWhat it measures
WMAPEAbsolute error weighted by actual demand
MASEAbsolute error scaled against a naive in-sample error
MAEAverage absolute error in demand units
RMSEError in demand units with more weight on larger misses
MAPEAverage percentage error; interpret carefully near zero demand

Lower scores generally indicate less error when DFUs use the same metric. Do not compare raw MAE or RMSE values across DFUs with very different volumes without considering scale.

The result does not include bias, a naive-baseline comparison, confidence intervals, or scores for every candidate model. best_model is the selected result, not a complete model leaderboard.

Cold starts

A cold start is a DFU that cannot use the normal historical evaluation. The engine reports cold_start: true and one of these methods:

  • client_relation — use the relation configured by the planner;
  • similarity_analog — borrow a pattern from a similar DFU;
  • hierarchy_share — derive demand from the hierarchy;
  • mean — use a mean-based fallback.

When the method is client_relation, confirm that the source DFU is still a meaningful analogue. A cold-start score can be null because there is not enough actual history for the selected metric.

Reviewing the portfolio

Start with DFUs that combine material forecast volume with a high score. An aggregate portfolio value can hide a small number of operationally important DFUs.

Then check:

  1. whether the historical series matches the cleaned monthly source data;
  2. whether the selected metric is appropriate for the DFU's scale and zero frequency;
  3. whether the DFU is a cold start;
  4. whether known-future regressors were populated in the uploaded future rows.

When the forecast disagrees with the commercial view

Check in this order:

  1. History — does monthly y reflect the business event correctly?
  2. Dimensions — does the selected hierarchy define the intended DFU?
  3. Regressors — was a planned event present in the input, including future months when known?
  4. Cold start — did the engine use a fallback or client relation?
  5. Configuration — were the intended models, horizon, metric, and reconciliation selected?

If the input itself is wrong, correct it in a new planning scenario. If only the run parameters need to change, use the adjust and reprocess flow.

Exporting

The forecast CSV contains the run, DFU identifier and dimensions, monthly ds, point forecast y_hat, best_model, metric, score, cold-start flag, and cold-start method. The exceptions export identifies cold starts, missing scores, and high-error results.

Exports do not contain confidence intervals, bias, baseline comparisons, or per-model comparison results.