Preserve forecast paths when planning for cumulative demand
PathFM generates complete future trajectories for forecasting. To use them for cumulative demand or peak-load decisions, preserve each path through the service interface and test the quantity the operational decision actually consumes.
Trace the interfaces between applications and records before changing a system.
A forecast of demand for each day does not necessarily answer how much stock will cover the whole replenishment period. That question depends on whether busy days tend to arrive together. A new model offers a way to represent that relationship directly, but a planning service must preserve it all the way to the calculation that uses it.
PathFM, released as a preprint on 15 October, generates complete future trajectories. A trajectory is one possible sequence across the forecast period. The model uses flow matching to transform noise into a future path, conditioned on history and known future inputs. Distillation shortens the generation process. (PathFM, section 3)
For a planning owner, the opportunity is to ask about accumulated demand, a peak or when a threshold is crossed. The architecture decision is what forecast information the downstream service receives. Returning only separate daily summaries would discard the relationship the path generator was chosen to model.
Identical daily forecasts can imply different stock requirements
Consider a constructed example for an item with no replenishment during a two-day window. Demand on either day is equally likely to be zero or 20 units. Compare two hypothetical patterns. In the clustered pattern, both days are quiet or both are busy. In the alternating pattern, a busy day is always paired with a quiet one. Each listed path has equal probability within its pattern.
| Illustrative pattern | Possible two-day paths and totals |
|---|---|
| Clustered | (0, 0) totals 0. (20, 20) totals 40. |
| Alternating | (0, 20) totals 20. (20, 0) totals 20. |
Both constructed patterns have mean demand of ten units on each day and expected total demand of 20 units. Yet with 30 units of starting stock, demand exceeds stock in half the clustered paths and none of the alternating paths. The calculation is exact under these invented assumptions. It is not a forecast, a measured stockout rate or a PathFM result.
A daily distribution is called a marginal distribution because it describes one part of the future separately. The example shows what it leaves unspecified: how outcomes belong together. The older Foundation Model Forecasts: Form and Function paper, published in October 2025, explains this distinction and the additional dependence assumptions needed to convert separate distributions into joint paths. It is a methodological foundation, not a new release or an independent evaluation of PathFM.
There is no need to generate full trajectories for every forecasting requirement. A single-period decision may need only that period’s distribution. If the sole requirement is expected total demand, adding the daily means is sufficient. The extra information matters when the decision depends on uncertainty in the total, simultaneous peaks or the sequence of events.
Keep path identity in the service contract
For a proposed replenishment assessment, retain a forecast identifier, the information cut-off, ordered time steps, units and a stable identifier for each path. If the model returns unequal weights, retain them too. Preserve the product or location dimension when several series are generated jointly. These are proposed interface requirements, not features verified in a released PathFM service.
The consumer should calculate the relevant total within each path and only then summarise those totals. Independently sorting each day’s values and joining them by rank can change which days belong together. Drawing a separate value for each day introduces a different dependence assumption. Either operation can preserve the daily distributions while changing the stock calculation.
Check that the service preserves each future path
Does processing the forecast retain the dependence between days?
- Fixture
- Clustered paths: (0, 0) and (20, 20). Alternating paths: (0, 20) and (20, 0). Each path has equal weight within its ensemble.
- Procedure
- Send each ensemble through the proposed storage and response format.
- Reconstruct paths using their identifiers and ordered time steps.
- Sum within each path before calculating the share above 30 units.
- Result to check
- Clustered demand exceeds 30 units in half the paths. Alternating demand never does.
- Evidence limit
- Passing checks the interface on this fixture. It cannot establish whether a learned forecast represents future demand.
- Next test
- Check the same path identifiers and derived quantities against the actual forecasting service output.
An exact arithmetic fixture for an interface test. It is not a PathFM run or evidence of forecast accuracy.
Pass two constructed demand ensembles through the forecasting interface. Each has identical daily distributions but different totals. The interface must retain those different totals when it returns the forecast paths.
Use the constructed fixture to check storage, export and reporting as well as the model adapter. A dashboard may display daily uncertainty bands while an inventory service consumes paths, provided both are derived from the intended forecast. The bands should not become the only retained output if a later decision needs the path structure.
This extends the archive’s inventory-policy argument. That article defines what a replenishment agent may order. Here the issue is the uncertainty information available before an ordering policy runs. It also differs from human forecast overrides, which concerns how planners contribute additional information.
Evaluate the quantity used by the planning decision
PathFM’s authors report stronger aggregate path scores than their evaluated baselines over 94 tasks, using 32 paths per forecast. Quantile-only comparators require explicit assumptions to create paths. Additional comparisons fit dependence to historical errors. These are author-reported results, not independent replication or measured inventory savings. (PathFM, sections 4–5 and appendix E)
For a local comparison, fix the replenishment window, stock position and operational question before choosing scores. Compare the candidate with the current forecasting process and an appropriate statistical method. If the incumbent already models dependence, preserve that capability in the comparison. Do not weaken it into independent daily draws merely because those are easier to construct.
Score the distribution of cumulative demand against observed totals and inspect the predicted probability of exhausting the available stock. Report calibration, whether predicted probabilities correspond to observed frequencies, over repeated forecast occasions. Keep results for ordinary demand and unusual periods visible. Retain the information available at each historical cut-off so future revisions do not improve the candidate’s inputs retrospectively.
More sampled paths reduce sampling uncertainty but cannot repair a mistaken model. Rare events can remain poorly estimated with a small ensemble. (PathFM, appendix C.1) The preprint’s reproducibility statement promises a future code and weight release. Verify availability, licence and the serving interface before selecting PathFM for a trial. No model experiment was run for this article.
The planning owner should choose the required forecast quantity and acceptable decision error. The architecture owner should demonstrate that the interface preserves it, then measure generation, transfer, storage and downstream calculation together. Fund a bounded trial where uncertainty over the whole period changes a real decision. Where separate daily forecasts already answer the operational question, retain the simpler output contract.