*Conceptual illustration of warehouse planning; not a PathFM result and not deployment data.*
PathFM is a research model that generates entire possible trajectories of time series. For procurement, this is a reason to check the demand distribution over the supplier lead time. However, the uncertain arrival date of goods and sales on out-of-stock days require separate rules: the trajectory itself does not determine the order.
On October 15, 2026, Jiachen Yao and co-authors posted the PathFM preprint on arXiv. The review was prepared by the AIrassvet editorial team on October 18 for planning leaders, analysts, and warehouse process owners. We studied the paper, but did not run the model. Below is an explanation of the news and a proposed review process, without any promise of reduced inventory or shortages.
Key takeaway: before choosing a forecasting model, define how far it must cover demand and which observations should count as demand at all.
Contents:
- What's new in PathFM
- What kind of forecast procurement needs
- Two time frames in one problem
- Why sales can hide demand
- What to record before comparing
- Questions about PathFM
- How AIrassvet helps prepare data
- Where to start
What's new in PathFM
The authors use flow matching: the model learns to transform noise into a joint future sequence. Distillation reduces generation to one or two model computations per output block. That is the number of computations, not the forecast duration in seconds. The mechanism is described in section 3 of the study.
In the authors' fps-bench, trajectory quality is compared across 94 tasks using 32 trajectories per forecast. PathFM achieves the best aggregate result among the external baselines considered. For quantile models, the authors explicitly define ways to reconstruct dependence across time points. Therefore, the result applies to this protocol, not to every way of using competing models. See sections 4–5.
The reproducibility statement promises a future release of code and weights. We did not verify the confirmed setup or the usage terms for any specific release. The authors also note a limitation: the model does not yet accept additional modalities such as text. See the discussion and reproducibility statement. The paper's results do not confirm any economic effect in the reader's warehouse.
What kind of forecast procurement needs
A trajectory is one possible future scenario with ordered values over time. A quantile, for example the upper bound of a chosen level, describes a separate part of the distribution. Dependence across future days matters when the decision depends on a sum, a peak, or a sequence of events. This distinction is explained in the PathFM problem formulation.
For procurement discussions, it helps to separate the questions. "How much do we expect to sell tomorrow?" is a question about a single period. "What is the expected total demand?" is a question about the expectation of a sum: with consistent units, daily means can be added. "How much uncertainty is associated with demand until replenishment?" is already a question about the distribution of the sum and the period boundary.
Suggested calculation workflow: within each scenario, define the needed interval, calculate demand over it, and then compare the resulting sums. Do not replace this object with the sum of upper daily quantiles without an explicit assumption about dependence. This is a recommendation for how to set up the review, not a measured PathFM result.
The basic questions of seasonality, promotions, and time-based validation are covered in the article on AI demand forecasting. Here the focus is already on where the demand horizon ends that a specific inventory level must cover.
Two time frames in one problem
Imagine a hypothetical warehouse of consumables. The planner runs the calculation in the morning; arrival is expected by the end of the week, but may be delayed. This is a fictional example for framing the problem, with no customer data and no implementation results.
It involves two different times. The calculation time determines what the system already knows. The replenishment availability time determines up to which day inventory is needed. If scenarios are always cut off at the planned date for convenience, the review stops answering the question of coverage in case of delay.
You can start with separate hypothetical cases: on-time arrival and late arrival. For each case, explicitly fix the availability date and calculate demand up to that point. If delay probabilities are unknown, the cases cannot be presented as a calibrated probabilistic distribution. They should be labeled as scenario analysis.
An additional complication arises if demand and delay are linked by a common cause — for example, a seasonal peak. Randomly mixing them as independent means accepting a separate assumption. Before using such a scheme, it needs to be justified with data or the dependence should be left as an open limitation. In the validated sections of PathFM, we did not find grounds to treat the random lead-time problem as a built-in function of the model.
It is equally important to define what "the goods arrived" means. Unloading, receiving, and being available for shipment may be different events. We suggest choosing one operational event and applying it consistently in the history and in the validation. This is a process-owner decision that cannot be derived from a forecast model ranking.
Why sales can hide demand
In the warehouse review proposed here, you cannot automatically equate zero sales with zero need. If the item was out of stock, the sales log shows completed purchases, but does not say how many customers abandoned their order. Therefore, a zero observation requires context about inventory and product availability.
The practical first step is to keep the original sales and separately mark periods of limited availability. Do not quietly replace such values with "reconstructed demand" and then evaluate the forecast on those same guesses as if they were facts. If a reconstruction method is used, its rule, input data, and uncertainty should be included in the review description.
When comparing historical results, another question matters: what information was available on the calculation date? A later revised arrival, corrected inventory, or final promotion classification may improve the historical result, even though the operational system did not know it at the time. The principle of validation on sequential time cutoffs is described in Forecasting: Principles and Practice, section on time series cross-validation.
From this, we propose two reports: performance over periods with observed sales when the item was in stock, and a separate analysis of periods with limited demand observability. Combining them into one accuracy claim is premature. Generating a large number of scenarios by itself does not remove uncertainty in the source data.
What to record before comparing
Below is a proposed worksheet for discussion. It is not a validated PathFM methodology. Its purpose is to make visible the conditions under which a forecast comparison answers the warehouse decision.
| Field | What to record | What otherwise remains unclear |
|---|---|---|
| Information cutoff | Calculation time and versions of the available data | Whether future information leaked into the past forecast |
| Demand object | Item, warehouse, unit, and time step | Whether values can be summed within the period |
| Coverage boundary | Arrival availability event; rule for including the day | Until When Safety Stock Is Needed |
| Delays | Conditional Scenarios or a Justified Probabilistic Model | What the Timing Uncertainty Is Based On |
| Observability | Sales, On-Hand Inventory, and Stockout Flags | Where Actual Demand Is Unknown |
| Decision Rule | Starting Inventory, Allowed Actions, and Error Criterion | Whether You Are Testing the Forecast or the Ordering Policy Already |
For comparison, we recommend keeping the current forecast and an appropriate statistical benchmark. If the existing system already accounts for time dependence, it should not be artificially weakened to independent daily values. The candidate and the baseline need the same time cutoffs, available features, and constraints.
You should evaluate exactly the chosen target: for example, the distribution of the total over an agreed period. For a probabilistic forecast, you should also check calibration — whether the stated probabilities match the frequencies of observed events across a series of historical calculations. The choice of criteria should be finalized before looking at the candidate's result.
Cost also applies to the entire process: data acquisition, generation, scenario storage, calculation of the final value, and handling exceptions. This article does not measure the time or cost of such an end-to-end workflow. Readiness for comparison starts with a reproducible setup, not with moving a research score into a financial plan.
Questions About PathFM
Can PathFM Be Used for Automatic Product Ordering?
The preprint describes a forecasting model, not a ready-to-use purchasing policy. Ordering separately requires available inventory, lead time, constraints, and rules for approving actions. We did not test such an operational service.
Do You Always Need to Forecast Full Trajectories?
No. For a single period, its distribution may be enough, and for average total demand, the sum of means may suffice. Trajectories are useful when the decision depends on uncertainty in the sum, the peak, or the order of events.
Does PathFM Account for Supplier Delays?
The reviewed study does not confirm a ready-made function for joint forecasting of demand and random lead time. The arrival rule should be specified separately; the relationship between delay and demand also needs justification.
Does fps-bench Prove a Reduction in Stockouts?
No. The author's benchmark compares forecasts under a defined protocol. A warehouse result requires a separate test on suitable data and with a fixed replenishment policy. This article contains no in-house measurements of stockouts.
How AIrassvet Helps Prepare Data for Forecasting
For this task, the relevant AIrassvet areas are ML forecasting, data preparation, integration, and testing. The scope of work can be defined concretely: align sales history with inventory and receipt events; prepare time slices for comparison; connect the validated forecast to the planning system. This is not a claim of ready-made support for PathFM or achieved savings.
The first step is to choose one replenishment process, describe the current forecast, data sources, constraints, and acceptance criteria. Discuss the Project.
Where to Start
PathFM is a reason to reconsider the forecasting target: for an inventory decision, a full future sequence may be needed. But before testing the model, you need to align the cutoff point, the event that makes inbound supply available, and the periods where sales do not reveal demand.
Start with one item and one warehouse: fill out the worksheet and make sure the historical task can be reconstructed from the data known at that time. If the coverage boundary or demand remain unknown, keep that limitation in the report. Then the subsequent model comparison will answer a clear operational question.
