Skip to main content

Inventory forecasting agent

A zero sale during a stockout is not zero demand.

An inventory forecasting agent estimates demand ranges for a defined item, location and planning horizon, then helps planners examine shortage, excess and expiry risks. It should account for stockouts, source timing and uncertainty before comparing options. Werkon would validate this pattern with historical baselines and planner review; service targets, purchasing and stock movements remain authorized business decisions.

Define what the forecast measures

Gross requests, accepted orders, shipments, point-of-sale units, consumption and return-adjusted sales are different targets. Preserve cancellations, substitutions, rationing, transfers and returns. Zero sales may reflect a stockout, closure, delisting or channel outage. Treat constrained demand as uncertain and let qualified owners choose any lost-sales estimation method.

Bind each series to exact item, pack, location, channel, unit, calendar and hierarchy versions. Separate physical and system stock from reserved, allocated, available-to-promise, in-transit, quality-held, damaged and expired quantities. Reconcile duplicate, delayed and corrected events; do not count a transfer twice or add incompatible units.

Evaluate only what was knowable then

Freeze the information available at each forecast origin. Planned price, promotions, assortment and supplier terms retain their historical versions; later execution, revised sales or weather observations cannot leak into an earlier prediction. Version transformations, lag windows, missing-data rules, outlier treatment and model configuration.

Use time-respecting evaluations and keep model selection separate from the final operating holdout. Compare simple and seasonal baselines. Report weighting and results for sparse, new, seasonal, volatile, promoted and discontinued items so aggregate performance does not hide weak segments. A more complex model needs evidence of useful improvement.

Separate uncertainty from stock policy

Label means, medians, quantiles, intervals, distributions and scenarios explicitly. Check calibration, coverage, sharpness, quantile crossing and tail stability by horizon and material segment. A quantile is not a guarantee. Hierarchy reconciliation cannot repair missing or weak source history.

Stock scenarios also depend on inventory position, open supply, lead times, supplier reliability, order calendars, minimum quantities, pack multiples, shelf life, storage, capacity, substitution and costs. Keep demand uncertainty separate from supply uncertainty and show the assumptions behind shortages, slow-moving stock and obsolescence candidates.

Follow the planning decision into operation

A planner override is a decision to evaluate, not a correction to historical truth. Preserve the alternatives, reason, assumptions, approvals and expiry of a proposal. Recheck supplier and inventory state before any separately authorized order or transfer; unusual, regulated, perishable and safety-critical items require the appropriate owner.

Reconcile later sales, backorders, receipts, stock counts, transfers, expiries and write-offs against the original planning packet. Distinguish forecast error from supply, policy, execution and data defects. Avoid evaluating only stocked or sold items, and do not treat rejected recommendations as observed counterfactual outcomes.

Forecast boundary

Forecast the target without erasing the constraint.

Observed sales, latent demand, inventory position and replenishment authority are different systems. Four boundaries preserve the planning decision.

01

Item, target, availability, and source history

Bind durable item-location-channel identity, units, time grain, hierarchy, target and forecast origin; reconstruct orders, sales, returns and inventory states from authoritative events and mark stockout or partial-availability intervals instead of treating every zero as demand.

Required evidence: Tenant and business scope, item, SKU, variant and pack, location and channel, target definition, unit and conversions, calendar and hierarchy version, source records and event times, on-hand, reserved, allocated, available, backordered and in-transit states, stockout and range intervals, corrections, substitutions and censoring decision.

02

Time-safe forecast and uncertainty evidence

Freeze every origin, use only known or legitimately planned features, compare simple baselines and qualified models across representative rolling cohorts and expose point, quantile, interval, bias, calibration, segment and abstention results.

Required evidence: Origin and horizon, training and validation cutoffs, feature availability and revisions, planned versus realized price and promotion, transformations and missingness, baseline and model versions, hierarchy reconciliation, loss and weights, point and quantiles, coverage, sharpness, crossing, bias, error by segment, drift and unsupported cases.

03

Stock-risk scenario and policy assumptions

Combine the forecast distribution with a versioned deterministic inventory model whose lead time, open supply, service objective, costs, order constraints, capacity, shelf life and substitution assumptions remain visible and owned by planners.

Required evidence: Inventory-position snapshot, open orders and due dates, lead-time source and distribution, supplier state, review and order calendar, service target, shortage and holding-cost assumptions, minimum and pack quantities, capacity, shelf life and expiry, transfer and substitution rules, demand-during-protection range, shortage, excess and expiry scenarios and sensitivity.

04

Planner decision, released action, and outcome

Keep model forecast, risk candidate, planner override, approved plan and external order distinct; recheck current sources before release and reconcile later sales, receipts, stockouts, service, excess and waste without crediting the model for the whole system.

Required evidence: Candidate action set, selected or rejected action, planner and reason, override, assumptions, freshness checks, approvers, expiry, order or transfer request, idempotency, external acceptance, supplier acknowledgment, receipt, sales and censored intervals, fill and backorder, inventory, expiry, write-off, waste, correction and contribution analysis.

Source-to-outcome path

Keep demand, forecast, risk, decision, and action separate.

A calibrated distribution can inform a planner. It cannot choose the service policy or place the order.

  1. 01

    Define the item-location demand contract

    Resolve tenant, product and location identities, target event, units, time grain, hierarchy, origin, horizon, planning decision and accountable owners; document substitutions, returns, cancellations, availability censoring and what the forecast excludes.

    Owner
    Demand-planning, product-master, inventory and data owners
    Evidence
    Stable identifiers, target and exclusions, units and conversions, calendar and local time, hierarchy and mapping version, origin schedule, horizons, demand event states, substitution and return policy, availability evidence, censoring method, affected decisions and owners.
  2. 02

    Reconstruct the decision-time source packet

    Join orders, shipments, sales, returns, counts, receipts, transfers, adjustments, promotions, prices and supplier records at their valid versions; deduplicate and order events, expose conflicts and build point-in-time inventory and availability without silent revision.

    Owner
    Order, sales, inventory, logistics, merchandising and supplier-data owners
    Evidence
    Source allowlist and field authority, record and event identifiers, event, received and valid times, schemas, units, corrections, order and fulfillment states, inventory positions, stockout intervals, promotion and price versions, lead-time observations, missingness, conflicts and lineage.
  3. 03

    Evaluate point and distribution forecasts

    Create rolling origin datasets with features available at each origin, run simple baselines and candidate methods, preserve hierarchy reconciliation and evaluate error, bias, quantiles, coverage, sharpness and material cohorts without future or selection leakage.

    Owner
    Forecasting, model-risk, data-science and planner-review owners
    Evidence
    Dataset and corpus digest, origin and horizon set, feature cutoff audit, target and transform versions, baselines, models and code, parameter search boundary, validation and holdout, metrics and weights, cohort results, quantile calibration and crossing, residuals, drift, explainability, abstention and reviewer decision.
  4. 04

    Simulate stock risk under owned policy

    Combine approved forecast distributions with current inventory, open supply and versioned lead-time and inventory-policy assumptions; produce sensitivity ranges for shortage, excess, expiry and service while keeping forecast, supply and execution uncertainty distinct.

    Owner
    Inventory, procurement, operations, merchandising and finance owners
    Evidence
    Forecast and inventory versions, open supply and supplier state, lead-time distribution, service and risk targets, review period, order calendar, minimum and pack quantities, capacity, shelf life, costs, substitution and transfer rules, scenario engine, sensitivity and unsupported constraint route.
  5. 05

    Decide, release, and measure realized service

    Present scenarios and assumptions to the planner, record action or override, re-read source and authority before any approved release and later reconcile orders, receipts, constrained demand, service, excess, expiry and waste against the frozen packet.

    Owner
    Named planner, procurement authority, inventory operations and outcome owners
    Evidence
    Candidate and comparison, planner decision and rationale, override and expiry, approvals, source freshness, released action and idempotency, external order and acknowledgments, receipt and count, sales and stockout state, fill, backorder, excess, age, expiry, write-off, waste, correction and contribution review.

Authority map

Separate forecast controls, analytical assistance, and inventory authority.

A model can estimate a tail quantile. It cannot choose a service target, supplier commitment or shelf-life risk.

01

Deterministic forecasting controls

Software owns tenant, identity and target boundaries, source authority, event deduplication, temporal cutoffs, units, availability censoring, hierarchy, transformations, known-at-origin features, calculations, inventory policy versions, access and receipts.

  • Item, location, target, source, event, forecast, scenario, decision, order and outcome identifiers
  • Forecast-origin, feature-availability, hierarchy, unit and correction enforcement
  • Stockout, quantile, interval, lead-time, constraint and action-precondition validation
  • Source, forecast, review, override, release, receipt and outcome receipts
02

Bounded AI assistance

Models can classify source anomalies, propose target mappings, forecast within evaluated boundaries, summarize source-linked drivers, explain baseline differences and draft planner questions, but cannot change policies or release actions.

  • Item and event mapping candidates with source evidence
  • Data-quality, stockout-censoring and future-leakage flags
  • Point, quantile, interval and baseline explanations
  • Stock-risk narrative, planner question and override-note drafts
03

Human planning and commercial authority

Qualified people own target meaning, data corrections, lost-sales method, promotion and assortment assumptions, model release, service and risk policy, lead-time judgment, supplier and capacity constraints, recommendations, overrides, orders and stop decisions.

  • Demand-target, censoring, substitution and hierarchy decisions
  • Forecast method, uncertainty, cohort and release decisions
  • Service target, cost, lead-time, shelf-life, capacity and scenario decisions
  • Planner override, procurement, transfer, allocation, supplier and retirement authority

Forecasting components

Build a demand-to-service ledger, not a single-number forecast.

Sales, stock, supply and policy change on different clocks. Four ledgers preserve what a planner actually knew.

01

Target, item, and availability ledger

Bind item, SKU, pack, location, channel, target event, units, time grain, hierarchy, origin and horizon to source orders, allocations, shipments, sales, returns, counts and inventory positions plus stockout, assortment and opening intervals.

Operating contract: Name is not identity, sale is not request, shipment is not consumption, return is not negative demand without policy, zero sale is not zero demand, system on-hand is not physical count, transfer is not net network demand and aggregate is not a compatible sum when units or mappings differ.

02

Feature, forecast, and uncertainty ledger

Version cutoff-safe features, planned and realized promotions and prices, calendars, transformations, missingness, baselines, models, hierarchy reconciliation, point forecasts, quantiles, intervals, bias, coverage, residuals, cohort results, drift and abstention.

Operating contract: Planned is not realized, revised history was not known, future observation is leakage, point is not distribution, quantile is not guarantee, narrow range is not calibration, aggregate accuracy is not item fitness, leaderboard result is not local evidence and model fluency is not forecast quality.

03

Inventory-policy and stock-risk ledger

Preserve inventory position, open supply, lead-time distribution, supplier reliability, review and order calendars, service and risk objectives, shortage and holding costs, order and pack constraints, capacity, shelf life, substitution, transfer, scenarios and sensitivity.

Operating contract: Forecast uncertainty is not lead-time uncertainty, open order is not receipt, supplier promise is not arrival, expected demand is not protection-period distribution, stock-risk scenario is not fact, service objective is not universal and recommendation is not authority.

04

Planner, action, and outcome ledger

Record candidate action, comparison, planner decision, override, assumptions, approval, expiry, source recheck, order or transfer release, external acceptance, receipt, count, constrained sales, service, backorder, excess, age, expiry, waste and corrections.

Operating contract: Candidate is not decision, approved plan is not released order, provider acceptance is not supplier commitment, supplier acknowledgment is not receipt, receipt is not available stock, observed service is not model causation and rejected action has no known counterfactual outcome.

Delivery path

Prove one item-location forecast through realized stock evidence.

Start with one category, network slice and planning cadence where availability, supply and service can be reconstructed at every forecast origin.

  1. 01

    Choose one bounded planning decision

    Select item-location-channel grain, demand target, units, hierarchy, horizons and one inventory decision; name product, sales, inventory, forecasting, procurement, operations, finance, supplier, data, model-risk and stop owners.

  2. 02

    Map sources and replay historical origins

    Inventory orders, shipments, sales, returns, counts, receipts, transfers, adjustments, promotions, prices, assortment and lead times; reconstruct sanitized point-in-time packets with event corrections, stockout censoring and planned versus realized features.

  3. 03

    Build baselines and uncertainty gates

    Encode identities, units, temporal cutoffs, known-at-origin features, hierarchy, censoring and evaluation cohorts; compare simple baselines and candidate point and quantile models and require bias, calibration, segment, drift and abstention evidence.

  4. 04

    Pilot scenarios under planner review

    Combine evaluated distributions with current stock, open supply and approved policy assumptions; let planners review shortage, excess and expiry ranges, record overrides and compare workload, service, inventory and waste without releasing orders automatically.

  5. 05

    Release narrowly and reconcile causes

    Permit only proven low-risk action preparation, require fresh source and authority checks for releases and join later receipts, counts, constrained demand, service and waste to forecast, policy, supply and execution contributions before expanding.

Release controls

Six controls before a forecast can inform stock action.

A better error score cannot repair a censored target, leaked promotion or impossible order constraint.

Target and units are exact
Bind item, pack, location, channel, demand event, time grain, hierarchy, origin, horizon, units, conversions, cancellations, substitutions and returns; reject mixed targets and incompatible aggregation.
Availability constrains observed sales
Reconstruct on-hand, reserved, allocated, available, backordered, in-transit, ranged and open states from authoritative point-in-time evidence; mark full and partial stockouts and require approved treatment of censored or lost demand.
Every feature existed at the origin
Freeze source versions at each forecast origin, distinguish planned from realized price, promotion, assortment and operations and block revised history, future inventory, later events and other leakage from training and evaluation.
Uncertainty is evaluated, not decorated
Compare baselines on rolling representative cohorts, expose point and quantile semantics, bias, coverage, sharpness, crossing, segment error and drift and widen or abstain for sparse, intermittent, novel, shifted and unsupported series.
Stock risk shows policy assumptions
Version inventory position, supply, lead-time distribution, service and cost choices, order calendar, minimums, packs, capacity, shelf life, transfer and substitution rules; separate demand, lead-time and execution uncertainty and show sensitivity.
A planner owns every external action
Keep forecast, scenario, candidate, override, approved plan and order distinct, preserve current source and authority checks, reasons, expiry and idempotency and reconcile later service, excess, expiry and waste without unsupported attribution.

Proof model

Measure calibrated demand ranges and realized stock tradeoffs.

Average accuracy can improve while sparse items stock out, promotions leak and perishable inventory expires.

Baseline

  • Business units, categories, items, packs, locations, channels, demand targets, units, grains, hierarchies, origins, horizons, calendars, planning decisions, suppliers, lead-time classes, service policies and planner groups
  • Current source coverage, event delay and correction, unit and hierarchy changes, stockout and partial-availability intervals, substitutions, returns, planned and realized promotion or price, sparse and intermittent series and source outages
  • Current baseline and model error, bias, quantile coverage and sharpness, crossing, residuals, drift, abstention and planner overrides by horizon and material segment
  • Current stockouts, fill, backorders, inventory, excess, age, expiry, write-off, waste, supplier and execution variance kept separate from forecast and policy evidence

Outcome evidence

  • Correct item, unit, target and hierarchy mapping, decision-time source completeness, duplicate and correction handling, availability-censoring coverage and planned-versus-realized feature integrity
  • Point and distribution performance against simple baselines, bias, calibration, sharpness, stable hierarchy reconciliation, drift detection and appropriate abstention by item, location, horizon and material cohort
  • Planner-visible stock-risk sensitivity, constraint-valid candidate actions, review and override quality, source freshness, idempotent release preparation and action-to-receipt reconciliation
  • Service, fill, backorder, inventory, excess, expiry, waste and planner burden measured separately, with supply and execution changes visible and causal value withheld without suitable comparison

Guardrails

  • Wrong tenant, SKU, pack, location, channel, target, unit, calendar or hierarchy; duplicate or missing event; late correction hidden; transfer double counted; phantom stock; stockout or delisting treated zero demand; substitution or return misclassified
  • Realized promotion used as plan, future sale or inventory leaked, revised history backfilled, sparse item hidden by aggregate, baseline omitted, selected holdout reused, point called certainty, quantile crossing, undercovered tail, drift or unsupported cohort ignored
  • Lead time treated fixed, open supply called receipt, service target invented, minimum or pack ignored, capacity or shelf life omitted, supply and demand uncertainty merged, shortage or excess scenario called fact, hidden cost or sensitivity
  • Model releases order, stale proposal used, concurrent override lost, supplier acceptance called receipt, receipt called available, constrained sale called true demand, rejected action assigned an outcome, service or saving attributed without evidence, rollback or stop fails

Fit test

Use this pattern when demand and availability can be reconstructed at each origin.

Good reason to begin

  • One category and network slice has named product, sales, inventory, forecasting, procurement, operations, finance, supplier, data, model-risk and stop owners plus planners with capacity to review scenarios and overrides.
  • Stable item, pack, location and channel identities, target events, units, hierarchies, origins, horizons, source authority, event corrections, availability states and planned feature versions are explicit.
  • Representative histories include stockouts, partial availability, returns, substitutions, promotions, sparse and intermittent demand, supplier delays, counts, corrections and realized service and waste without future leakage.
  • Inventory position, open supply, lead-time evidence, service policy, order constraints, capacity, shelf life, planner decisions, released actions, receipts and outcomes can be replayed from the decision-time packet.

Resolve before beginning

  • The item-location target, unit, hierarchy, forecast origin, availability source, stockout treatment, feature cutoff, lead-time owner, service policy, planner authority or outcome evidence is undefined.
  • The process cannot distinguish order from sale, sale from demand, system from physical stock, planned from realized promotion, point from distribution, forecast from scenario, candidate from decision or supplier acceptance from receipt.
  • Success is defined by one aggregate accuracy score or lower inventory without sparse-item bias, interval calibration, stockouts, service, backorders, excess, expiry, waste, planner burden and supply or execution effects.
  • The agent is expected to invent lost demand, trust stale inventory, use future data, guarantee quantiles, choose service levels or costs, bypass order constraints, select suppliers, place orders, change prices or promise savings and availability.

Source basis

Sources behind the control model.

  • 01

    GS1

    EPCIS 2.0.1

    GS1 lists EPCIS 2.0.1, published 1 July 2025, as the latest version of its standard for sharing supply-chain visibility event data with business context. It does not authenticate an event producer, guarantee identifiers, event completeness, order or current inventory, define the demand target, recover stockout-censored demand, supply a forecast or lead time, authorize an inventory action or prove service or outcome.

  • 02

    United Nations Economic Commission for Europe

    Buy-Ship-Pay Reference Data Model

    UN/CEFACT describes the Buy-Ship-Pay Reference Data Model as a generic semantic business standard for contextualized supply-chain data exchange across organizations and industries. It does not authenticate a party or record, make every implementation complete or current, define local order, inventory or demand truth, resolve units or corrections, supply a forecasting dataset or model, choose inventory policy or prove an action or outcome.

  • 03

    International Journal of Forecasting

    M5 accuracy competition: Results, findings, and conclusions

    This primary study reports the M5 Accuracy competition on point forecasts for 42,840 hierarchical Walmart unit-sales series and discusses its methods and findings. That competition evidence does not establish fitness for another organization, target, horizon, item mix or operating process; resolve stockout censoring, source quality, lead time or inventory constraints; select a model here; guarantee future accuracy; or prove service or business outcome.

  • 04

    INSEAD

    The M5 Uncertainty Competition: Results, Findings and Conclusions

    This primary research page describes the M5 Uncertainty competition, which evaluated nine quantiles for uncertainty distributions across 42,840 hierarchical Walmart unit-sales series. It does not guarantee calibration outside that dataset and evaluation, define the meaning of another interval, recover unconstrained demand, choose a service level or risk appetite, validate inventory or supplier inputs, authorize replenishment or prove stock, waste or financial outcomes.

[ WORKFLOW / SYSTEMS AUDIT ]
THE FIRST ENGAGEMENT

Start with one real workflow

A Systems Audit is the usual starting point. If the opportunity is already clear, we can move directly into a focused build.

Show Us the WorkflowStart with the free automation readiness checklist

OBSERVEQUANTIFYDECIDEBUILD

Continue with the work

Choose the next useful path.

Stay with the operating question. The technology can wait until the work is clear.

01

Explore governed AI agents

Compare other bounded agent patterns and the evidence, authority and recovery contracts they require.

Open path02

Design supply-chain workflows

Connect planning, inventory, logistics and supplier decisions through time-valid data and accountable execution.

Open path03

Build retail and commerce systems

Frame product, inventory, order and customer workflows around operational truth and measurable outcomes.

Open path04

Compare network-level intelligence

Inspect a broader supply-chain pattern that separates source facts, forecasts, risks, scenarios, decisions and released actions.

Open path