What we build
Every layer of inventory decisioning — modelled.
Forecasts drive allocations. Allocations drive transfers. Transfers feed replenishment. Replenishment performance feeds back into forecasts. Each layer documented, versioned, and monitored.
Demand forecasting · SKU × store
LSTM demand models with promotional uplift per SKU × store × day — probabilistic forecasts with explainable drivers, benchmarked weekly against actuals.
Transfer recommendations
Graph-based transfer engine matches excess inventory at one store to gaps at another — respects size curves, service-level targets, and carrier lead times.
Replenishment optimisation
Right-time, right-quantity replenishment from DC to store — lead-time aware, supplier-SLA aware, and aligned to marketing calendar and weather outlook.
Stockout alerts + recovery plans
Real-time stockout detection at the shelf — with a recommended recovery plan (nearest-store transfer, emergency replen, or markdown avoidance) surfaced immediately.
New-launch allocation · cold start
First-time SKUs with zero historicals allocated using attribute-similarity regression, wishlist-match signals, and early sell-through feedback within 48 hours.
Markdown scheduling + shrinkage
Seasonal markdown cadence optimised against carry cost, recovery curves, and brand equity. Shrinkage analytics flag loss hotspots by store, SKU, and hour.
Use cases we model
Inventory decisions happen every hour. We model every one.
Fashion lifecycles, beauty velocity, regional demand curves, new-launch cold-start, markdown cadence, shrinkage patterns — each is a distinct decisioning problem, and each is wired into the same shared feature store and orchestration plane.
Fashion · seasonal collections
SS / FW / resort drops with short lifecycles — cold-start allocation, rapid transfer, and markdown scheduling that protect full-price sell-through.
Beauty · high-velocity SKUs
Skincare and fragrance with fast turnover, gift-with-purchase cadence, and dense store grids — hour-level stockout detection and recovery routing.
Lifestyle · 50+ store networks
Multi-city, multi-country networks where local demand shape differs per store — allocation respects regional size curves, local events, and weather outlooks.
New-launch planning
Pre-season planning and first-season allocation for SKUs with no historical signal — attribute similarity, wishlist match, and 48-hour early-read feedback.
Markdown + clearance
End-of-season markdown optimisation — cadence, depth, and location routing tuned against carry cost, recovery, and brand-equity constraints.
Shrinkage analytics
Shrinkage attribution by store, SKU family, and hour — pattern detection across checkout, stockroom, and transfer events with investigation-ready audit trails.
Model families we deploy
No single model covers every inventory decision. So we compose them.
Deep sequence models forecast, graph networks route, linear programming solves constraints, regression handles cold-start. Each model is evaluated and retrained independently, each contribution logged on every decision.
Long-short-term-memory network over SKU × store × day features — promotional uplift, weather, footfall, local events. Quantile-regression output for probabilistic planning.
Graph over stores with edges weighted by transfer cost and historical flow — computes optimal moves to rebalance inventory while respecting size curves and service levels.
Classical LP solver constrained by store capacity, size curve, service-level targets, and promotion calendar — fast enough to re-optimise on every receipt cycle.
Regression over product attributes (category, fabric, price band, design codes) and early sell-through feedback — produces first-week allocations for items with no history.
Data sources wired into every forecast
Every signal that moves a stock decision — integrated.
Pulled in real time, normalised into one SKU × store × day feature schema, versioned alongside the model that consumes them.
Explainable forecasts, not black boxes
A number isn't enough. A merchant needs the drivers behind it.
Every forecast ships with its top drivers — promotion, weather, trend, cannibalisation, cold-start feedback — ranked and quantified. Merchants can trust it, challenge it, and override it with a full audit trail attached.
- Top drivers per forecast with contribution weight
- MAE / MAPE / bias reported per SKU × store × week
- Override reasoning logged with approver chain
- Weekly cycle review with merchandising leadership
Frameworks we align to
Why Axccelerate for inventory analytics
Not a forecasting module.
A decisioning plane.
A forecasting module gives you a number. Our decisioning plane gives you the forecast, the allocation, the transfer, the replenishment, and the reason codes — all wired to your existing RELEX / Blue Yonder / Oracle / SAP estate.
Pricing
Priced to the catalogue and fleet, not the forecast volume.
Inventory deployments are custom — we scope against your category count, store count, and planning-system shape before quoting.
Glossary
The vocabulary behind every stock decision.
A quick reference for the language your merchandising, allocation, and planning teams use — and the terminology your planning-system vendors will reach for in conversations.
- SKU
- Stock Keeping Unit
The most granular product identifier — one variant (colour × size × pack) of a style. Forecasts, allocations, and replenishment all operate at SKU grain.
- Sell-through
- Units sold / units received
The share of received inventory that has actually been sold in a window. The headline health metric for seasonal and collection-driven categories.
- Lead time
- Order-to-delivery duration
Days between placing a replenishment order and receiving it at the destination. Drives safety-stock calculations and replenishment timing.
- Safety stock
- Buffer against demand variance
Extra inventory held to cover the uncertainty in demand between replenishment cycles. Sized from demand σ, service-level target, and lead time.
- Replenishment
- Restocking flow from DC to store
The regular cycle of moving inventory from distribution centres to stores to match forward demand — typically daily for fast-movers, weekly for slow ones.
- Allocation
- First-receipt store distribution
The decision of how many units of a newly received shipment to send to each store. Drives opening sell-through for new launches.
- Transfer
- Store-to-store inventory move
A sideways move of inventory between stores to correct allocation mistakes or respond to localised demand surges. Cheaper than markdown; faster than DC replen.
- Markdown
- Price reduction to clear
A deliberate price cut to accelerate sell-through on slow-moving or end-of-season inventory. Tradeoff between recovered revenue and brand-equity cost.
- WOS
- Weeks of Supply
On-hand inventory divided by average weekly demand — the headline measure of how long current stock will last. Key input to replenishment triggers.
- Stockout
- Zero on-hand at shelf
A SKU × store at zero on-hand while demand still exists. Direct lost sales, and a leading indicator of basket abandonment and channel switching.
- Cold start
- No historical signal
Allocation or forecasting for items with no (or near-zero) sales history. Requires attribute-similarity modelling, early-read feedback, and human-plus-AI review.
- Cannibalisation
- Uplift stealing from self-portfolio
When a new or promoted SKU's sales come at the expense of another SKU in the same portfolio rather than from new demand. Must be netted out of uplift calcs.
- Size curve
- Distribution of units across sizes
The expected proportion of demand by size within a style — varies by category, region, and customer profile. Wrong size curves cause breakage and lost sales.
- Shrinkage
- Inventory loss not explained by sale
The gap between book inventory and physical stock that isn't accounted for by legitimate sales, returns, or transfers — theft, damage, or process error.
Your inventory plane, engineered.
30-minute scoping with a senior engineer and an inventory-systems operator. You'll leave with a forecast plan, integration sketch, and realistic timeline — not a sales pitch.