— What we hear
"Our forecast worked beautifully until something changed, and now we're $11M out of balance."
Every retail forecasting model is built on the assumption that the future looks like the past. When the past stops being a good predictor — regulation, geopolitics, viral demand shifts — the model doesn't know it's wrong until the inventory team does.
"Personalisation lifts conversion but kills lifetime value, and nobody on the team knows why."
Optimising on a single objective produces predictable distortions. The recommender that maximises click-through is rarely the one that maximises retention. The metric you choose is the strategy you've adopted, whether you meant to or not.
"Returns are eating our margin and we can't predict them at order time."
Returns are the silent killer of apparel and consumer goods economics. Models that predict them at order time enable pricing, fulfillment, and packaging decisions that compound. Most teams haven't built that layer because the cross-functional alignment is too hard.
"Marketing wants dynamic pricing and finance won't let them ship it."
Dynamic pricing in retail is a governance problem masquerading as a technical one. Without explicit rails, finance is right to push back. With them, pricing AI becomes the highest-ROI deployment in the stack.
— Pattern recognition
— What fails
Forecasting models without regime detection. Every model fails on regime shifts. Building a "better" model is the wrong response — building a system that knows when its model is about to fail is the right one.
Single-metric optimisation. Conversion alone. CTR alone. AOV alone. Each produces a recommender that destroys some other part of the funnel. Multi-objective from day one.
Treating recommendation as a model problem. The hard part isn't the model — it's the cold start, the diversity controls, the merchandising overrides, the seasonal calibration. The model is 20% of the work.
Pricing AI without finance rails. If finance doesn't have a kill switch and override pathway built in, pricing AI gets unplugged within 90 days of the first surprise. Build the rails first, the model second.
— What works
Regime-detection layers above forecasting. Distribution shift monitoring across 10–20 input signals. When the regime layer fires, predictions degrade weight and fall back to conservative envelopes the team trusts.
Multi-objective recommenders. Conversion, retention proxy, margin, and diversity as joint objectives. Trade-offs explicit and tunable per surface.
Returns prediction at the SKU-customer level. Input to pricing, fulfillment routing, and bundle composition. The compounding model — small accuracy gains move several P&L lines at once.
Governance-first pricing. Bounds, audit trails, manual overrides, and a kill switch built in week one. The model deploys when finance signs off, which they do faster when the rails are already in place.
— Engagement patterns
— 01
Forecasting with regime detection. Fallback envelopes for shock events. SKU-store-day granularity. Connected to allocation and replenishment decisions, not just dashboards.
— 02
Multi-objective recommenders. Diversity controls, merchandising overrides, cold-start handling. Built for the retention surface, not just the conversion surface.
— 03
Governance-first dynamic pricing. Elasticity modeling, bounds, kill-switches, audit trails. Finance-approved before it ships.
— 04
Returns prediction at order time. Routing optimisation. Fraud-vs-friction balance on return acceptance. Touches pricing, packaging, fulfillment.
— Pre-IPO specialty retailer (illustrative)
A pre-IPO specialty retailer's forecasting model worked — until the category collapsed during a regulatory shift. We didn't rebuild the model. We built a regime-detection layer that knew when the model was about to be wrong, and what to fall back on. Stockout days fell 23% in the year following deployment.
— Composite case drawn from engagement patterns; anonymized for client confidentiality. Real metrics, real outcome arcs.
— If this resonated
If you're a RetailTech leader sitting with one of the problems on this page, applications are open for next quarter. We read every one.