Sector

Forecasting that survives a festival season.

Retail demand in India is dominated by events that a general-purpose forecasting model treats as anomalies to be smoothed away — which is exactly the demand you most needed to predict.

01 / What is hard here

The problem underneath the problem.

The hard part is not the average week. It is Diwali, a school-holiday shift, a regional festival that moves against the Gregorian calendar, a competitor's promotion, and an unseasonal monsoon — often overlapping.

Compounding it, most retail data records what was sold rather than what was wanted. A stockout looks identical to low demand in the sales log, so a naive model learns to under-forecast precisely the lines that keep selling out.

02 / What we build

Work we take on in retail.

Demand forecasting by SKU and location: Hierarchical forecasts reconciled to the aggregate, with festivals, promotions, local holidays and weather as explicit drivers rather than absorbed as noise.
Stockout-corrected history: Censored demand reconstruction so that periods of unavailability are not learned as periods of low interest. Frequently the single largest accuracy gain available.
Assortment and allocation: Which lines to carry per store format and catchment, and how to distribute limited stock across locations under real replenishment constraints.
Pricing and markdown support: Elasticity estimation and markdown timing on ageing stock, with the commercial guardrails applied as an explicit layer.
Recommendations and search: Two-stage ranking across a large catalogue with cold-start handling, so new stock is visible before it accumulates interaction history.

03 / Constraints

What shapes the build.

These are design constraints agreed at the start, not compliance work bolted on before launch. Retrofitting them is expensive and usually incomplete.

Master data quality: The same product under three SKUs across channels defeats every downstream model. Reconciling this is usually the first phase, not an assumption.
Channel fragmentation: Store, own site and marketplace data arrive in different shapes on different schedules. Unifying them honestly is real work.
Planner adoption: Experienced planners know things absent from the data. Systems they cannot override get worked around, so overrides are designed in and recorded.
Latency at the shelf: In-store and on-site personalisation has a hard response-time budget that shapes the architecture.

04 / What we advise against

Where we would say no.

We would advise against a recommendation engine before event tracking is trustworthy, and against promising a headline forecast-accuracy improvement before seeing how stockouts and promotions are recorded.

We would rather lose the work than build something that creates a liability for you or a harm for the people on the other side of it. If that rules out what you had in mind, it is better established now than at delivery.

05 / Other sectors

Adjacent work.

Different domains, the same delivery discipline — a measured baseline, a real evaluation set, and monitoring that catches decay early.

Next step

Working in retail?

Tell us which process is slow, expensive or error-prone. The diagnostic will tell you whether this is a problem worth solving with machine learning — including if it is not.

Request a diagnostic All industries