E-commerce

Demand forecasting.
Asymmetric by nature.

Most forecasting projects optimise a symmetric error metric, then are judged on outcomes that are wildly asymmetric. A stockout on a hero line and an overstock on a slow one are not the same size of mistake, and the model should know that.

THE MISTAKE THAT COSTS THE MOST

A forecast trained to minimise mean absolute error treats a unit of over-forecast and a unit of under-forecast as equally bad. In a real retail business they almost never are. Understocking a high-margin, high-demand line costs the margin and sometimes the customer. Overstocking a perishable or seasonal line costs the markdown and the working capital. The ratio between those differs by category, by season and by product lifecycle stage.

Building that asymmetry into the objective — forecasting a quantile rather than a mean, with the quantile chosen per category from the actual cost ratio — in our experience typically produces a larger commercial improvement than any change of model architecture. It is also the step most often skipped, because it requires a conversation with finance and merchandising rather than with data science.

The second most valuable step is equally unglamorous: modelling promotions and price properly. A forecast that treats a promotional week as an outlier to be smoothed is discarding the signal that matters most for planning.

WHERE LANGUAGE MODELS ACTUALLY FIT

Not the forecast itself

Time-series demand forecasting is well served by statistical and gradient-boosting methods that are faster, cheaper, more accurate and interpretable. Using a language model for the numeric forecast is a fashionable and expensive mistake.

Feature extraction from unstructured data

Product attributes, supplier notes, competitor listings and review text turned into features the forecaster can use. This is where a language model contributes real signal to a classical model.

New-product cold start

Forecasting a product with no history by finding genuinely comparable products through semantic similarity of description and attributes, rather than by category-average heuristics that ignore what the item is.

Explaining the forecast

Translating a model output and its drivers into a narrative a planner can challenge. Planner trust is what decides whether a forecast is used or overridden, and an unexplained number gets overridden.

WHAT WE MEASURE

Not forecast accuracy in isolation. Accuracy is an input; the outputs that matter are service level, markdown, stockout rate and working capital tied up in inventory. A forecast that is marginally less accurate but biased in the commercially cheaper direction is a better forecast for the business, and reporting only error metrics conceals that.

We also measure the override rate — how often planners overrule the system, and whether they were right when they did. Persistent, correct overrides in one category are the clearest possible signal of a missing feature, and they usually point at something the planners know and the data does not record.

MODELLING THE COST, NOT THE ERROR

FORECAST PATH

History

sales · price · promotion
Features

attributes from unstructured data
Quantile forecast

chosen from the cost ratio
Plan

inventory and replenishment
Planner review

override recorded with reason
Correct overrides

the missing-feature signal

Solid: the forecast path Dashed: the override loop, which is where missing features surface

The quantile is a commercial decision taken with finance and merchandising, not a hyperparameter left at its default.

QUESTIONS

QUESTIONS — 3

They are a real and improving research area and worth benchmarking, particularly for cold-start. As a wholesale replacement for a well-tuned classical forecaster on a business with history, we have not yet seen the case made on real retail data — so we would benchmark rather than assume, in either direction.

Enough to see the seasonality you care about, which usually means at least two cycles. Less than that and you are relying on cross-product signal, which is exactly where the semantic similarity work earns its place.

It can detect that the regime has changed and flag it. It cannot forecast an unprecedented event, and a system presented as able to is misleading. What good design gives you is fast detection and a clear handover to human judgement, not prescience.

Keep reading

Supply Chain Planning

→

30-minute technical call · no deck