Skip to content
varun.rout

Jump to

Search pages, projects and repos

← Back to work

Retail · demand forecasting

ProfessionalReal data

Multi-Horizon Demand Forecasting

Data Scientist · Manor Park Trading Company · 2024

What this is

Demand forecasting across ~7,000 SKUs over Shopify, Amazon, eBay and The Range at three horizons, paired with a seven-cluster customer and product segmentation.

Metrics

Accuracy

+28% weekly · +19% monthly · +37% quarterly

Segmentation

Seven clusters for prioritisation

Campaign test

~7% ROAS

vs a randomly assigned holdout group

Associated revenue

~£78k, association only

Reporting

Automation cut per-draft time ~50%

Stack

PythonForecastingClusteringRandomised holdout testingCampaign analytics

Provenance

Two separate claims, deliberately kept apart. The ~7% ROAS improvement was measured against a randomly assigned control group held back while the new segment-based targeting ran concurrently, so it is a controlled comparison. The ~£78k is not: it is revenue associated with the analysis with no holdout behind it, and is never a causal claim.

What I would not claim

  • The holdout covered the campaign targeting test only. It does not make the ~£78k a causal figure.
  • No sample size, split ratio, test duration, statistical power, MDE, p-value or number of variants is published, because none is on file.

Explore the deltas

The same before/after figures, interactive. Toggle between the three roles.

Residual-load framework improved forward-curve accuracy on both near and far seasons; the Christmas-period fix followed a holiday-specific error diagnosis. Relative improvements only, against an internal baseline.