Sources & Assumptions Register

Full transparency on data provenance, modelling decisions, and known limitations.

Data Sources

Source	Description	Link
UCI Online Retail Dataset	UK gift/novelty wholesaler transactions, Dec 2010–Dec 2011. B2B proxy dataset.	UCI ML Repository
Bhavya Kaushal — Tableau Public	RFM segmentation dashboards built on the cleaned dataset	Tableau Public
Klaviyo Win-Back Benchmark Report	Consumer retail email win-back campaign response rates (2022–2024). Used for 10–20% recovery rate assumptions.	Klaviyo Blog
Salesforce Marketing Cloud Research	Email campaign performance benchmarks used to cross-validate recovery rate assumptions.	Salesforce Research

Assumptions Register

#	Assumption	Value / Basis	Confidence	Risk if Wrong
1	UCI dataset used as a B2C proxy	B2B dataset, B2C recommendations	High (disclosed)	Segment dynamics and win-back rates differ for true B2C retailers
2	RFM scores by quintile binning (1–5)	Standard methodology	High	Thresholds may not reflect true behavioural breaks
3	CLV proxy formula	(revenue / lifespan_days) × 365 × frequency	Medium	Single-transaction customers inflated; corrected via P99 winsorisation
4	CLV regression features: R, F, M only	R² = 0.1426	High (disclosed)	Low explanatory power — geography, category, timing would improve accuracy
5	Signal thresholds data-driven	P95 CLV; r_score==3	High	Require recalibration on new datasets
6	Win-back recovery rate — Conservative	10% (consumer retail benchmark)	Medium	B2B win-back is typically 5–12%
7	Win-back recovery rate — Base	15% (consumer retail benchmark)	Medium	Optimistic for wholesale context
8	Win-back recovery rate — Optimistic	20% (consumer retail upper quartile)	Low	Upper bound; requires strong personalisation and offer
9	Email campaign cost (India)	₹58,305 total (₹5/send + ₹55K fixed)	Medium	Agency rates vary; premium creative would increase cost
10	INR/GBP conversion	₹107 per £1	Medium (rate as of project date)	Exchange rate fluctuates; figures are indicative only
11	DMart store count	~350 stores (2024 annual report)	High	Publicly reported figure; may change
12	Customers per DMart store (loyalty enrolled)	5,000 (modelled estimate)	Low	DMart does not publish per-store enrolment — significant uncertainty
13	Average At-Risk revenue per customer (India)	₹8,000/year (DMart), ₹6,500/year (Reliance)	Low	Calibrated to FMCG basket data — not validated against actual retailer data
14	Demand forecast — modelled stock	4× average weekly demand (full-period baseline)	Low	If actual stock is lower, more SKUs would be flagged as at risk
15	Reorder risk threshold	< 14 days (wholesale 2-week lead time)	Medium	Lead times vary by supplier — should be calibrated in production
16	Audit trail stored in browser localStorage	No database — prototype only	High (disclosed)	Data lost on browser clear; not suitable for production
17	AI recommendations generated at build time	Claude Sonnet, 4 calls, cached system prompt	High	Static recommendations — not dynamically updated with new data

Known Limitations

—Dataset covers B2B resellers, not B2C consumers — segment dynamics and win-back rates differ
—13-month window limits cohort analysis depth and seasonal decomposition
—CLV regression R² = 0.1426 — RFM scores alone explain limited CLV variance; geography, product category, and purchase timing would improve accuracy
—Signals computed on a static snapshot — a production system would recompute on live transaction feeds
—Demand forecast uses a 4-week SMA which ignores trend and seasonality — a Q4 seasonal spike would cause under-forecasting in October/November
—Network extrapolation (DMart, Reliance Retail) relies on modelled estimates for customers-per-store and revenue-per-customer — figures are indicative and not validated against actual retailer data
—DPDP Act 2023 compliance requirements for email marketing are noted but not implemented in this prototype