FORECASTING UK BRANDED MEDICINES SPEND UNDER VPAG: A MACHINE-LEARNING MODEL INTEGRATING NHS PRESCRIBING AND COMMERCIAL SALES DATA
Author(s)
Ben Richardson, Consulting1, Ioannis Katsoulis, Consulting1, Ben Wetherell, Consulting1, Guy Phillips, BASc2, Kim Assender, BASc2.
1CF, London, United Kingdom, 2Association of the British Pharmaceutical Industry, London, United Kingdom.
1CF, London, United Kingdom, 2Association of the British Pharmaceutical Industry, London, United Kingdom.
OBJECTIVES: UK NHS branded medicines spend grew from £13bn to £16bn (2019-2024, 4.7% CAGR), and rebate rates under the Voluntary Scheme for Branded Medicines Pricing, Access and Growth (VPAG) create commercial uncertainty for manufacturers. We developed a predictive model to forecast 12-month spend growth at product, BNF chapter and total-market levels to support rebate planning and portfolio benchmarking.
METHODS: NHS prescribing datasets (EPD, HPDC, SCMD, PCHC; ~1.5bn rows) were integrated with commercial sales data (29,358 products, 2019-2025) and mapped to BNF codes. Products were classified newer/older using loss-of-exclusivity (LOE) dates, and an actual-cost proxy for England’s Department of Health and Social Care measured sales was derived from BNF section-level implied discounts. A gradient-boosting model (XGBoost) trained on 2016-2024 data (~1,100 products with ≥£100k average monthly spend, ~95% of market revenue) predicted a log-transformed 12-month forward growth ratio from features capturing momentum, product type, competition, loss of exclusivity (LOE) timing, market positioning, pricing and healthcare activity. A 2-year buffer period separated training and prediction; performance was validated on Nov 2024-Oct 2025.
RESULTS: The model predicted aggregate branded medicines spend to within ~1% over 12 months (R²≈0.95). Product-level R² was ~0.5, median absolute percentage error (APE) 10-15%, directional accuracy >70%. Predictability varied by lifecycle stage: pre-LOE products were most accurate (median APE 8.2%); the 0-24-month post-LOE window least predictable (12.9%). Forecast error for the top-50 newer products was £50m against £6.5bn actual spend. Momentum contributed 32% of feature importance, product type 17%, competition 13%.
CONCLUSIONS: A gradient-boosting model integrating NHS and commercial data forecasts UK branded medicines spend to ~1% aggregate accuracy, providing a more robust and granular understanding of ongoing growth dynamics informed by the latest data. Quarterly retraining and iterative refinement of LOE assumptions will further improve product-level accuracy.
METHODS: NHS prescribing datasets (EPD, HPDC, SCMD, PCHC; ~1.5bn rows) were integrated with commercial sales data (29,358 products, 2019-2025) and mapped to BNF codes. Products were classified newer/older using loss-of-exclusivity (LOE) dates, and an actual-cost proxy for England’s Department of Health and Social Care measured sales was derived from BNF section-level implied discounts. A gradient-boosting model (XGBoost) trained on 2016-2024 data (~1,100 products with ≥£100k average monthly spend, ~95% of market revenue) predicted a log-transformed 12-month forward growth ratio from features capturing momentum, product type, competition, loss of exclusivity (LOE) timing, market positioning, pricing and healthcare activity. A 2-year buffer period separated training and prediction; performance was validated on Nov 2024-Oct 2025.
RESULTS: The model predicted aggregate branded medicines spend to within ~1% over 12 months (R²≈0.95). Product-level R² was ~0.5, median absolute percentage error (APE) 10-15%, directional accuracy >70%. Predictability varied by lifecycle stage: pre-LOE products were most accurate (median APE 8.2%); the 0-24-month post-LOE window least predictable (12.9%). Forecast error for the top-50 newer products was £50m against £6.5bn actual spend. Momentum contributed 32% of feature importance, product type 17%, competition 13%.
CONCLUSIONS: A gradient-boosting model integrating NHS and commercial data forecasts UK branded medicines spend to ~1% aggregate accuracy, providing a more robust and granular understanding of ongoing growth dynamics informed by the latest data. Quarterly retraining and iterative refinement of LOE assumptions will further improve product-level accuracy.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
PT31
Topic
Health Policy & Regulatory, Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas