Definition
Forecast accuracy stratified by product lifecycle stage rather than reported in aggregate. Prevents mature-product accuracy from masking forecast failure at launch, patent cliff, and post-LOE stages where the decisions are largest.
The Measurement Gap
A top-20 pharma company reports 87.2% aggregate AI forecast accuracy at the 3-month horizon, above the 85% target. Underneath that number: 94 mature products forecast at 93.1% accuracy because year-8 demand looks like year-7. Meanwhile, 11 patent cliff products forecast at 61.3% because the AI has no reliable model for generic entry timing and pricing behavior. Six launch products forecast at 54.8% because demand depends on formulary access, physician adoption curves, and competitive dynamics the model has never seen for this molecule.
The aggregate number is mathematically correct and operationally useless. Mature products that are easy to forecast drown out signal from products where accuracy matters most. A patent cliff product at 61% accuracy produces either $40 million in excess inventory or a critical supply gap. A launch product at 55% means millions spent on unnecessary safety stock or a supply disruption during the most commercially critical window.
Legacy forecast metrics compute a single number across all products. No current metric decomposes accuracy by lifecycle stage to reveal where the model fails at moments of highest financial exposure.
Formula
LSFA(s) = 1 - (1/N_s) * SUM(|forecast_i - actual_i| / actual_i)
Where s is one of: launch, growth, mature, patent_cliff, post_LOE
Worked Example:
Q4 analysis at 3-month forecast horizon:
Mature products (78 SKUs): LSFA(mature) = 0.932. MAPE 6.8%. Excellent but expected.
Growth products (14 SKUs): LSFA(growth) = 0.886. MAPE 11.4%. Reasonable with established trajectory.
Launch products (4 SKUs): LSFA(launch) = 0.587. MAPE 41.3%. Two products launched into a competitive market with demand 30-50% below forecast. One exceeded forecast by 60% due to a competitor supply disruption.
Patent cliff products (7 SKUs): LSFA(patent_cliff) = 0.628. MAPE 37.2%. First generic priced at 85% discount, demand dropped 42% in month one versus the model's gradual erosion prediction.
Post-LOE products (9 SKUs): LSFA(post_LOE) = 0.813. MAPE 18.7%. Better than patent cliff because the initial shock has passed.
Aggregate: 0.908 (MAPE 9.2%). The aggregate looks fine while the supply chain commits $180 million on patent cliff products using a model wrong by 37% on average.
Proposed formula, pending validation pilots.