Lessons In Forecasting Discipline From A $40B Supply Chain
By Eshaan Jain, Senior Consultant, Mphasis
At Amazon, I spent years on a supply chain contract portfolio worth more than $40 billion a year in last-mile logistics. The forecasts that held up under pressure were the ones built on data that was actually usable and current, regardless of which model sat behind them.
Clinical trial supply planning has a version of the same problem, and a widely cited McKinsey & Company estimate, reported by Clinical Leader and Suvoda, puts investigational medicinal product (IMP) waste from poor forecasting and planning as high as 50% of a trial's supply.1 2 Most of that waste gets blamed on the forecasting model. In my experience, the bigger driver is upstream: the clinical decisions that change supply and demand happen on a different clock than the supply plan does.
I have never built a demand forecasting model for a clinical trial. But the pattern I saw at Amazon, where forecast accuracy depended more on how current the inputs were than on which algorithm processed them, shows up here in a specific and measurable way: protocol amendments, enrollment pace, site activation timing, and treatment duration all move supply and demand, and most supply teams find out about each one after the decision has already been made.
Where The Forecast Actually Breaks
A demand forecast for a clinical trial pulls from several sources: the protocol and its amendments, site activation schedules, enrollment projections, randomization ratios, and depot inventory positions inside an interactive response technology (IRT) system. Each of those sources is usually owned by a different team, updated on a different schedule and reconciled by hand.
At Amazon, we hit a version of this same wall with supply contracts locked inside PDFs across a portfolio too large for manual review. We built a machine learning system that extracted clause-level data from those contracts with 95% accuracy, turning an unusable archive into a queryable data set. Forecasting and downstream exception handling improved immediately, not because the forecasting logic changed but because the inputs stopped being stale.
The Clinical Decisions Nobody Tells The Forecast About
Protocol amendments are the clearest example, and they are far more common than most supply teams plan for. Tufts Center for the Study of Drug Development found that 76% of Phase 1 through 4 trials now require at least one amendment, up from 57% in 2015, with oncology trials reaching 90%.3 Each amendment costs an average of $141,000 to $535,000 in direct expenses, takes roughly 260 days from initiation to final approval, and leaves sites operating under different protocol versions for around 215 days while it works through the system.3 About 23% of amendments are potentially preventable with better initial protocol planning, which means most are not, and supply teams have to absorb them as they come.3
A protocol amendment that adds a cohort, changes a dosing schedule, or expands eligibility criteria changes exactly how much investigational product a trial needs and when. That change is first and foremost a clinical and regulatory decision. It reaches the supply forecast last, often only after someone notices the numbers no longer match what shipping is.
Enrollment pace and site activation timing create the same kind of gap in the opposite direction. A site that activates six weeks late does not need its full allocation on the original schedule. A trial enrolling faster than projected needs replenishment sooner than the standing reorder cycle assumes. Treatment duration changes, common in trials that extend the dosing period based on interim safety or efficacy data, shift the total volume needed per patient without necessarily changing the patient count. None of these are supply chain decisions. All of them are supply chain inputs, and in most organizations, nobody owns the job of feeding them into the forecast the same day they are made.
The Cost Of Getting This Wrong
The market context makes the stakes clear. The clinical trial supply market is projected to grow at roughly 5.8% annually from 2023 through 2030, according to Suvoda's analysis of industry data, which means the waste problem scales right alongside the spend.4 Suvoda also reported a trial that moved from static to dynamic IRT configuration, adjusting supply parameters based on real enrollment and site data instead of fixed assumptions, saving roughly $6 million, about 25% of that trial's total supply budget.5
Layer the amendment cost on top of that. A trial averaging even one amendment at the low end of Tufts' range, $141,000, is absorbing a cost that a forecast tied to protocol and enrollment data in real time would have anticipated, rather than discovering after the fact.
3 Steps To Get It Right
Three moves matter more than the forecasting algorithm itself.
First, put the supply chain in the protocol amendment review, not after the amendment is approved. If a change to dosing, eligibility, or cohort structure is being discussed, the forecast should be updated the same week, not the same quarter.
Second, pick one system of record for enrollment and site status, and make every other system read from it instead of maintaining a parallel copy. If your IRT and your clinical trial management system (CTMS) disagree about how many patients are active at a site, your forecast is a guess.
Third, treat protocol amendments as structured events with a defined supply impact field, not PDFs that get read and manually keyed in weeks later. An amendment that changes dosing frequency or adds a cohort should update the forecast the same day it is approved.
A Worked Example
Pull three numbers on the same day: the active patient count from the IRT, the enrolled patient count from the CTMS, and the most recent approved protocol amendment's supply impact, if one exists. If the IRT says 340 active patients, the CTMS says 328, and an amendment approved three weeks ago added a cohort that neither system reflects, you already know the source of your next forecast error before it happens. That gap is a specific, traceable data lag that can be fixed in a week rather than a quarter.
Run this test monthly, publish the gap size next to the forecast, and treat a shrinking gap as the real measure of forecasting maturity. The algorithm underneath can stay the same the entire time.
Check Your Forecast With The Tools You Already Have
Before adding another forecasting tool to the stack, run one test: pull the current protocol version, the current enrollment count, and the current site activation status, and check whether all three match what your supply forecast assumed when it last ran. If they do not, the next dollar belongs to closing the timing gap between clinical decisions and supply planning, not to a better algorithm.
References:
- Clinical Leader, "Digitalized Drug Forecasting Minimizes Waste In Clinical Trial Supply Chain," citing McKinsey & Company: www.clinicalleader.com/doc/digitalized-drug-forecasting-minimizes-waste-in-clinical-trial-supply-chain-0001
- Suvoda, "Strategies to Optimize Clinical Supply," citing McKinsey & Company (November 2021): www.suvoda.com/insights/blog/strategies-to-optimize-clinical-supply
- Tufts Center for the Study of Drug Development, protocol amendment frequency, cost, and timeline research, cited by Precision For Medicine, "The Amendment Trap: Why 76% Of Clinical Trials Face Six-Figure Protocol Changes": www.precisionformedicine.com/blog/the-amendment-trap-why-76-of-clinical-trials-face-six-figure-protocol-changes
- Suvoda, "Strategies to Optimize Clinical Supply," industry market growth analysis, 2023 to 2030: www.suvoda.com/insights/blog/strategies-to-optimize-clinical-supply
- Suvoda, "How to Cut Drug Supply Costs With Seamless Forecasting and IRT Systems," dynamic IRT case study: www.suvoda.com/insights/blog/how-to-cut-drug-supply-costs-with-seamless-forecasting-and-irt-systems
About The Author:
Eshaan Jain is a senior product consultant at Mphasis and serves as the lead product owner for Salesforce/Vlocity CPQ and CLM at T-Mobile, engaged through Mphasis’s consulting services. He previously worked as a senior technical program manager at Amazon, where he co-built a machine learning system that extracted clause-level data across a $40 billion annual supply chain contract portfolio with 95% accuracy. He is an IEEE senior member and holds professional membership with Forbes Tech Council, ACM, IEEE, Isaca, and AAAI.