How it works
Know today which medicines will be hard to find in 90 days.
Pharmendium forecasts the availability of every medicine pack in Italy and lets buyers and sellers trade against that forecast. Here is how the forecast is made, how accurate it is, and what it can and cannot tell you.
Why forecast shortages
AIFA’s official shortage list is the source of truth, but it shows a problem that has already started. By the time a medicine is listed, alternatives are already in demand. Most shortages build up from signals that are visible earlier: a reason that historically lasts long, other packs of the same molecule already short, a manufacturer with several products affected, a seasonal pattern. Pharmendium reads those signals and turns them into a forecast you can act on.
How the forecast works
- Collect
- AIFA's official shortage list, its archive since 2018, the national registry of ~160,000 packs, and news.
- Learn
- 26,000+ past shortages show how long each kind of shortage really lasts.
- Forecast
- The chance each pack is unavailable in 30, 60, 90, 180 and 365 days.
- Act
- Buy, stock up or list stock against forecast availability, not just today's.
- For a medicine on AIFA's shortage list, an AI model compares it with thousands of past shortages that looked similar on the same day of their life: same reason, same ingredient or manufacturer, same age, same number of related packs short, same seasonal pattern. It answers directly how likely the shortage is to still be running at each horizon.
- Shortages that were never resolved are part of what it learns from. They are the longest ones, and leaving them out would make every forecast too optimistic.
- When AIFA has announced a shortage or discontinuation for a future date, the pack is shown as available until then and flagged now, not when it starts.
- Recent news about a manufacturer or active ingredient lengthens the forecast by a small, fixed amount, with the source linked on the forecast.
- For medicines that are not short, the risk is the measured rate at which similar packs have gone short before: higher for packs with a history of repeated shortages.
- The marketplace itself feeds the forecast: what sellers record in their Depot (stock on hand, lots about to expire) and what buyers order are reported as signals, so a pack that the market is running out of, or asking for, is seen before the regulator lists it.
- Every forecast lists the factors behind it, so you can see why a pack is rated the way it is.
What the risk levels mean
Each medicine gets a risk level from its chance of still being unavailable 90 days from now. The forecast is calibrated: of the packs it rated 80–90% in testing, 82% were indeed still unavailable 90 days later, so the percentages can be read literally.
| Risk level | Chance unavailable in 90 days | What it means for you |
|---|---|---|
| Critical | 80% or more | Secure stock or an alternative now. |
| High | 50–80% | More likely than not still missing next quarter: plan around it. |
| Moderate | 25–50% | Probably back within the quarter, but worth watching. |
| Low | Under 25% | No action needed. |
How accurate it is
We tested nine forecasting models the way they would be used for real: each learned only from shortages that began between 2018 and 2024, then forecast the 1,887 shortages that began in 2025, and was scored against what actually happened. The model with the best results on every measure powers Pharmendium.
| Model | Type | Ranking accuracy ↑ | Forecast error ↓ | Avg. gap to reality ↓ |
|---|---|---|---|---|
| TabICLv2, one forecast per horizonIn use | AI foundation model | 0.676 | 0.204 | 1.9 pts |
| Gradient-boosted survival model | Statistical | 0.656 | 0.216 | 5.0 pts |
| Random survival forest | Statistical | 0.643 | 0.219 | 4.3 pts |
| TabICLv2, duration estimate | AI foundation model | 0.630 | 0.218 | 2.9 pts |
| TabPFN-3.5, duration estimate | AI foundation model | 0.620 | 0.249 | 15.7 pts |
| SurvPFN | AI survival model | 0.612 | 0.226 | 2.2 pts |
| XGBoost survival | Statistical | 0.589 | 0.247 | 11.4 pts |
| Typical duration by reason | Rule of thumb | 0.575 | 0.230 | 2.1 pts |
| SurvivalPFN | AI survival model | 0.560 | 0.232 | 4.9 pts |
- Ranking accuracy: how often the model correctly tells which of two shortages will last longer (0.5 is a coin toss).
- Forecast error: how far the 90-day probability was from what happened, averaged over all shortages (Brier score).
- Avg. gap to reality: difference in percentage points between the predicted and the observed share of shortages still running at 90 days.
The chosen model also stays accurate for shortages that have already lasted months, which make up most of the current list. The model it replaced underestimated how long those would last by 10 to 15 points.
Built for buyers and sellers
Data sources
Forecasts are built from public, official data, refreshed on every update: AIFA’s live shortage list; AIFA’s archive of past lists since 2018, reconstructed into 26,236 shortage episodes; the national registry of every authorized medicine pack (~160,000); and public news coverage of manufacturers and ingredients. The model runs on Pharmendium’s own infrastructure: no shortage or customer data is sent to an outside AI service.
Limitations
- A forecast is a probability, not a guarantee. Shortage duration is only partly predictable from public data: the model ranks two shortages correctly about two times in three.
- Data is Italian and regulatory: AIFA's lists and registry. Stock levels, production sites and supplier networks are not yet visible to the model.
- News coverage is partial. A linked article is a real signal; finding none does not mean a medicine is safe.