Methodology

What Confluence is

Confluence forecasts demand and capacity across Jordan’s medical sector and surfaces, for each actor in that sector, the decisions that need to be made this week, this quarter, this year. It is not a clinical decision support system. It does not make patient-level recommendations. It tells inventory managers, operations teams, governorate directors, and ministry strategists what their corner of the sector is about to need — and what other corners are about to need — in time to act.

The models we use

Three forecasting models, used together:

• SARIMAX — Seasonal ARIMA with Ramadan and Eid as exogenous regressors. The workhorse. Picked over deep models because hospital signals are dense, mostly weekly-seasonal, and easy to interpret. We add Ramadan and Eid as features because they shift demand by 10–25% in Jordan and ignoring them would invalidate every forecast.
• Prophet — used as a sanity check on long-horizon trend. We prefer SARIMAX’s tighter confidence intervals at 28-day horizon but cross-check against Prophet’s decomposition before publishing.
• 28-day SMA — the “dumb baseline.” If SARIMAX can’t beat a simple moving average on a series, that series is too noisy or too sparse to forecast and we say so on the screen rather than hiding it.

How we score honesty

Every forecast we display ships with its MAPE (mean absolute percentage error) on rolling-origin backtest — usually 8–15% for high-volume series, 20–40% for sparse ones. We show the number. We don’t round it down. The /admin → model metrics page lets any auditor compare SARIMAX vs SMA vs Prophet per series.

What Confluence is not

• Not a clinical decision-support system. Don’t use it to decide what to transfuse for a patient.
• Not real Hakeem data. Every number is synthesised using the patterns Hakeem records but no real patient or facility-private data has touched this system.
• Not an “AI system” in the buzzword sense. It is interpretable statistical forecasting plus rule-based decision logic. We chose this on purpose, because the audience includes clinicians who will (correctly) ask “but why does the model say that?”
• Not certified for PDPL compliance. We follow its principles — data minimisation, no patient identifiers, audit logs — but no third-party audit has been conducted.

Honest limitations

The synthetic data covers 10 hospitals + 30 PHCs/pharmacies/labs. Real deployment would extend to the ~120-facility national network with no architectural change but real data quality problems we haven’t solved yet. Some of the “districts” in the heatmap are synthetic placeholders flagged as such in the database. Cold-start facilities (Jerash, Princess Basma Maan) don’t have enough history for SARIMAX and fall back to a national-pooled forecast scaled by bed count — we widen confidence intervals to ±35% to flag this visually.