SNF Census Forecasts: Reconcile Arrivals, Departures and Capacity Scenarios

A forecast earns a place in a purchasing decision when it changes a concrete choice: whether to extend admissions coverage, prepare for a census decline or investigate a capacity bottleneck. A smoothly rising occupancy chart does not establish that value. The buyer needs the underlying movements, assumptions and history of missed predictions.

CMS's October 2026 nursing-home dictionary distinguishes federally certified beds from average daily residents derived from MDS data. Those public fields help describe a market, but neither establishes tomorrow's usable beds or today's referrals. Keep their dates and definitions separate from internal operating data. CMS data dictionary.

Reconcile the starting point

Choose an observation time and define which residents count. Reconcile admission, discharge, hospital-transfer and return events against that definition. A hospital leave may affect physical occupancy differently from a reserved bed; the forecast must not silently treat both as the same stock.

Use resident-level expected movements for the immediate horizon where available, then aggregate them. Separate confirmed arrangements, tentative plans and historically estimated arrivals. Remove duplicate referrals before applying any arrival probability. An admission already included as confirmed cannot also remain in the weighted pipeline.

Clinical and facility leaders supply applicable capacity constraints. The forecasting model represents those decisions; it does not establish staffing adequacy or override an admission assessment.

Work a seven-day example

This synthetic case starts with 84 residents. Each closing census becomes the following day's opening census. All numbers are illustrative, not market benchmarks.

The arithmetic is 84 + 20 arrivals − 17 departures = 87. Now test two assumptions. A low-arrival case with 14 arrivals and the same departures ends at 81. A delayed-departure case with 20 arrivals and only 12 departures ends at 92. These are conditional scenarios, not statistical confidence intervals.

If the operational ceiling is 90, the second scenario requires a day-by-day review of which arrivals can actually occur. Do not clip 92 to 90 and erase the two unresolved placements. Show unconstrained demand, accepted arrivals and remaining demand separately. A room constraint can bind before the facility-wide total does.

Backtest the decision horizon

Save the input snapshot and model version when the forecast is issued. For four weekly forecasts of 87, 88, 86 and 90, suppose actual closing census is 85, 89, 84 and 88. Errors, defined as forecast minus actual, are +2, −1, +2 and +2. Mean absolute error is 7/4 = 1.75 residents; average signed error is +1.25. The model tends to overpredict in this small example.

Compare those results with a dated baseline using the same weeks. Four observations cannot establish lasting accuracy. Review longer periods, weekday patterns and disruption cases, and keep a forecast made seven days ahead separate from one revised the previous evening.

Require a reproducible forecast

Ask a platform supplier to replay a historical date without using later information. Trace one revised discharge date through the forecast, then restore the prior version. Check whether a closed unit changes usable capacity without rewriting the historical certified-bed count. Export the assumptions and prediction history so finance can independently reproduce the calculation.

Purchase against the decisions the model supports: a visible scenario range, explainable movement counts and backtests at the required horizon. Ease's skilled nursing offering is planned; any forecasting workflow would need to be evaluated against these requirements as its scope becomes available.