EXECUTIVE SUMMARY / THE DECISION IN PLAIN LANGUAGE
Lower staffing cost comes with a service decision.
Evaluated synthetic study · no actual client outcomes
The uncertainty-aware plan reduced simulated 52-week cost by $59,028 (7.0%) versus the buffer rule. It served 99.38% of demand, versus 99.83% for the buffer rule.
What a leader should do next
Set an acceptable service floor, validate shortage costs, then test the schedule in shadow mode before operational use.
What could change the decision
After a sudden sustained demand increase, the estimated savings fell to about $80 per week and the interval included zero. This is a one-site synthetic study with simplified labor constraints.
01 / BUSINESS DECISION
Start with the decision.
Lock seven days of regular staffing for a single service site, then use limited daily overtime after demand is observed. Choose a feasible plan under a weekly staffing budget and daily ramp limits, with explicit costs for labor and unserved requests.
02 / PRACTICAL IMPLICATION
What the evidence can support.
Over 52 held-out synthetic weeks, the stochastic plan costs $59,028 less than a feasible 20% buffer rule—a 7.0% reduction. Service decreases from 99.8% to 99.4%. After a sustained demand surge, the savings interval includes zero. Leaders must decide what service floor and shortage costs are acceptable before preferring the lower-cost plan.
03 / EVIDENCE
Staffing cost and service tradeoffs.
A complete rolling forecast-and-decision evaluation over 156 synthetic weeks, with 52 final test weeks. All policies obey the same daily staffing bounds, weekly worker-day budget and within-week ramp limits. The optimizer is checked against exhaustive enumeration on small cases.
Interactive results · the executive summary above remains the fixed headline comparison.
SYNTHETIC SERVICE SITE · 52 HELD-OUT WEEKS · IDENTICAL RESOURCE CONSTRAINTS
The stochastic plan lowers cost against the buffer rule, with a small reduction in demand served. The perfect-information result is an unattainable lower bound.
| Policy | 52-week cost | Demand served | Weekly savings vs buffer | 95% savings interval |
|---|---|---|---|---|
| 20% buffer rule | $848,316 | 99.83% | $0 | $0 to $0 |
| Point-forecast plan | $791,208 | 99.46% | $1,098 | $782 to $1,412 |
| Stochastic plan | $789,288 | 99.38% | $1,135 | $805 to $1,474 |
| Perfect-information bound | $764,904 | 99.40% | $1,604 | $1,223 to $1,967 |
Savings intervals use a paired four-week moving-block bootstrap over one 52-week synthetic history. The perfect-information planner sees future demand and cannot be deployed. No actual business savings are claimed.
Forecasts and service economics
The demand model’s aggregate error is 6.0%, versus 7.3% for the seasonal-naive forecast. Its nominal 80% daily ranges cover 73.6% of held-out demand, showing that uncertainty calibration still needs attention.
| Penalty per unit | 52-week cost¹ | Demand served | Regular worker-days | Overtime worker-days |
|---|---|---|---|---|
| $12 | $768,732 | 92.73% | 5,860 | 0 |
| $24 | $789,288 | 99.38% | 6,176 | 206 |
| $48 | $795,492 | 99.83% | 6,233 | 231 |
¹Each row uses a different cost definition; compare service/resource choices rather than interpreting cross-row cost differences as savings.
04 / DATA
Know where the evidence comes from.
Original synthetic daily service demand, version synthetic-capacity-v1. Demand includes weekday and annual seasonality, trend and correlated weekly shocks. There are no real staffing records. Unserved demand is lost rather than queued; staffing does not change demand. Calibration uses archived seven-day forecast-error vectors from earlier weeks.
The source includes a data dictionary, fixed configuration and seed, original data generator, observed CSVs, evaluation outputs and source/data hashes. Read the data dictionary ↗
05 / METHODOLOGY
Match the method to the design.
Fit a regularized trend/seasonal/recent-level model at each weekly cutoff. Add up to 52 past weekly error vectors to the new forecast, retaining within-week error dependence. Minimize expected labor plus shortage cost with exact dynamic programming over day, cumulative worker-days and previous-day staffing. Evaluate all feasible overtime choices as daily recourse. Compare the realized decisions with practical baselines under identical constraints.
Evaluation design
- Compare ridge demand forecasts with seasonal-naive forecasts using only information available before each week.
- Evaluate a projected 20% buffer rule, point-forecast optimization and stochastic planning on the same held-out weeks.
- Report realized cost, service, shortages, overtime and unused capacity; use a perfect-information lower bound, demand-shock tests and shortage-cost sensitivity.
References motivate the methods. These study-specific implementations do not claim every assumption or theoretical guarantee of the cited work.
06 / LIMITATIONS
Conditions that matter.
- Regular staffing is 8–26 workers daily, at most 135 worker-days weekly, and changes by at most four workers between adjacent days within a week.
- One worker supplies 12 units per day. Regular cost is $120 per worker-day; up to five overtime workers cost $180 each; unserved units cost $24 each.
- No cross-week ramp, skill mix, labor-law, absence, shift-length or multisite constraints are modeled.
- All demand and calibration errors available to a plan occur before its cutoff. The perfect-information planner is an evaluator-only benchmark.
This completed simulation does not establish actual cost savings or staffing feasibility for an employer. Marginal 80% demand intervals cover only 73.6% of stable held-out days, and shifts can degrade performance. The four-week block-bootstrap cost interval describes one synthetic history. A lower-cost plan may deliver lower service; neither forecast accuracy nor cost alone is a sufficient operating criterion.
07 / CODE & NEXT STEPS
Reproduce the result.
git clone https://github.com/mpgibb/operational-planning.git
cd operational-planning
uv sync --frozen
uv run python -m unittest discover -s tests -v
uv run python study.py
git diff --exit-code -- data resultsAutomated checks and synthetic evaluation are complete; Michael’s independent technical review is pending.