Find your next learning step
A separate 90th-percentile forecast for every horizon does not describe one possible future. Capacity planning needs whole demand paths whose timing, persistence and ramp constraints can coexist.
Open the chapter contents · Queueing Theory root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The chapters remain open on the page; these grouped routes are an index.
Route 1 · Separate marginals from paths
Route 2 · Work the scenario ledger
Route 3 · Build feasible actions
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · SEPARATE MARGINALS FROM PATHS
1. Marginal quantiles are not a time path
A forecaster may provide the 50th and 90th percentiles of queue demand for 09:00, 09:15 and 09:30. Each column describes a marginal distribution at one horizon. The columns do not say which values occur together. Pairing every 90th percentile produces a severe-looking path, but it need not have probability ten per cent and may not resemble any observed episode.
For operational planning, a scenario is an ordered vector such as (12,18,25) jobs. Cross-horizon dependence describes whether a high value now tends to persist, reverse or arrive later. Marginal calibration and noncrossing quantiles are necessary, but neither supplies that joint structure.
CHAPTER 2 OF 12 · SEPARATE MARGINALS FROM PATHS
2. Separate three kinds of coherence
Distributional coherence requires lower quantiles not to exceed higher quantiles at one horizon. Temporal coherence requires a collection of scenarios to preserve credible dependence across horizons. Operational coherence requires the resulting capacity path to obey readiness delays, ramp limits, minimum run times and replica bounds.
A forecast can pass one layer and fail another. Sorted quantiles can be marginally coherent while independently sampled horizons create implausible zigzags. A credible demand scenario can still demand an impossible capacity jump. Keep the three checks in separate columns of the evidence ledger.
CHAPTER 3 OF 12 · WORK THE SCENARIO LEDGER
3. Work a four-scenario dependence ledger
Suppose four equally weighted demand scenarios cover three fifteen-minute horizons. Ensemble A contains persistent paths; ensemble B uses the same values at each horizon but changes how they are paired.
| Scenario | Persistent ensemble A | Re-paired ensemble B |
|---|---|---|
| 1 | 8, 9, 10 | 8, 17, 10 |
| 2 | 10, 11, 12 | 10, 9, 20 |
| 3 | 16, 17, 18 | 16, 21, 12 |
| 4 | 20, 21, 20 | 20, 11, 18 |
At horizon one both ensembles contain 8, 10, 16 and 20; the same equality holds for the later columns. Any marginal quantile table is therefore identical. Yet A represents persistent load, whereas B contains sharp reversals. A staffing plan sensitive to duration and ramps will value the two ensembles differently.
CHAPTER 4 OF 12 · WORK THE SCENARIO LEDGER
4. Reorder calibrated samples rather than inventing pairings
Ensemble copula coupling is one way to combine calibrated univariate predictive distributions with a dependence template. Draw or choose calibrated values at each horizon, then reorder them to match the rank pattern of a raw multivariate ensemble. The operation preserves each calibrated marginal sample while importing a rank-dependence structure.
The template is evidence, not truth. A small or biased raw ensemble can carry the wrong persistence, and ranks do not restore tail dependence that was never represented. Compare dependence diagnostics and operational scores before and after coupling.
CHAPTER 5 OF 12 · WORK THE SCENARIO LEDGER
5. Turn scenarios into feasible capacity paths
Let c_t be ready capacity and let a scenario prescribe demand d_t. A simple path constraint might require 0≤c_t≤24 and |c_t−c_{t−1}|≤4 units per interval. If c_0=12, the path 12,20,24 is infeasible because the first jump is eight.
One feasible repair is 16,20,24. It respects the ramp but leaves four units of demand uncovered at the first horizon. Starting capacity earlier may avoid that shortage at an idle-cost penalty. The decision is therefore a constrained optimisation across the whole path, not a separate quantile lookup at each time.
| Horizon | Demand scenario | Naive capacity | Feasible capacity | Uncovered demand |
|---|---|---|---|---|
| 1 | 20 | 20 | 16 | 4 |
| 2 | 22 | 22 | 20 | 2 |
| 3 | 24 | 24 | 24 | 0 |
CHAPTER 6 OF 12 · BUILD FEASIBLE ACTIONS
6. Changed context: dependence reverses the preferred action
Suppose two scenarios have the same high marginal demand at the middle horizon. In one, the spike persists for three intervals; in the other, it is followed by a rapid fall. With a minimum run time, starting a slow resource is sensible for the persistent case and wasteful for the transient case.
This is the transfer check: hold marginal quantiles fixed and change only the scenario ordering. If the recommended action never changes, the planner may be ignoring cross-horizon dependence. If it changes wildly, test whether the dependence estimate is stable enough to support that sensitivity.
CHAPTER 7 OF 12 · BUILD FEASIBLE ACTIONS
7. Use scenario coverage, not a mythical percentile path
Choose a collection of scenarios and evaluate cost, shortage, delay and service outcomes across them. A chance constraint may require the capacity policy to meet a target in at least a stated proportion of weighted scenarios. A robust policy may protect against a defined ambiguity set rather than every imaginable path.
Do not call the componentwise 90th-percentile vector a “90% path.” Simultaneous coverage depends on dimension and dependence. Estimate pathwise coverage directly on held-out trajectories, including peak duration and timing errors.
CHAPTER 8 OF 12 · BUILD FEASIBLE ACTIONS
8. Connect forecast verification to queue outcomes
Marginal scores test each horizon; multivariate scores and rank diagnostics test joint structure. Queue outcomes add another layer: simulated waiting tails, abandonment, overload duration, cold starts and activation regret under the proposed policy.
Use rolling-origin evaluation so every forecast uses only information available at issue time. Compare an independent-horizon baseline, the raw ensemble, the coupled scenarios and at least one stress dependence. A method that improves a statistical score but worsens service outcomes needs explanation before deployment.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
A copula or rank template does not repair biased marginals. More scenarios do not guarantee better tail coverage if they repeat the wrong dependence. A feasible capacity path does not prove queue stability, because service times, retrials, priorities and feedback can still dominate.
Quantiles at different horizons are not ordered by time: tomorrow’s median may be lower than today’s. Noncrossing applies across probability levels at a fixed horizon. Scenario weights, operational costs and constraints must be declared rather than hidden inside one recommended curve.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Build an evidence check
Archive issue times, horizons, marginal distributions, calibration windows, raw scenarios, reordering rules, scenario weights and capacity constraints. Verify marginal coverage, quantile order, rank correlations, transition distributions, peak duration and simultaneous path coverage on held-out periods.
Re-run the capacity decision under independent, persistent and adverse dependence templates. Record action changes, infeasibility, shortage and idle cost. Preserve the exact scenario set that supported each decision so later reviewers can reproduce the path calculation.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Questions readers should ask
Do marginal quantiles determine a joint forecast?
No. Many dependence structures share the same marginal distributions and can imply different peak duration and capacity decisions.
Is the vector of 90th percentiles a 90% scenario?
Not generally. Its simultaneous probability depends on cross-horizon dependence and must be measured, not assumed.
Does ensemble copula coupling create the correct dependence?
It transfers a dependence template. Its quality depends on whether that template represents the future conditions of interest.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Continue through verified routes
Begin with the Queueing Theory root, Gold Standard of Education and How X Works. Build the marginal layer with Quantile Forecasts, Asymmetric Loss and Service-Level Protection and Quantile Crossing, Coherent Forecasts and Monotone Capacity Actions. Connect action feasibility through Rate-Limited Controllers, Replica Caps and Anti-Windup Checks, Cold Starts, Readiness Gates and Capacity That Exists Only on Paper and Forecast Error, Early Starts and Activation Regret.
Primary reading includes Schefzik, Thorarinsdottir and Gneiting’s ensemble copula coupling framework, Chernozhukov, Fernández-Val and Galichon’s rearrangement of quantile curves, and Gneiting and Ranjan’s assessment of probabilistic forecasts. The scenario and capacity ledgers are original teaching examples.
Top · Contents · Previous · How X Works