Find your next learning step
A forecast vector is not coherent when its 90th percentile falls below its median. Repair the distribution first, then test whether the capacity action preserves the order that the forecast claims.
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 · Find the crossing
Route 2 · Work the repair ledger
Route 3 · Connect forecasts to actions
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · FIND THE CROSSING
1. Recognise a forecast that contradicts itself
For a fixed case, horizon and conditioning information, quantiles must be ordered: if 0.1 < 0.5 < 0.9, then Q10 ≤ Q50 ≤ Q90. A forecast with Q90 below Q50 is not an unusual distribution. It violates the definition of a quantile function.
This matters in a queue because different quantiles often drive different protection levels. A controller cannot explain why a more cautious percentile requests less capacity than a central percentile unless some separate rule overrides the forecast. First make the predictive distribution coherent; then make the action rule explicit.
CHAPTER 2 OF 12 · FIND THE CROSSING
2. See why independently fitted quantiles can cross
Teams often fit one regression for each quantile level. Sampling noise, sparse regimes, flexible models, regularisation and separate optimisation can make those fitted curves intersect. Crossing can also appear between input conditions: a model may behave sensibly for ordinary load yet cross during a rare burst where data are thin.
Do not repair a crossing by relabelling the outputs. Preserve the issued values, model version, horizon and conditioning variables. The failure is evidence about estimation and coverage, even when a post-processing step makes the delivered vector usable.
CHAPTER 3 OF 12 · WORK THE REPAIR LEDGER
3. Work a crossing-and-repair ledger
Suppose a six-minute demand forecast reports Q10 = 38, Q50 = 52, Q80 = 49 and Q95 = 71 jobs per minute. The 0.8 quantile falls below the median. A simple increasing rearrangement sorts the four values to 38, 49, 52 and 71 while retaining the multiset of predicted levels.
If one ready replica safely serves 20 jobs per minute, ceiling division maps the repaired vector to two, three, three and four replicas. The action sequence is nondecreasing. The raw sequence would have mapped Q50 and Q80 to three and three here, hiding the crossing by chance; a different threshold could have exposed a reversed action.
| Level | Raw forecast | Rearranged forecast | Rounded replicas |
|---|---|---|---|
| 0.10 | 38 | 38 | 2 |
| 0.50 | 52 | 49 | 3 |
| 0.80 | 49 | 52 | 3 |
| 0.95 | 71 | 71 | 4 |
CHAPTER 4 OF 12 · WORK THE REPAIR LEDGER
4. Compare rearrangement with isotonic projection
Increasing rearrangement sorts the estimated curve over the quantile index. Chernozhukov, Fernández-Val and Galichon show that this monotonisation can improve approximation to a monotone target under stated conditions. It is attractive because it is transparent and preserves the empirical distribution of the fitted values over the index grid.
Isotonic projection instead finds a nondecreasing sequence close to the original values under a declared loss, often squared error with optional weights. For 38, 52, 49 and 71, equal-weight isotonic regression pools the middle violation at 50.5, giving 38, 50.5, 50.5 and 71. Rearrangement and projection are therefore not interchangeable. Record which operator was used and why.
| Repair | Q50 | Q80 | What it preserves |
|---|---|---|---|
| rearrangement | 49 | 52 | sorted multiset over the grid |
| equal-weight isotonic projection | 50.5 | 50.5 | least-squares closeness under monotonicity |
CHAPTER 5 OF 12 · WORK THE REPAIR LEDGER
5. Recheck calibration after every repair
Monotonicity is a logical constraint, not proof of statistical calibration. After repair, Q80 should still be evaluated by the fraction of comparable outcomes at or below it, pinball loss, sharpness and conditional coverage across regimes. A repair can remove crossings while leaving every upper quantile biased low.
Compare raw and repaired forecasts on held-out or rolling-origin data. If repairs are frequent or large, change the estimation procedure rather than treating post-processing as harmless polish. Report crossing rate, total adjustment and the regimes in which they concentrate.
CHAPTER 6 OF 12 · CONNECT FORECASTS TO ACTIONS
6. Changed context: do not force order across horizons
Suppose Q90 is 70 jobs per minute at six minutes and 62 at twelve minutes because a scheduled event is expected to end. This decline is not a quantile crossing: the conditioning horizon changed. Quantile order is required across probability levels for the same predictive distribution, not necessarily across time horizons.
If an operator imposes “longer horizon must mean higher demand,” it can manufacture capacity that the model never claimed. Temporal coherence requires a joint scenario or declared dynamics, not a generic monotone constraint in clock time.
CHAPTER 7 OF 12 · CONNECT FORECASTS TO ACTIONS
7. Map coherent forecasts through a monotone action rule
Let a(q) be the capacity action obtained from demand quantile q. If all other state variables are fixed and more forecast demand cannot justify less service capacity, a should be nondecreasing. Ceiling division by positive per-replica capacity has that property. Caps, minimums and monotone lookup tables can preserve it too.
Hysteresis introduces state dependence: the same q may produce different actions depending on current ready and pending capacity. That does not violate monotonicity when the state is held fixed. Test the complete policy with identical state, varying only q, and log any exception such as a shared resource cap.
CHAPTER 8 OF 12 · CONNECT FORECASTS TO ACTIONS
8. Keep probability order separate from policy priority
A high quantile is not automatically the correct control input. The selected level comes from shortage and surplus consequences, service objectives, setup delay and model calibration. Once selected, it should enter a coherent vector and a traceable action mapping.
Continue through Quantile Forecasts, Asymmetric Loss and Service-Level Protection for the cost ratio, then Confidence Bands, Decision Margins and Forecast-Triggered Hysteresis for delayed control.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
Sorting four reported quantiles does not reconstruct a complete conditional distribution. Noncrossing forecasts can still have atoms, implausible tail shape or poor joint dependence across horizons. Independently repaired horizons may not form realistic sample paths.
Monotone capacity actions do not guarantee stable autoscaling, low wait or feasible service. Setup time, pending replicas, shared bottlenecks and rate limits remain. A capped action can be nondecreasing while failing every service target above the cap.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Build an evidence check
Archive raw and delivered quantile vectors, levels, horizons, conditioning features, model version, repair operator, weights and adjustment size. Store ready, pending and capped capacity plus every threshold, dwell rule and action reason.
On replay, report crossing frequency before repair; empirical coverage and pinball loss after repair; replica monotonicity tests at fixed state; service-level attainment; tail wait; starts; cancellations; cap hits; and idle-ready time. Stratify by data density, burst regime and forecast horizon.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Questions readers should ask
Can a true quantile function cross?
No. For one distribution, quantiles are nondecreasing in their probability level. Estimated curves can cross because estimation is imperfect.
Is sorting always the best repair?
No. Rearrangement, isotonic projection and jointly constrained estimation solve different optimisation problems. Choose and validate the operator rather than hiding it.
Must capacity increase with forecast horizon?
No. Demand may be expected to fall. Monotonicity across probability levels is not monotonicity across time.
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. Continue through Forecast Error, Early Starts and Activation Regret, Rate-Limited Controllers, Replica Caps and Anti-Windup Checks and Cold Starts, Readiness Gates and Capacity That Exists Only on Paper.
Primary routes include Chernozhukov, Fernández-Val and Galichon’s Quantile and Probability Curves Without Crossing, Koenker and Bassett’s Regression Quantiles, Gneiting’s Making and Evaluating Point Forecasts, and Pan and colleagues’ MagicScaler. The crossing ledger is an original teaching example.
Top · Contents · Previous · How X Works