Chapter 17 Prediction Intervals
Previous chapters established the directional framework for probability, dependence, and statistical inference. Chapter 13 derived conditional probability and Bayes’ theorem from co-partial moments, and Chapter 15 developed methods for comparing entire distributions without relying on parametric hypothesis tests — including the distinction between the degree-zero CDF, which counts observations, and the degree-one ratio, which measures deviation area.
A natural next step is prediction.
In statistical analysis it is often necessary to estimate the range within which future observations are likely to occur. Classical statistics addresses this problem using confidence intervals and prediction intervals. These concepts are frequently confused, but they serve fundamentally different purposes.
This chapter clarifies the distinction and develops empirical intervals and severity bands based on partial moments. Their construction requires no specified parametric family. Degree-zero quantiles target event frequency; degree-one expectiles target deviation-area balance. Claims about future or repeated-sample coverage still require sampling assumptions and an appropriate calibration procedure.
17.1 Confidence Intervals versus Prediction Intervals
A confidence interval estimates an unknown population parameter.
For example, a classical confidence interval for the mean takes the form
\[ \bar{X} \pm z_{\alpha/2} \frac{s}{\sqrt{n}}, \]
where
- \(\bar{X}\) is the sample mean,
- \(s\) is the sample standard deviation,
- \(n\) is the sample size.
This interval reflects uncertainty about the parameter \(\mu = E[X]\).
A prediction interval, by contrast, estimates the range within which a future observation will fall.
For a normally distributed population, a classical prediction interval is
\[ \bar{X} \pm t_{\alpha/2,n-1} s \sqrt{1 + \frac{1}{n}}. \]
Prediction intervals are wider because they incorporate two sources of uncertainty:
- uncertainty about the population parameters, and
- natural variability of individual observations.
In practice, however, classical prediction intervals rely heavily on distributional assumptions, particularly normality.
The directional framework allows prediction intervals to be constructed directly from the empirical distribution, avoiding these assumptions entirely.
17.2 The Discrete–Continuous Distinction: Recap and Application to Intervals
Chapter 15 (Section 15.1.8) separated two readings of the mass below a benchmark. The empirical CDF is a degree-zero lower partial moment, a step function that counts observations. Its value at the sample mean,
\[ \hat{F}_X(\bar{x}) = \frac{1}{n}\sum_{i=1}^n \mathbf{1}_{\{x_i \le \bar{x}\}}, \]
estimates \(P(X \le \mu)\), which is 0.5 for a continuous distribution symmetric about its mean, but need not be 0.5 otherwise. The degree-one partial moment ratio
\[ F_1(t; X) = \frac{LPM_1(t, X)}{LPM_1(t, X) + UPM_1(t, X)} \]
is the share of deviation area below \(t\). It is continuous in \(t\), and \(F_1(\bar{x}; X) = 0.5\) for every non-constant sample by construction.
This distinction matters for interval estimation. Coverage, the share of observations an interval contains, is a frequency, so coverage intervals invert the degree-zero CDF. Inverting the degree-one ratio gives severity-weighted thresholds. Their event-frequency tail proportions can be above or below the nominal area level, depending on the distribution.
17.3 Prediction as a Quantile Problem
Prediction intervals can be understood as quantile intervals of the underlying distribution.
Let \(X\) be a random variable with cumulative distribution function \(F_X(t)\). The \(p\)-quantile is defined as
\[ Q_X(p) = \inf \{t : F_X(t) \ge p\}. \]
A prediction interval with coverage probability \(1-\alpha\) is therefore
\[ [Q_X(\alpha/2), \, Q_X(1-\alpha/2)]. \]
For example, a 95% prediction interval corresponds to
\[ [Q_X(0.025),\, Q_X(0.975)]. \]
Classical parametric methods estimate these quantiles using assumed distributions. The directional framework estimates them directly from partial moments, with the additional capability to choose between discrete (degree-zero) and continuous (degree-one) probability representations.
Prediction intervals are one application of quantile inversion, but they are not the only one. The same inversion logic can also be used to select benchmark thresholds for directional decision analysis.
A lower-tail threshold chosen by \[ \inf\{t:F_X(t)\ge \alpha\} \] can be interpreted as the lower endpoint of a tail-quantile construction. In finance this object is often called Value-at-Risk, but the underlying mathematics is much more general. It may represent an acceptable forecast error, a reliability boundary, a minimum service level, or any other adverse threshold defined relative to a benchmark.
This distinction is conceptual rather than mathematical. Prediction intervals ask for a range likely to contain future observations. Threshold analysis asks where the lower tail begins to contain a specified fraction of directional mass. In both cases the central task is quantile inversion.
Once higher degrees are introduced, the interpretation broadens further. Degree-zero thresholds partition observations by frequency. Higher-degree thresholds partition them by severity-weighted directional mass. Thus the quantile framework supports both interval estimation and benchmark-sensitive threshold design.
17.4 Partial-Moment Quantile Functions: LPM.VaR and UPM.VaR
The NNS package provides two complementary quantile functions that invert the partial moment CDF:
LPM.VaR(percentile, degree, variable) — returns a lower-tail threshold \(t\) for the level \(p\). At degree = 0 it is the type-7 (linearly interpolated) empirical quantile. Because the empirical CDF is a step function, that quantile need not satisfy \(F_0(t; X) = p\): LPM.VaR(0.1, 0, 1:4) is 1.3, where \(F_0(1.3) = 0.25\). At positive degrees \(F_\text{degree}(t; X)\) is continuous in \(t\), and LPM.VaR returns the \(t\) with \(F_\text{degree}(t; X) = p\); \(p\) is then a share of directional mass rather than a probability (at degree 1, \(t\) is the \(p\)-expectile).
UPM.VaR(percentile, degree, variable) — returns the matching upper-tail threshold: the type-7 empirical quantile at \(1 - p\) when degree = 0, and the \(t\) with \(1 - F_\text{degree}(t; X) = p\) at positive degrees.
Both functions accept a degree argument. Setting degree = 0 returns classical (type-7) empirical quantiles. Setting degree = 1 uses the area-based ratio of Chapter 15, so \(p\) is a share of deviation area rather than of observations. Higher degrees extend the same inversion principle to severity-weighted directional probability. Thus LPM.VaR is best interpreted not as a finance-specific tool, but as a general lower-tail threshold operator generated by the partial-moment framework.
Coverage is a frequency statement, so a plug-in interval targeting coverage \(1 - \alpha\) uses degree-zero quantiles. Finite-sample future coverage still depends on the sampling design and quantile estimation:
# Plug-in interval targeting 95% coverage: degree = 0 uses frequency
lower <- LPM.VaR(percentile = 0.025, degree = 0, x = x)
upper <- UPM.VaR(percentile = 0.025, degree = 0, x = x)
# Magnitude-calibrated band: 2.5% of the area in each tail (degree = 1)
lower_1 <- LPM.VaR(percentile = 0.025, degree = 1, x = x)
upper_1 <- UPM.VaR(percentile = 0.025, degree = 1, x = x)In Python the same quantile inversions are lpm_var and upm_var:
from nns import lpm_var, upm_var
# Plug-in interval targeting 95% coverage: degree = 0 uses frequency
lower = lpm_var(0.025, 0, x)
upper = upm_var(0.025, 0, x)
# Magnitude-calibrated band: 2.5% of the area in each tail (degree = 1)
lower_1 = lpm_var(0.025, 1, x)
upper_1 = upm_var(0.025, 1, x)The degree-one curve is continuous in the benchmark, whereas the empirical degree-zero CDF is a step function. The package also interpolates degree-zero quantiles between order statistics, so continuity of a computed endpoint is not unique to degree one. Expectile bounds calibrate area rather than frequency. At \(\alpha = 0.05\), the population bands contain about 72% of uniform, 76% of exponential, and 84% of normal observations. These are examples, not a universal coverage ordering: rare large extremes can make a degree-one band wider and give coverage above the nominal level.
17.4.1 Generalized Threshold Operators
The notation LPM.VaR can appear more domain-specific than the underlying mathematics actually is. In the present framework, the function returns the benchmark \(t\) that solves a lower-tail threshold problem under a chosen directional degree.
When \(d=0\), the threshold is
\[
t_\alpha^{(0)}=\inf\{t:L_0(t;X)\ge \alpha\},
\]
which is the ordinary empirical lower quantile. LPM.VaR interpolates linearly between adjacent order statistics instead (R’s quantile(x, alpha, type = 7)), so the two agree at the order statistics and differ between them: for \(\{1, 2, 3, 4, 5, 100\}\) the lower 5% threshold is 1 by the infimum and 1.25 from LPM.VaR, and the upper 5% threshold is 100 and 76.25.
When \(d=1\), the threshold instead solves \[ t_\alpha^{(1)}=\inf\left\{t:\frac{L_1(t;X)}{L_1(t;X)+U_1(t;X)}\ge \alpha\right\}. \] This no longer counts all observations equally. Larger deviations below the benchmark contribute more heavily to the lower-tail mass. The threshold is the \(\alpha\)-expectile of \(X\) (Newey and Powell, 1987): \(L_1/(L_1+U_1)=\alpha\) is the same equation as \(\alpha\,E[(X-t)_+]=(1-\alpha)\,E[(t-X)_+]\). Expectiles are elicitable on suitable moment classes. With losses taken as positive, an upper expectile at level \(\tau \in [1/2,1)\) is a coherent risk measure; in the payoff convention the corresponding risk is the negative lower expectile at level \(1-\tau\). Coherence is not an all-level statement under an unspecified sign convention. Expectiles are not quantiles and carry no event-frequency guarantee.
Similarly, when \(d=2\), \[ t_\alpha^{(2)}=\inf\left\{t:\frac{L_2(t;X)}{L_2(t;X)+U_2(t;X)}\ge \alpha\right\}, \] so extreme adverse deviations receive quadratic weight.
These thresholds therefore define a family of directional calibration rules: \[ \text{degree 0: frequency-calibrated threshold}, \] \[ \text{degree 1: magnitude-calibrated threshold}, \] \[ \text{degree 2: extreme-deviation-calibrated threshold}. \]
This interpretation is fully general. It applies whenever one wishes to choose a threshold not only by how often a process crosses it, but also by how severely the process behaves once crossed.
17.5 Link to Stochastic Dominance (Chapter 15)
Chapters 15 and 17 use the same degree-one quantile geometry from different angles.
- In stochastic dominance (Chapter 15), the ordering is defined by integrated quantile/CDF behavior.
- In prediction intervals (this chapter), interval endpoints are degree-zero or degree-one quantiles from
LPM.VaRandUPM.VaR.
The methods are mathematically unified: the degree-one objects used to construct magnitude-calibrated bands are the same area-based objects used to diagnose dominance relations.
17.6 Comparison with Bootstrap Confidence Intervals
The difference between discrete and continuous partial moment quantiles becomes especially apparent when applied to bootstrap confidence intervals.
Consider the correlation statistic from the law dataset (Efron and Tibshirani, 1993). Standard bootstrap methods produce a range of confidence intervals depending on the method chosen:
library(bootstrap); library(boot)
data("law")
get_r <- function(data, indices, x, y) {
d <- data[indices, ]
round(as.numeric(cor(d[x], d[y])), 3)
}
set.seed(12345)
boot_out <- boot(law, x = "LSAT", y = "GPA", R = 500, statistic = get_r)
boot.ci(boot_out)
## BOOTSTRAP CONFIDENCE INTERVAL CALCULATIONS
## Level Normal Basic
## 95% ( 0.5247, 1.0368 ) ( 0.5900, 1.0911 )
## Level Percentile BCa
## 95% ( 0.4609, 0.9620 ) ( 0.3948, 0.9443 )The distribution of bootstrapped correlations is visibly asymmetric — there is a long left tail and the upper bound exceeds 1.0, which is impossible for a correlation coefficient.
17.7 Degree-Zero Partial Moment Intervals
The degree-zero LPM quantile is the percentile bootstrap interval up to the interpolation rule. The discrete CDF assigns equal weight to each bootstrap replicate; LPM.VaR interpolates linearly between order statistics while boot.ci interpolates on the normal-quantile scale, so the endpoints agree closely but not exactly (0.4684 and 0.9615 here, against 0.4609 and 0.9620 from boot.ci):
17.8 Degree-One Partial Moment Intervals
The degree-one quantile weights replicates by their distance from the boundary, so each tail beyond its bounds holds 2.5% of the replicates’ deviation area rather than 2.5% of the replicates. The band is narrower, and it is not a 95% interval:
# Area-calibrated band: 2.5% of the deviation area in each tail
LPM.VaR(percentile = 0.025, degree = 1, x = boot_out$t)
## [1] 0.5612616
UPM.VaR(percentile = 0.025, degree = 1, x = boot_out$t)
## [1] 0.913772The bootstrap comparison in Python (the 15-observation law dataset is inlined; the resampling loop replaces boot):
import numpy as np
from nns import lpm_var, upm_var
lsat = np.array([576, 635, 558, 578, 666, 580, 555, 661,
651, 605, 653, 575, 545, 572, 594.0])
gpa = np.array([3.39, 3.30, 2.81, 3.03, 3.44, 3.07, 3.00, 3.43,
3.36, 3.13, 3.12, 2.74, 2.76, 2.88, 2.96])
rng = np.random.default_rng(12345)
boot_t = np.array([
np.round(np.corrcoef(lsat[idx], gpa[idx])[0, 1], 3)
for idx in rng.integers(0, lsat.size, size=(500, lsat.size))
])
# Discrete lower and upper CI: the percentile method, up to interpolation
lpm_var(0.025, 0, boot_t), upm_var(0.025, 0, boot_t)
# Area-calibrated band: 2.5% of the deviation area in each tail
lpm_var(0.025, 1, boot_t), upm_var(0.025, 1, boot_t)The degree-one band \((0.561,\, 0.914)\) is tighter than the percentile interval \((0.461,\, 0.962)\), and its upper bound is below 1.0. The tightening comes from the change of measure: the degree-one bounds leave 2.5% of the replicates’ area, not 2.5% of the replicates, in each tail, so the band holds about 80% of the bootstrap replicates rather than 95% (see the next section). It describes where the replicates’ deviation area concentrates; it is not a bias-corrected 95% confidence interval.
Classical bootstrap corrections (BCa, studentized) adjust the percentile interval for bias and skewness through acceleration constants or additional resampling passes, and keep its 95% coverage target. The degree-one band does not attempt that correction: it changes what is measured, from a share of replicates to a share of their deviation area. Use the degree-zero (percentile) interval, or BCa, when the goal is 95% confidence, and the degree-one band to describe where the replicates’ deviation area lies.
17.9 Distribution-Free Prediction Intervals
Using the NNS quantile functions, a band is constructed as
\[ [\text{LPM.VaR}(\alpha/2,\, d,\, X),\; \text{UPM.VaR}(\alpha/2,\, d,\, X)] \]
for a chosen degree \(d \in \{0, 1\}\).
Degree 0 returns R’s type-7 interpolated empirical quantiles at levels \(\alpha/2\) and \(1-\alpha/2\). The generalized-inverse ECDF convention instead selects order statistics at ceiling ranks. Neither interpolation nor an observed sample’s tail fraction alone guarantees exact finite-sample coverage for a future observation.
Degree 1 is continuous in the data, but it calibrates area rather than frequency: each tail beyond the bounds holds \(\alpha/2\) of the degree-one partial moment mass, not \(\alpha/2\) of the observations. On uniform data the degree-one band at \(\alpha = 0.05\) contains about 72% of the observations, so it should be read as a magnitude-calibrated band, not as a \(1-\alpha\) prediction interval.
Neither calculation requires a parametric distribution or a variance estimate. This describes how the endpoints are computed; it does not establish exact predictive or confidence coverage without sampling and calibration assumptions.
More generally, one may define a lower-tail directional threshold by \[ t_\alpha^{(d)}=\mathrm{LPM.VaR}(\alpha,\text{ degree }=d,X), \] where \(d=0\) corresponds to frequency-based probability and \(d\ge 1\) corresponds to severity-weighted directional probability. The interpretation of the threshold changes with \(d\), but the computational principle remains unchanged: invert the chosen directional probability representation.
17.10 Example
Consider the fully specified sample
\[ X = -2,\,-1,\,0,\,1,\,2,\,3,\,5,\,7,\,9,\,10. \]
For a 90% prediction interval (\(\alpha = 0.10\)):
x <- c(-2, -1, 0, 1, 2, 3, 5, 7, 9, 10)
# Degree 0: classical empirical quantiles
LPM.VaR(percentile = 0.05, degree = 0, x = x)
## 5%
## -1.55
UPM.VaR(percentile = 0.05, degree = 0, x = x)
## 95%
## 9.55
# Degree 1: area-based thresholds
LPM.VaR(percentile = 0.05, degree = 1, x = x)
## [1] -0.4347826
UPM.VaR(percentile = 0.05, degree = 1, x = x)
## [1] 8.173913In Python:
import numpy as np
from nns import lpm_var, upm_var
x = np.array([-2, -1, 0, 1, 2, 3, 5, 7, 9, 10.0])
# Degree 0: classical empirical quantiles
lpm_var(0.05, 0, x), upm_var(0.05, 0, x)
# Degree 1: area-based thresholds
lpm_var(0.05, 1, x), upm_var(0.05, 1, x)The degree-zero interval is \([-1.55, 9.55]\), interpolated between the two lowest and the two highest observations. The degree-one interval \([-0.43, 8.17]\) is tighter because it leaves 5% of the deviation area in each tail rather than 5% of the observations, so it holds fewer of them. With only \(n=10\) observations, this difference is practically meaningful.
17.10.1 Worked Threshold Example: Frequency versus Severity
To see how degree changes interpretation, consider a sample \(X\) and a lower-tail calibration level \(\alpha\).
A degree-zero threshold is \[ t_\alpha^{(0)}=\mathrm{LPM.VaR}(\alpha,\text{ degree }=0,X), \] which partitions the sample by event frequency. Approximately an \(\alpha\) fraction of observations lies below the selected threshold.
Now compare this with degree-one and degree-two thresholds: \[ t_\alpha^{(1)}=\mathrm{LPM.VaR}(\alpha,\text{ degree }=1,X), \qquad t_\alpha^{(2)}=\mathrm{LPM.VaR}(\alpha,\text{ degree }=2,X). \] These thresholds are chosen in a different geometry. They do not partition simple counts. Instead they partition severity-weighted directional mass. Large adverse deviations therefore influence the threshold more strongly than small ones.
This leads to an important practical phenomenon. A severity-weighted threshold may occur at a milder benchmark than a frequency-based threshold, yet still produce a less severe set of threshold violations in realized use. The reason is that the rule is designed to trigger earlier, before the most extreme adverse deviations dominate the lower tail. The probability-bounds literature demonstrates this effect empirically in finance examples, but the mechanism is general and should be understood as an early-intervention property of higher-degree threshold calibration.
This distinction matters in many domains:
- in forecasting, a severity-weighted threshold warns before very large misses accumulate,
- in operations, it triggers replenishment before severe shortages emerge,
- in engineering, it signals intervention before large safety-margin breaches dominate the tail.
What appears paradoxical under degree-zero counting becomes natural under directional weighting. The threshold is solving a different partition problem.
17.11 Interpretation and Coverage
Prediction intervals constructed from partial moment quantiles possess a simple probabilistic interpretation.
Degree zero targets event frequency in the estimated distribution. Higher degrees target weighted adverse exposure. At the population level, continuous CDF quantiles give the stated event probabilities; estimated quantiles add sampling uncertainty and interpolation choices. Area calibration at higher degrees is a different target.
For independent identically distributed observations, with a continuous distribution and identifiable continuous quantiles at the two target levels, a plug-in degree-zero interval has the large-sample target
\[ P\bigl(\text{LPM.VaR}(\alpha/2, 0, X) \le X_{n+1} \le \text{UPM.VaR}(\alpha/2, 0, X)\bigr) \approx 1 - \alpha, \]
as the empirical quantiles converge to the population quantiles. This is an asymptotic statement under the stated conditions; ties, dependence, distribution shifts, and small samples require their own coverage analysis.
For degree \(d = 1\), each endpoint leaves \(\alpha/2\) of the deviation area on its corresponding side. Its ordinary frequency coverage depends on the distribution and can be below or above \(1-\alpha\). For example, a sample with ten values at \(-100\), 980 at zero, and ten at \(100\) has type-7 95% bounds \([0,0]\) and degree-one bounds approximately \([-27.5362,27.5362]\); the latter contain 98% of the observations. The directional bands remain useful measures of weighted exposure, with a target distinct from frequency coverage.
Neither construction requires choosing a parametric family. Valid claims about future coverage still require sampling assumptions and calibration; a distribution-free construction does not make those requirements disappear.
17.12 Residual Prediction Bands via NNS Regression
The regression estimator is developed in Chapter 22. Its fitted location function can be nonlinear and locally adaptive; the current NNS.reg interval implementation uses a global residual distribution.
NNS.reg(..., confidence.interval = ...) computes two type-8 quantiles of all in-sample residuals and adds the same offsets to every fitted or predicted value. Its band therefore has constant width across predictor locations; it does not estimate partition-specific residual distributions or automatically calibrate conditional coverage under heteroskedasticity. Local or held-out residual calibration is a separate extension.
set.seed(12345)
x_train <- runif(1000, -2, 2)
y_train <- sin(pi * x_train) + rnorm(1000, sd = 0.2)
fit <- NNS.reg(x = x_train, y = y_train, order = NULL,
point.est = c(-1, 0, 1), confidence.interval = .95)
fit$pred.int # common residual-quantile offsets around each predictionIn Python:
import numpy as np
from nns import nns_reg
rng = np.random.default_rng(12345)
x_train = rng.uniform(-2, 2, 1000)
y_train = np.sin(np.pi * x_train) + rng.normal(scale=0.2, size=1000)
fit = nns_reg(x_train, y_train, order=None, point_est=np.array([-1, 0, 1]),
confidence_interval=0.95)
fit["pred.int"] # check the port's documented residual-band implementation
NNS.reg(..., confidence.interval = 0.95) visualization with nonlinear fit and global residual-quantile bands.17.13 Advantages of the Directional Approach
Prediction intervals based on partial moments possess several advantages over classical methods.
Explicit calibration.
Degree-zero bounds calibrate frequency, which is what coverage requires; degree-one bounds calibrate deviation area, which is continuous in the benchmark and anchored at 0.5 at the mean for every nondegenerate integrable distribution (Chapter 15, Section 15.1.8). The choice is stated, not hidden in a distributional assumption.
Distribution-free construction.
No assumptions about normality or parametric form are required.
Robustness depends on the degree.
Degree-zero bounds are empirical quantiles, so an extreme observation moves them only by its rank. Degree-one bounds weight deviations by their size, as the mean does, so one extreme value moves them: for x <- seq(0, 1, length.out = 21), moving the largest value from 1 to 10 leaves the degree-zero upper 5% point UPM.VaR(0.05, 0, x) at 0.95 but moves the degree-one point UPM.VaR(0.05, 1, x) from 0.83 to 5.12. The continuity of the degree-one ratio is smoothness, not robustness.
Single pass.
Degree-zero intervals and degree-one bands both come from one sample, or one set of bootstrap replicates, with no resampling of the resamples. They do not replace BCa or double-bootstrap corrections when the goal is a bias-corrected 95% confidence interval.
Flexible location with explicit residual calibration. NNS regression adapts the fitted location to nonlinear structure. Its current global residual bands require separate validation when residual spread changes with predictor location.
Interpretability.
Intervals correspond directly to probability statements about future observations, rooted in directional partial-moment representations.
A further advantage of the directional approach is that it unifies interval estimation and threshold-based decision analysis. The analyst need not switch theories when moving from prediction intervals to adverse-threshold selection. In both cases the task is to invert a directional probability representation. The only difference is whether mass is counted equally, as in degree zero, or weighted by severity, as in higher degrees.
A second advantage is that threshold analysis can be separated from domain-specific naming conventions. Lower-tail quantiles, conditional lower-tail means, semivariance, and higher-order partial moments can all be interpreted inside a single benchmark-relative framework. The estimation-error literature supports this broader view by treating partial moments as nonparametric statistical objects with useful asymptotic behavior, rather than merely as specialized measures for one field.
17.14 Summary
Prediction intervals describe the range within which future observations are expected to occur.
Parametric, bootstrap, and other calibrated interval procedures offer different assumptions and coverage properties. A percentile bootstrap uses one collection of resamples; BCa additionally estimates acceleration, commonly by jackknife, while studentized methods can require further resampling. The directional framework represents empirical quantiles through degree-zero partial moments and severity thresholds through higher degrees. Computing these endpoints is distinct from validating their repeated-sample or future coverage.
Key results:
- The discrete CDF (\(d=0\)) is a degree-zero partial moment: it counts observations, so its quantiles give coverage intervals.
- The degree-one ratio (\(d=1\)) is the share of deviation area, continuous in the benchmark and equal to 0.5 at the mean of every non-constant sample by construction; it is not \(P(X \le t)\).
LPM.VaRandUPM.VaRinvert these CDFs to produce distribution-free quantile intervals at either degree.- Degree-one bands calibrate deviation area. Their width and event-frequency coverage relative to percentile intervals depend on the distribution; neither has a universal ordering.
NNS.regcombines a flexible fitted location with global residual-quantile offsets. Conditional heteroskedastic coverage requires additional calibration.
These ideas connect directional threshold analysis to statistical estimation. Chapter 18 develops recursive mean-split estimation, Chapter 19 interprets its effective bandwidth, and Chapter 20 develops synthetic data generation and bootstrap methods.