-
Notifications
You must be signed in to change notification settings - Fork 8
Glossary
This page defines terms that look similar but carry intentionally distinct meanings in greybox and its documentation. They are not interchangeable — each name encodes either a different surface (API argument vs. string value vs. math symbol), a different language (R vs. Python), or a different scope.
When the wiki uses a particular spelling, the spelling carries information. If you are unsure which one you mean, look it up here.
| Term | What it is | Where it appears |
|---|---|---|
distribution |
The argument of ALM / stepwise / CALM selecting the assumed error distribution. |
alm(..., distribution="dnorm") (R) / ALM(distribution="dnorm") (Python). |
"dnorm", "dlnorm", "dlaplace", … |
String values of that argument. The d-prefixed name is reused as a label — it names the distribution, not a call to a density function here. |
The distribution argument. |
dnorm(), pnorm(), qnorm(), rnorm()
|
Actual density / CDF / quantile / random functions for a distribution family. See distributions. | Called directly, e.g. qlaplace(0.9, 0, 1). |
"plogis", "pnorm" (as distribution or occurrence) |
Cumulative-function names used to denote occurrence / binary models, not continuous error distributions. |
occurrence="plogis". |
| Prefix | Meaning | Example |
|---|---|---|
d |
Density (PDF / PMF). | dgnorm() |
p |
Cumulative distribution function (CDF). | pgnorm() |
q |
Quantile function (inverse CDF). | qgnorm() |
r |
Random generation. | rgnorm() |
Note the prefix collision: "dnorm" as a distribution value (a label) is
not the same as calling dnorm() the function.
The loss argument of ALM picks the estimation objective. These are
not the same as distribution (which sets the likelihood):
| Value | Meaning |
|---|---|
"likelihood" |
Maximise the likelihood of the chosen distribution (default). |
"MSE", "MAE", "HAM"
|
Mean Squared / Absolute / Half-Absolute-Moment error. |
"LASSO", "RIDGE"
|
Penalised regression objectives. |
"ROLE", "QUALE"
|
Robust / quantile likelihood objectives (R). |
"AIC", "AICc", "BIC", "BICc" are selection criteria (used by
stepwise / CALM), not loss functions. AICc / BICc are the
small-sample-corrected variants and are the defaults.
| Term | What it is | Where |
|---|---|---|
lags (in EDA stick) |
The seasonal periodicities of a series, e.g. 12 (monthly) or c(24, 168) (hourly). |
stick(y, lags=12). |
lags (in xregExpander) |
The set of lead/lag shifts to apply to a regressor, e.g. -1:1. |
manipulations. |
| seasonal lag | General term for one repeating cycle length. | Prose. |
Arguments of measures() (see measures) — distinct roles:
| Term | What it is |
|---|---|
holdout |
The observed out-of-sample values being scored. |
forecast |
The predicted values to evaluate. |
actual |
The in-sample (training) data, used to scale errors (MASE, etc.). |
greybox results are R lists (subscripted with $) and Python objects
(attribute access with .). The wiki writes both, e.g. result$strength
(R) and result.strength (Python). The names usually match; where they
differ it is noted on the page. See R-Python-differences.
R uses camelCase functions (alm, xregExpander, outlierdummy); Python
uses snake_case functions and PascalCase classes (ALM, xreg_expander,
outlier_dummy). The full mapping is on Home under Naming Conventions.
In EDA (stick): strength is the share of total sum of squares
explained by each component. Seasonal, trend, and irregular
(the residual share) are the three STI components; their strengths sum to 1.
See Resources for the full reference list.