,
Florian Ziel [aut] | Title: | Probabilistic Forecast Combination Using CRPS Learning |
|---|---|
| Description: | Combine probabilistic forecasts using CRPS learning algorithms proposed in Berrisch, Ziel (2021) <doi:10.48550/arXiv.2102.00968> <doi:10.1016/j.jeconom.2021.11.008>. The package implements multiple online learning algorithms like Bernstein online aggregation; see Wintenberger (2014) <doi:10.48550/arXiv.1404.1356>. Quantile regression is also implemented for comparison purposes. Model parameters can be tuned automatically with respect to the loss of the forecast combination. Methods like predict(), update(), plot() and print() are available for convenience. This package utilizes the optim C++ library for numeric optimization <https://github.com/kthohr/optim>. |
| Authors: | Jonathan Berrisch [aut, cre] ,
Florian Ziel [aut] |
| Maintainer: | Jonathan Berrisch <[email protected]> |
| License: | GPL (>= 3) |
| Version: | 1.3.3 |
| Built: | 2024-11-21 05:37:43 UTC |
| Source: | https://github.com/berrij/profoc |
Use multiple online-aggregation algorithms to combine probabilistic forecasts using CRPS Learning as described in Berrisch, Ziel: "CRPS Learning", 2021. The primary function of this package is called online.
doi:10.1016/j.jeconom.2021.11.008
Index of help topics:
autoplot.batch Autoplot method for batch models
autoplot.online Autoplot method for online models
batch Probabilistic Forecast Combination - Batch
conline Create an conline Object from the conline C++
Class
init_experts_list Create experts list to be used in conline class
make_basis_mats Create a List of Basis Matrices
make_hat_mats Create a List of Hat Matrices
make_knots Create a vector of knots for splines
online Probabilistic Forecast Combination - Online
oracle Probabilistic Forecast Combination - Oracle
penalty B-Spline penalty
plot.batch Plot method for batch models
plot.online Plot method for online models
post_process_model Post Process Data from conline Class
predict.online Predict method for online models
print.batch Print method for batch models
print.online Print method for online models
profoc-package Package Info
splines2_basis Create B-Spline basis
summary.online Summary method for online models
tidy.online.experts_loss
Tidy the Experts' losses of an Online object
tidy.online.forecaster_loss
Tidy the Experts' losses of an Online object
tidy.online.predictions
Tidy the Predictions of an Online object
tidy.online.weights Tidy the Weights of an Online object
update.online Update method for online models
Maintainer: Jonathan Berrisch mailto:[email protected]
Co-Author: Florian Ziel
Berrisch, Ziel: "CRPS Learning", 2021
doi:10.1016/j.jeconom.2021.11.008
Source Code: https://github.com/BerriJ/profoc
BugReports: https://github.com/BerriJ/profoc/issues
Plots the most recent weights in each quantile using ggplot2.
## S3 method for class 'batch' autoplot(object, ...)## S3 method for class 'batch' autoplot(object, ...)
object |
Object of class inheriting from 'batch' |
... |
further arguments are ignored |
Plots the most recent weights in each quantile using ggplot2.
## S3 method for class 'online' autoplot(object, ...)## S3 method for class 'online' autoplot(object, ...)
object |
Object of class inheriting from 'online' |
... |
further arguments are ignored |
Returns predictions and weights calculated by sequential numeric optimization. The optimization is done stepwise, always calculating a one-step-ahead forecast.
![[Experimental]](data:image/svg+xml;base64,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)
batch( y, experts, tau = 1:dim(experts)[2]/(dim(experts)[2] + 1), affine = FALSE, positive = FALSE, intercept = FALSE, debias = TRUE, lead_time = 0, initial_window = 30, rolling_window = initial_window, loss_function = "quantile", loss_parameter = 1, qw_crps = FALSE, b_smooth = list(knots = length(tau), mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, periodic = FALSE), p_smooth = list(knots = length(tau), mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, ndiff = 1.5, lambda = -Inf, periodic = FALSE), forget = 0, soft_threshold = -Inf, hard_threshold = -Inf, fixed_share = 0, parametergrid_max_combinations = 100, parametergrid = NULL, forget_past_performance = 0, allow_quantile_crossing = FALSE, trace = TRUE )batch( y, experts, tau = 1:dim(experts)[2]/(dim(experts)[2] + 1), affine = FALSE, positive = FALSE, intercept = FALSE, debias = TRUE, lead_time = 0, initial_window = 30, rolling_window = initial_window, loss_function = "quantile", loss_parameter = 1, qw_crps = FALSE, b_smooth = list(knots = length(tau), mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, periodic = FALSE), p_smooth = list(knots = length(tau), mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, ndiff = 1.5, lambda = -Inf, periodic = FALSE), forget = 0, soft_threshold = -Inf, hard_threshold = -Inf, fixed_share = 0, parametergrid_max_combinations = 100, parametergrid = NULL, forget_past_performance = 0, allow_quantile_crossing = FALSE, trace = TRUE )
y |
A numeric matrix of realizations. In probabilistic settings a matrix of dimension Tx1, in multivariate settings a TxD matrix. In the latter case, each slice of the expert's array gets evaluated using the corresponding column of the y matrix. |
experts |
An array of predictions with dimension (Observations, Quantiles, Experts). |
tau |
A numeric vector of probabilities. |
affine |
Defines whether weights are summing to 1 or not. Defaults to FALSE. |
positive |
Defines if a positivity constraint is applied to the weights. Defaults to FALSE. |
intercept |
Determines if an intercept is added, defaults to FALSE. If true, a new first expert is added, always predicting 1. |
debias |
Defines whether the intercepts weight is constrained or not. If TRUE (the default), the intercept weight is unconstrained. Only affects the results if affine and or positive is set to TRUE. If FALSE, the intercept is treated as an expert. |
lead_time |
offset for expert forecasts. Defaults to 0, which means that experts forecast t+1 at t. Setting this to h means experts predictions refer to t+1+h at time t. The weight updates delay accordingly. |
initial_window |
Defines the size of the initial estimation window. |
rolling_window |
Defines the size of the rolling window. Defaults to the value of initial_window. Set it to the number of observations to receive an expanding window. |
loss_function |
Either "quantile", "expectile" or "percentage". |
loss_parameter |
Optional parameter scaling the power of the loss function. |
qw_crps |
Decides whether the sum of quantile scores (FALSE) or the quantile weighted CRPS (TRUE) should be minimized. Defaults to FALSE. Which corresponds to Berrisch & Ziel (2021) |
b_smooth |
A named list determining how the B-Spline matrices for probabilistic smoothing are created. Default corresponds to no probabilistic smoothing. See details. |
p_smooth |
A named list determining how the hat matrices for probabilistic P-Spline smoothing are created. Default corresponds to no smoothing. See details. |
forget |
Adds an exponential forgetting to the optimization. Past observations will get less influence on the optimization. Defaults to 0, which corresponds to no forgetting. |
soft_threshold |
If specified, the following soft threshold will be applied to the weights: w = sgn(w)*max(abs(w)-t,0) where t is the soft_threshold parameter. Defaults to -inf, which means that no threshold will be applied. If all expert weights are thresholded to 0, a weight of 1 will be assigned to the expert with the highest weights prior to thresholding. Thus soft_threshold = 1 leads to the 'follow the leader' strategy if method is set to "ewa". |
hard_threshold |
If specified, the following hard thresholding will be applied to the weights: w = w*(abs(w)>t) where t is the threshold_hard parameter. Defaults to -inf, which means that no threshold will be applied. If all expert weights are thresholded to 0, a weight of 1 will be assigned to the expert with the highest weight prior to thresholding. Thus hard_threshold = 1 leads to the 'follow the leader' strategy if method is set to "ewa". |
fixed_share |
Amount of fixed share to be added to the weights. Defaults to 0. 1 leads to uniform weights. |
parametergrid_max_combinations |
Integer specifying the maximum number of parameter combinations that should be considered. If the number of possible combinations exceeds this threshold, the maximum allowed number is randomly sampled. Defaults to 100. |
parametergrid |
User supplied grid of parameters. Can be used if not all combinations of the input vectors should be considered. Must be a matrix with 13 columns (online) or 12 columns batch with the following order: basis_knot_distance, basis_knot_distance_power, basis_deg, forget_regret, soft_threshold, hard_threshold, fixed_share, p_smooth_lambda, p_smooth_knot_distance, p_smooth_knot_distance_power, p_smooth_deg, p_smooth_ndiff, gamma. |
forget_past_performance |
Share of past performance not to be considered, resp. to be forgotten in every iteration of the algorithm when selecting the best parameter combination. Defaults to 0. |
allow_quantile_crossing |
Shall quantile crossing be allowed? Defaults to false, which means that predictions are sorted in ascending order. |
trace |
Print a progress bar to the console? Defaults to TRUE. |
batch selects various parameters automatically based on the past loss. For this, the parameters smoothing parameters (see below) can be specified as numeric vectors containing values to consider.
This package offers two options for smoothing (Basis Smoothing
and P-Splines).
Parameters b_smooth and p_smooth take named lists to
create the corresponding basis and hat matrices. The arguments are:
knots which determines the number of knots to be created, mu,
sigma, sigma, nonc, tailweight correspond to
to parameters of the beta distribution, which defines how the knots are
#distributed (see ?make_knots for details) the defaults will create
an equidistant knot sequence, deg sets the degree of the spline
function and also influences how many outer knots will be used and
periodic which determines whether the spline basis will be periodic.
It's possible to provide vectors of values for each of these parameters.
In that case, all parameter combinations will be used to create the
respective matrices and all candidates will be considered during
online-learning. In addition to the inputs mentioned
before p_smooth requires ndiff which determines the degree
of differentiation applied to the basis-matrix (can take any value
between and including 1 and 2), lambda which determines the degree
of penalization applied to the smoothing, higher values will give
smoother weight functions. As for the other parameters, it is possible
to provide multiple values.
Returns weights and corresponding predictions. It is possible to impose a convexity constraint to the weights by setting affine and positive to TRUE.
## Not run: T <- 50 # Observations N <- 2 # Experts P <- 9 # Quantiles prob_grid <- 1:P / (P + 1) y <- rnorm(n = T) # Realized experts <- array(dim = c(T, P, N)) # Predictions for (t in 1:T) { experts[t, , 1] <- qnorm(prob_grid, mean = -1, sd = 1) experts[t, , 2] <- qnorm(prob_grid, mean = 3, sd = sqrt(4)) } model <- batch( y = matrix(y), experts = experts, p_smooth = list(lambda = 10) ) print(model) plot(model) autoplot(model) ## End(Not run)## Not run: T <- 50 # Observations N <- 2 # Experts P <- 9 # Quantiles prob_grid <- 1:P / (P + 1) y <- rnorm(n = T) # Realized experts <- array(dim = c(T, P, N)) # Predictions for (t in 1:T) { experts[t, , 1] <- qnorm(prob_grid, mean = -1, sd = 1) experts[t, , 2] <- qnorm(prob_grid, mean = 3, sd = sqrt(4)) } model <- batch( y = matrix(y), experts = experts, p_smooth = list(lambda = 10) ) print(model) plot(model) autoplot(model) ## End(Not run)
Allows for the creation of a Online Object in C++ from R using the C++ conline class.
A conline object from the C++ conline Class.
conline_obj <- new(conline)conline_obj <- new(conline)
This function works in conjunction with the conline class. It takes a matrix of experts and a matrix of outcomes and returns a list of experts which fulfills all properties that are needed for passing it to the an instance of conline.
init_experts_list(experts, y, output_with_names = FALSE)init_experts_list(experts, y, output_with_names = FALSE)
experts |
array of predictions with dimension T x D x P x K (Observations x Variables x Quantiles x Experts) or T x D x K or T x P x K. |
y |
A matrix of outcomes with dimension T x D. |
output_with_names |
Defaults to FALSE. If TRUE, the function returns a list with the experts list, the names of the variables (dnames) and the names of the experts (enames). |
This function creates a list of basis matrices and the corresponding
parameters. It is used in online() to create the basis matrices
for basis smoothing.
make_basis_mats( x, n = length(x), mu = 0.5, sigma = 1, nonc = 0, tailw = 1, deg = 1, periodic = FALSE, idx = NULL, params = NULL )make_basis_mats( x, n = length(x), mu = 0.5, sigma = 1, nonc = 0, tailw = 1, deg = 1, periodic = FALSE, idx = NULL, params = NULL )
x |
The predictor variable |
n |
Number of knots |
mu |
Beta distribution location parameter |
sigma |
Beta distribution scale parameter |
nonc |
Beta distribution noncentrality parameter |
tailw |
Tailweight |
deg |
Degree of splines |
periodic |
Create periodic basis |
idx |
|
params |
Instead of the arguments above, a grid (data.frame or named matrix) of parameters can be passed directly. |
This function creates a list of hat matrices and the corresponding
parameters. It is used in online() to create the hat matrices
for penalized smoothing.
make_hat_mats( x, n = length(x), mu = 0.5, sigma = 1, nonc = 0, tailw = 1, deg = 1, ndiff = 1.5, lambda = -Inf, periodic = FALSE, idx = NULL, params = NULL )make_hat_mats( x, n = length(x), mu = 0.5, sigma = 1, nonc = 0, tailw = 1, deg = 1, ndiff = 1.5, lambda = -Inf, periodic = FALSE, idx = NULL, params = NULL )
x |
The predictor variable |
n |
Number of knots |
mu |
Beta distribution location parameter |
sigma |
Beta distribution scale parameter |
nonc |
Beta distribution noncentrality parameter |
tailw |
Tailweight |
deg |
Degree of splines |
ndiff |
Sets the degree of the differencing matrix for creating the penalty |
lambda |
Penalty parameter (higher values lead to higher penalty) |
periodic |
Create periodic penalty |
idx |
|
params |
Instead of the arguments above, a grid (data.frame or named matrix) of parameters can be passed directly. |
This function creates a knot vector for splines. The knots are distributed
according to a beta distribution. The first input defines the number of inner
knots. The total number of knots is n + 2 * order.
make_knots(n, mu = 0.5, sig = 1, nonc = 0, tailw = 1, deg = 1)make_knots(n, mu = 0.5, sig = 1, nonc = 0, tailw = 1, deg = 1)
n |
Number of knots |
mu |
Beta distribution location parameter |
sig |
Beta distribution scale parameter |
nonc |
Beta distribution noncentrality parameter |
tailw |
Tailweight |
deg |
Degree of splines |
Returns predictions and weights calculated by online-learning algorithms using CRPS Learning.
![[Stable]](data:image/svg+xml;base64,PHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIHhtbG5zOnhsaW5rPSJodHRwOi8vd3d3LnczLm9yZy8xOTk5L3hsaW5rIiB3aWR0aD0iOTYiIGhlaWdodD0iMjAiPjxsaW5lYXJHcmFkaWVudCBpZD0iYiIgeDI9IjAiIHkyPSIxMDAlIj48c3RvcCBvZmZzZXQ9IjAiIHN0b3AtY29sb3I9IiNiYmIiIHN0b3Atb3BhY2l0eT0iLjEiLz48c3RvcCBvZmZzZXQ9IjEiIHN0b3Atb3BhY2l0eT0iLjEiLz48L2xpbmVhckdyYWRpZW50PjxjbGlwUGF0aCBpZD0iYSI+PHJlY3Qgd2lkdGg9Ijk2IiBoZWlnaHQ9IjIwIiByeD0iMyIgZmlsbD0iI2ZmZiIvPjwvY2xpcFBhdGg+PGcgY2xpcC1wYXRoPSJ1cmwoI2EpIj48cGF0aCBmaWxsPSIjNTU1IiBkPSJNMCAwaDUzdjIwSDB6Ii8+PHBhdGggZmlsbD0iIzRjMSIgZD0iTTUzIDBoNDN2MjBINTN6Ii8+PHBhdGggZmlsbD0idXJsKCNiKSIgZD0iTTAgMGg5NnYyMEgweiIvPjwvZz48ZyBmaWxsPSIjZmZmIiB0ZXh0LWFuY2hvcj0ibWlkZGxlIiBmb250LWZhbWlseT0iRGVqYVZ1IFNhbnMsVmVyZGFuYSxHZW5ldmEsc2Fucy1zZXJpZiIgZm9udC1zaXplPSIxMTAiPjx0ZXh0IHg9IjI3NSIgeT0iMTUwIiBmaWxsPSIjMDEwMTAxIiBmaWxsLW9wYWNpdHk9Ii4zIiB0cmFuc2Zvcm09InNjYWxlKC4xKSIgdGV4dExlbmd0aD0iNDMwIj5saWZlY3ljbGU8L3RleHQ+PHRleHQgeD0iMjc1IiB5PSIxNDAiIHRyYW5zZm9ybT0ic2NhbGUoLjEpIiB0ZXh0TGVuZ3RoPSI0MzAiPmxpZmVjeWNsZTwvdGV4dD48dGV4dCB4PSI3MzUiIHk9IjE1MCIgZmlsbD0iIzAxMDEwMSIgZmlsbC1vcGFjaXR5PSIuMyIgdHJhbnNmb3JtPSJzY2FsZSguMSkiIHRleHRMZW5ndGg9IjMzMCI+c3RhYmxlPC90ZXh0Pjx0ZXh0IHg9IjczNSIgeT0iMTQwIiB0cmFuc2Zvcm09InNjYWxlKC4xKSIgdGV4dExlbmd0aD0iMzMwIj5zdGFibGU8L3RleHQ+PC9nPiA8L3N2Zz4=)
online( y, experts, tau, lead_time = 0, loss_function = "quantile", loss_parameter = 1, loss_gradient = TRUE, method = "bewa", b_smooth_pr = list(knots = P, mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, periodic = FALSE), p_smooth_pr = list(knots = P, mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, ndiff = 1.5, lambda = -Inf, periodic = FALSE), b_smooth_mv = list(knots = D, mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, periodic = FALSE), p_smooth_mv = list(knots = D, mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, ndiff = 1.5, lambda = -Inf, periodic = FALSE), forget_regret = 0, soft_threshold = -Inf, hard_threshold = -Inf, fixed_share = 0, gamma = 1, parametergrid_max_combinations = 100, parametergrids = list(general = NULL, b_smooth_pr = NULL, p_smooth_pr = NULL, b_smooth_mv = NULL, p_smooth_mv = NULL), forget_past_performance = 0, save_past_performance = FALSE, save_predictions_grid = FALSE, allow_quantile_crossing = FALSE, init = NULL, loss = NULL, regret = NULL, trace = TRUE, get_timings = FALSE )online( y, experts, tau, lead_time = 0, loss_function = "quantile", loss_parameter = 1, loss_gradient = TRUE, method = "bewa", b_smooth_pr = list(knots = P, mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, periodic = FALSE), p_smooth_pr = list(knots = P, mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, ndiff = 1.5, lambda = -Inf, periodic = FALSE), b_smooth_mv = list(knots = D, mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, periodic = FALSE), p_smooth_mv = list(knots = D, mu = 0.5, sigma = 1, nonc = 0, tailweight = 1, deg = 1, ndiff = 1.5, lambda = -Inf, periodic = FALSE), forget_regret = 0, soft_threshold = -Inf, hard_threshold = -Inf, fixed_share = 0, gamma = 1, parametergrid_max_combinations = 100, parametergrids = list(general = NULL, b_smooth_pr = NULL, p_smooth_pr = NULL, b_smooth_mv = NULL, p_smooth_mv = NULL), forget_past_performance = 0, save_past_performance = FALSE, save_predictions_grid = FALSE, allow_quantile_crossing = FALSE, init = NULL, loss = NULL, regret = NULL, trace = TRUE, get_timings = FALSE )
y |
A numeric matrix of realizations. In probabilistic settings a matrix of dimension Tx1, in multivariate settings a TxD matrix. In the latter case, each slice of the expert's array gets evaluated using the corresponding column of the y matrix. |
experts |
An array of predictions with dimension T x D x P x K (Observations x Variables x Quantiles x Experts) or T x D x K or T x P x K. |
tau |
A numeric vector of probabilities. |
lead_time |
offset for expert forecasts. Defaults to 0, which means that experts forecast t+1 at t. Setting this to h means experts predictions refer to t+1+h at time t. The weight updates delay accordingly. |
loss_function |
Either "quantile", "expectile" or "percentage". |
loss_parameter |
Optional parameter scaling the power of the loss function. |
loss_gradient |
Determines if a linearized version of the loss is used. |
method |
One of "boa", "bewa", "ml_poly" or "ewa". Where "bewa" refers to a mixture of boa and ewa, including the second order refinement of boa, but updating weights with the simple exponential weighting. |
b_smooth_pr |
A named list determining how the B-Spline matrices for probabilistic smoothing are created. Default corresponds to no probabilistic smoothing. See details. |
p_smooth_pr |
A named list determining how the hat matrices for probabilistic P-Spline smoothing are created. Default corresponds to no smoothing. See details. |
b_smooth_mv |
A named list determining how the B-Spline matrices for multivariate smoothing are created. Default corresponds to no probabilistic smoothing. See details. |
p_smooth_mv |
A named list determining how the hat matrices for probabilistic P-Spline smoothing are created. Default corresponds to no smoothing. See details. |
forget_regret |
Share of past regret not to be considered, resp. to be forgotten in every iteration of the algorithm. Defaults to 0. |
soft_threshold |
If specified, the following soft threshold will be applied to the weights: w = sgn(w)*max(abs(w)-t,0) where t is the soft_threshold parameter. Defaults to -inf, which means that no threshold will be applied. If all expert weights are thresholded to 0, a weight of 1 will be assigned to the expert with the highest weights prior to thresholding. Thus soft_threshold = 1 leads to the 'follow the leader' strategy if method is set to "ewa". |
hard_threshold |
If specified, the following hard thresholding will be applied to the weights: w = w*(abs(w)>t) where t is the threshold_hard parameter. Defaults to -inf, which means that no threshold will be applied. If all expert weights are thresholded to 0, a weight of 1 will be assigned to the expert with the highest weight prior to thresholding. Thus hard_threshold = 1 leads to the 'follow the leader' strategy if method is set to "ewa". |
fixed_share |
Amount of fixed share to be added to the weights. Defaults to 0. 1 leads to uniform weights. |
gamma |
Scaling parameter for the learning rate. |
parametergrid_max_combinations |
Integer specifying the maximum number of parameter combinations that should be considered. If the number of possible combinations exceeds this threshold, the maximum allowed number is randomly sampled. Defaults to 100. |
parametergrids |
User supplied grids of parameters. Can be used if not all combinations of the input vectors should be considered. Must be a named list of five matrices. The matrices in list must be named as: "general", "b_smooth_pr", "b_smooth_mv", "p_smooth_pr", "p_smooth_mv". The "general" matrix must contain 11 named columns: "forget_regret", "soft_threshold", "hard_threshold", "fixed_share", "basis_pr_idx", "basis_mv_idx", "hat_pr_idx", "hat_mv_idx", "gamma", "loss_share", "regret_share". The matrices determining the basis smoothing (b_smooth_pr, b_smooth_mv) must contain the following named columns: n, mu, sigma, nonc, tailw, deg, periodic. In addition to the columns of the basis smoothing matrices, the matrices determining the penalized smoothing (p_smooth_pr, p_smooth_mv) must contain the following columns: diff, lambda. The *_idx columns in the general matrix determine which row of the corresponding smoothing matrix is used. |
forget_past_performance |
Share of past performance not to be considered, resp. to be forgotten in every iteration of the algorithm when selecting the best parameter combination. Defaults to 0. |
save_past_performance |
Whether or not the past performance w.r.t to the
considered parameter grid should be reported or not. Defaults to |
save_predictions_grid |
Whether or not all predictions w.r.t to the
considered parameter grid should be reported or not. Defaults to |
allow_quantile_crossing |
Shall quantile crossing be allowed? Defaults to false, which means that predictions are sorted in ascending order. |
init |
A named list containing "init_weights": Array of dimension DxPxK used as starting weights. "R0" a matrix of dimension PxK or 1xK used as starting regret. |
loss |
User specified loss array. Can also be a list with elements "loss_array" and "share", share mixes the provided loss with the loss calculated by profoc. 1 means, only the provided loss will be used. share can also be vector of shares to consider. |
regret |
User specified regret array. If specific, the regret will not be calculated by profoc. Can also be a list with elements "regret_array" and "share", share mixes the provided regret with the regret calculated by profoc. 1 means, only the provided regret will be used. share can also be vector of shares to consider. |
trace |
Print a progress bar to the console? Defaults to TRUE. |
get_timings |
Whether or not to return timings. Defaults to |
online selects various parameters automatically based on the past loss. For this, lambda, forget, fixed_share, gamma, and the smoothing parameters (see below) can be specified as numeric vectors containing values to consider.
This package offers two options for smoothing (Basis Smoothing
and P-Splines). Both options can be used to smooth the weights
over dimension D (covariates) or P (quantiles) or both.
Parameters b_smooth_pr and b_smooth_mv take named lists to
create the corresponding basis matrices. The arguments are:
knots which determines the number of knots to be created, mu,
sigma, sigma, nonc, tailweight correspond to
to parameters of the beta distribution, which defines how the knots are
#distributed (see ?make_knots for details) the defaults will create
an equidistant knot sequence, deg sets the degree of the spline
function and also influences how many outer knots will be used and
periodic which determines whether the spline basis will be periodic.
It's possible to provide vectors of values for each of these parameters.
In that case, all parameter combinations will be used to create the
respective matrices and all candidates will be considered during
online-learning.
Parameters p_smooth_pr and p_smooth_mv determine the hat-matrix
creation for P-Spline smoothing. In addition to the inputs mentioned
before, they require to provide ndiff which determines the degree
of differentiation applied to the basis-matrix (can take any value
between and including 1 and 2), lambda which determines the degree
of penalization applied to the smoothing, higher values will give
smoother weight functions. As for the other parameters, it is possible
to provide multiple values.
Returns weights and corresponding predictions.
## Not run: T <- 50 # Observations N <- 2 # Experts P <- 9 # Quantiles prob_grid <- 1:P / (P + 1) y <- rnorm(n = T) # Realized experts <- array(dim = c(T, P, N)) # Predictions for (t in 1:T) { experts[t, , 1] <- qnorm(prob_grid, mean = -1, sd = 1) experts[t, , 2] <- qnorm(prob_grid, mean = 3, sd = sqrt(4)) } model <- online( y = matrix(y), experts = experts, tau = prob_grid, p_smooth_pr = list(lambda = 10) ) print(model) plot(model) new_y <- matrix(rnorm(1)) # Realized new_experts <- experts[T, , , drop = FALSE] # Update will update the models weights etc if you provide new realizations model <- update(model, new_y = new_y, new_experts = new_experts) # Predict will expand `model$predictions` by default model <- predict(model, new_experts = new_experts, update_model = TRUE) ## End(Not run)## Not run: T <- 50 # Observations N <- 2 # Experts P <- 9 # Quantiles prob_grid <- 1:P / (P + 1) y <- rnorm(n = T) # Realized experts <- array(dim = c(T, P, N)) # Predictions for (t in 1:T) { experts[t, , 1] <- qnorm(prob_grid, mean = -1, sd = 1) experts[t, , 2] <- qnorm(prob_grid, mean = 3, sd = sqrt(4)) } model <- online( y = matrix(y), experts = experts, tau = prob_grid, p_smooth_pr = list(lambda = 10) ) print(model) plot(model) new_y <- matrix(rnorm(1)) # Realized new_experts <- experts[T, , , drop = FALSE] # Update will update the models weights etc if you provide new realizations model <- update(model, new_y = new_y, new_experts = new_experts) # Predict will expand `model$predictions` by default model <- predict(model, new_experts = new_experts, update_model = TRUE) ## End(Not run)
Returns predictions and weights calculated by numeric optimization. The optimization is done in hindsight. This means all observations are used.
oracle(y, experts, tau, affine = FALSE, positive = FALSE, intercept = FALSE, debias = TRUE, loss_function = "quantile", loss_parameter = 1, forget = 0)oracle(y, experts, tau, affine = FALSE, positive = FALSE, intercept = FALSE, debias = TRUE, loss_function = "quantile", loss_parameter = 1, forget = 0)
y |
A numeric matrix of realizations. In probabilistic settings a matrix of dimension Tx1, in multivariate settings a TxD matrix. In the latter case, each slice of the expert's array gets evaluated using the corresponding column of the y matrix. |
experts |
An array of predictions with dimension (Observations, Quantiles, Experts). |
tau |
A numeric vector of probabilities. |
affine |
Defines whether weights are summing to 1 or not. Defaults to FALSE. |
positive |
Defines if a positivity constraint is applied to the weights. Defaults to FALSE. |
intercept |
Determines if an intercept is added, defaults to FALSE. If true, a new first expert is added, always predicting 1. |
debias |
Defines whether the intercepts weight is constrained or not. If TRUE (the default), the intercept weight is unconstrained. Only affects the results if affine and or positive is set to TRUE. If FALSE, the intercept is treated as an expert. |
loss_function |
Either "quantile", "expectile" or "percentage". |
loss_parameter |
Optional parameter scaling the power of the loss function. |
forget |
Adds an exponential forgetting to the optimization. Past observations will get less influence on the optimization. Defaults to 0, which corresponds to no forgetting. |
Returns weights and corresponding predictions. It is possible to calculate the best convex combination of weights by setting affine and positive to TRUE.
## Not run: T <- 50 # Observations N <- 2 # Experts P <- 9 # Quantiles prob_grid <- 1:P / (P + 1) y <- rnorm(n = T) # Realized experts <- array(dim = c(T, P, N)) # Predictions for (t in 1:T) { experts[t, , 1] <- qnorm(prob_grid, mean = -1, sd = 1) experts[t, , 2] <- qnorm(prob_grid, mean = 3, sd = sqrt(4)) } model <- oracle( y = matrix(y), experts = experts ) ## End(Not run)## Not run: T <- 50 # Observations N <- 2 # Experts P <- 9 # Quantiles prob_grid <- 1:P / (P + 1) y <- rnorm(n = T) # Realized experts <- array(dim = c(T, P, N)) # Predictions for (t in 1:T) { experts[t, , 1] <- qnorm(prob_grid, mean = -1, sd = 1) experts[t, , 2] <- qnorm(prob_grid, mean = 3, sd = sqrt(4)) } model <- oracle( y = matrix(y), experts = experts ) ## End(Not run)
This function calculates the B-Spline basis penalty. It follows the procedure outlined in the paper by Zheyuan Li, Jiguo Cao, 2022 "General P-Splines for Non-Uniform B-Splines" doi:10.48550/arXiv.2201.06808. For equidistant knots it coincides with the usual penalty based on the identitiy. For non-equidistant knots it is a weighted penalty with respect to the knot distances. In addition to the above, we added the possibility to calculate periodic penalties which are based on the periodic differencing matrices.
penalty(knots, order, periodic = FALSE, max_diff = 999L)penalty(knots, order, periodic = FALSE, max_diff = 999L)
knots |
Vector of knots. |
order |
Order of the Basis (degree + 1). |
periodic |
Whether the penalties should be periodic or not. |
max_diff |
Maximum difference order to calculate. |
Returns a list of (order - 1) penalty matrices.
## Not run: # Equidistant knots with order 2 knots <- 1:10 P <- penalty(knots, order = 2) print(P[[1]]) # First differences # Non equidistant knots knots <- c(0, 0, 0, 0, 1, 3, 4, 4, 4, 4) P <- penalty(knots, order = 4) print(P[[1]]) # First differences print(P[[2]]) # Second differences print(P[[3]]) # Third differences # Periodic penalty for equidistant knots oder <- 4 deg <- order - 1 knots <- 1:15 penalty(knots, order = order, periodic = TRUE)[[1]] penalty(knots, order = order, periodic = TRUE)[[2]] penalty(knots, order = order, periodic = TRUE)[[3]] ## End(Not run)## Not run: # Equidistant knots with order 2 knots <- 1:10 P <- penalty(knots, order = 2) print(P[[1]]) # First differences # Non equidistant knots knots <- c(0, 0, 0, 0, 1, 3, 4, 4, 4, 4) P <- penalty(knots, order = 4) print(P[[1]]) # First differences print(P[[2]]) # Second differences print(P[[3]]) # Third differences # Periodic penalty for equidistant knots oder <- 4 deg <- order - 1 knots <- 1:15 penalty(knots, order = order, periodic = TRUE)[[1]] penalty(knots, order = order, periodic = TRUE)[[2]] penalty(knots, order = order, periodic = TRUE)[[3]] ## End(Not run)
Plots the most recent weights in each quantile.
## S3 method for class 'batch' plot(x, ...)## S3 method for class 'batch' plot(x, ...)
x |
Object of class inheriting from 'batch' |
... |
further arguments are ignored |
Plots the most recent weights in each quantile.
## S3 method for class 'online' plot(x, ...)## S3 method for class 'online' plot(x, ...)
x |
Object of class inheriting from 'online' |
... |
further arguments are ignored |
This function works in conjunction with the conline class. After the main learning task, it takes the output of the conline class and returns an object suitable for, visualization, further, and deployment. analysis.
post_process_model(model_instance, names)post_process_model(model_instance, names)
model_instance |
An instance of conline. |
names |
A named list with dimnames of |
Calculates predictions based on new expert advice. This does not update weights. If new observations are available use update instead. The latter updates and weights and computes predictions.
## S3 method for class 'online' predict(object, new_experts, update_model = TRUE, ...)## S3 method for class 'online' predict(object, new_experts, update_model = TRUE, ...)
object |
Object of class inheriting from 'online' |
new_experts |
new expert predictions |
update_model |
Defines whether the model object should be updated or not. If TRUE, new forecaster and expert predictions are appended onto the respective object items. Defaults to TRUE. |
... |
further arguments are ignored |
predict.online produces an updated model object.
Prints the average loss of all and the forecast combination.
## S3 method for class 'batch' print(x, ...)## S3 method for class 'batch' print(x, ...)
x |
Object of class inheriting from 'batch' |
... |
further arguments are ignored |
Prints the average loss of all experts and the forecast combination.
## S3 method for class 'online' print(x, ...)## S3 method for class 'online' print(x, ...)
x |
Object of class inheriting from 'online' |
... |
further arguments are ignored |
This function creates a B-Spline matrix.
splines2_basis(x, knots, deg, periodic = FALSE, intercept = TRUE)splines2_basis(x, knots, deg, periodic = FALSE, intercept = TRUE)
x |
Vector of values. |
knots |
Vector of knots. |
deg |
Degree of the Spline functions. |
periodic |
Whether the basis should be periodic or not. |
intercept |
Whether the firs column should be kept. |
Returns a matrix of B-Spline basis functions.
n <- 9 deg <- 3 mu <- 0.35 x <- 0:1000 / 1000 knots <- make_knots(n, mu = mu, deg = deg) B <- splines2_basis(x, knots, deg) ts.plot(B, col = 1:dim(B)[2]) # Periodic Case B <- splines2_basis(x, knots, deg, periodic = TRUE) ts.plot(B, col = 1:dim(B)[2])n <- 9 deg <- 3 mu <- 0.35 x <- 0:1000 / 1000 knots <- make_knots(n, mu = mu, deg = deg) B <- splines2_basis(x, knots, deg) ts.plot(B, col = 1:dim(B)[2]) # Periodic Case B <- splines2_basis(x, knots, deg, periodic = TRUE) ts.plot(B, col = 1:dim(B)[2])
Calculates parameters chosen during optimization and aggregates losses.
## S3 method for class 'online' summary(object, ...)## S3 method for class 'online' summary(object, ...)
object |
Object of class inheriting from 'online' |
... |
further arguments are ignored |
tidy will transform the experts_loss array of an online object
into a tibble that is better suited for plotting and analysis.
## S3 method for class 'online.experts_loss' tidy(x, ...)## S3 method for class 'online.experts_loss' tidy(x, ...)
x |
The experts_loss of an |
... |
Not currently used. |
A tibble with columns t d p k and w corresponding
to the time, marginals, probabilities, and experts_loss
of the online-learning computation.
tidy will transform the 'forecaster_loss“ array of an online object
into a tibble that is better suited for plotting and analysis.
## S3 method for class 'online.forecaster_loss' tidy(x, ...)## S3 method for class 'online.forecaster_loss' tidy(x, ...)
x |
The forecaster_loss of an |
... |
Not currently used. |
A tibble with columns t d p k and w corresponding
to the time, marginals, probabilities, and forecaster_loss
of the online-learning computation.
tidy will transform the predictions array of an online object
into a tibble that is better suited for plotting and analysis.
## S3 method for class 'online.predictions' tidy(x, ...)## S3 method for class 'online.predictions' tidy(x, ...)
x |
The predictions of an |
... |
Not currently used. |
A tibble with columns t d p k and w corresponding
to the time, marginals, probabilities, and predictions
of the online-learning computation.
tidy will transform the weights array of an online object
into a tibble that is better suited for plotting and analysis.
## S3 method for class 'online.weights' tidy(x, ...)## S3 method for class 'online.weights' tidy(x, ...)
x |
The weights of an |
... |
Not currently used. |
A tibble with columns t d p k and w corresponding
to the time, marginals, probabilities, experts, and weights
of the online-learning computation.
Continues learning using new observations and new expert advice.
## S3 method for class 'online' update(object, new_y, new_experts = NULL, trace = FALSE, ...)## S3 method for class 'online' update(object, new_y, new_experts = NULL, trace = FALSE, ...)
object |
Object of class inheriting from 'online' |
new_y |
new observations |
new_experts |
new expert predictions. This must be left unspecified |
trace |
If a progress bar shall be shown. Defaults to FALSE if the model already contains the expert predictions corresponding to new_y. |
... |
further arguments are ignored |
update.online produces an updated model object.