Skip to contents

Wrapper function for optimizing any combination of a given seasonal.factor vector in NNS.ARMA. The objective function obj.fn on a validation window (the last length(variable) - training.set observations) is used to determine the optimum across all NNS.ARMA methods.

Usage

NNS.ARMA.optim(
  variable,
  h = NULL,
  training.set = NULL,
  seasonal.factor,
  lin.only = FALSE,
  negative.values = FALSE,
  obj.fn = expression({
     .denom <- NNS::Co.LPM(1, predicted, actual, target_x =
    mean(predicted), target_y = mean(actual)) + NNS::Co.UPM(1, predicted, actual,
    target_x = mean(predicted), target_y = mean(actual))
     if (.denom == 0) Inf else
mean((predicted - actual)^2)/.denom
 }),
  objective = "min",
  linear.approximation = TRUE,
  ncores = NULL,
  pred.int = 0.95,
  print.trace = TRUE,
  plot = FALSE,
  bias.correction = c("none", "split")
)

Arguments

variable

a numeric vector.

h

integer; NULL (default) Number of periods to forecast out of sample. If NULL, h = length(variable) - training.set.

training.set

integer; NULL (default) Sets the number of variable observations as the training set. See Note below for recommended uses.

seasonal.factor

integers; Multiple frequency integers considered for NNS.ARMA model, i.e. (seasonal.factor = c(12, 24, 36)).

lin.only

logical; FALSE (default) For fast optimization of the linear regression method. More robust than lin.only = TRUE.

negative.values

logical; FALSE (default) If the variable can be negative, set to (negative.values = TRUE). It will automatically select (negative.values = TRUE) if the training data variable[1 : training.set] has a negative value, and for the out-of-sample forecast if variable does.

obj.fn

expression; The default is the mean squared error divided by the degree-1 co-movement of predicted and actual, mean((predicted - actual)^2) / (Co.LPM(1, predicted, actual, mean(predicted), mean(actual)) + Co.UPM(1, predicted, actual, mean(predicted), mean(actual))) (Inf when the co-movement is 0), to be minimized. Any expression(...) using the specific terms predicted and actual can be used.

objective

options: ("min", "max") "min" (default) Select whether to minimize or maximize the objective function obj.fn.

linear.approximation

logical; TRUE (default) Uses the best linear output from NNS.reg to generate a nonlinear and mixture regression for comparison. FALSE is a more exhaustive search over the objective space.

ncores

integer; value specifying the number of cores to be used in the parallelized procedure. If NULL (default), the number of cores to be used is equal to the number of cores of the machine - 1.

pred.int

numeric [0, 1]; 0.95 (default) Returns the associated prediction intervals for the final estimate. Constructed from the returned model's internal validation residuals (actual - forecast), adding their LPM.VaR and UPM.VaR at (1 - pred.int)/2 to the final estimates.

print.trace

logical; TRUE (default) Prints current iteration information. Suggested as backup in case of error, best parameters to that point still known and copyable!

plot

logical; FALSE (default)

bias.correction

"none" (default) returns the selected model's forecasts without an additive shift. "split" estimates a gravity bias from the first half of the validation errors and accepts it only if the objective improves on the later half; at least four validation observations are required. On rolling-origin forecasts of nine series (AirPassengers, UKgas, LakeHuron, a random walk, sunspot.month, nottem, USAccDeaths, co2 and Nile; six expanding origins, h = 12), the uncorrected forecasts had the lower mean MASE on eight of the nine series, both against "split" and against estimating and accepting the shift on the same window, a rule that is no longer offered. Model selection still uses the validation window, so these diagnostics are not independent out-of-sample performance estimates.

Value

Returns a list containing:

  • $period a vector of optimal seasonal periods

  • $weights the optimal weights of each seasonal period between an equal weight or NULL weighting

  • $obj.fn the obj.fn value of the returned model's internal validation forecasts, after $bias.shift.

  • $method the method identifying which NNS.ARMA method was used.

  • $shrink whether to use the shrink parameter in NNS.ARMA.

  • $nns.regress whether to smooth the variable via NNS.reg before forecasting.

  • $bias.shift the accepted additive correction (0 when rejected or disabled). Add it to the forecast from NNS.ARMA with the returned model parameters, then apply the nonnegative support clamp if required.

  • $bias.validation correction policy, estimated/applied shifts, chronological estimate and acceptance rows, and raw/adjusted acceptance scores.

  • $raw.results forecasts before the additive bias correction.

  • $validation.forecast the final internal validation forecasts, including correction and support clamp.

  • $errors a vector of the returned model's internal validation errors (forecast - actual), before $bias.shift.

  • $results a vector of length h.

  • $lower.pred.int a vector of lower prediction intervals per forecast point.

  • $upper.pred.int a vector of upper prediction intervals per forecast point.

Note

By default no bias shift is added. With bias.correction = "split" the shift is estimated on the first half of the validation window and accepted on the second. Base model selection still uses the internal validation window, so these tuning scores are not an independent evaluation of the whole forecasting procedure. Evaluate the complete procedure at fresh chronological origins.

  • Typically, (training.set = 0.8 * length(variable)) is used for optimization. Smaller samples could use (training.set = 0.9 * length(variable)) (or larger) in order to preserve information.

  • The number of combinations will grow prohibitively large, they should be kept as small as possible. seasonal.factor containing an element too large will result in an error. Please reduce the maximum seasonal.factor.

  • Set (ncores = 1) if routine is used within a parallel architecture.

References

Viole, F. and Nawrocki, D. (2013) "Nonlinear Nonparametric Statistics: Using Partial Moments" (ISBN: 1490523995, 2nd edition: https://ovvo-financial.github.io/NNS/book/)

Author

Fred Viole, OVVO Financial Systems

Examples


## Nonlinear NNS.ARMA period optimization using 2 yearly lags on AirPassengers monthly data
if (FALSE) { # \dontrun{
nns.optims <- NNS.ARMA.optim(AirPassengers[1:132], training.set = 120,
seasonal.factor = seq(12, 24, 6))

## To predict out of sample using best parameters:
NNS.ARMA.optim(AirPassengers[1:132], h = 12, seasonal.factor = seq(12, 24, 6))

## Incorporate any objective function from external packages (such as \code{Metrics::mape})
NNS.ARMA.optim(AirPassengers[1:132], h = 12, seasonal.factor = seq(12, 24, 6),
obj.fn = expression(Metrics::mape(actual, predicted)), objective = "min")
} # }