Conformalized Forecasting using Machine Learning (and statistical) models with ARCH effects

mlarchf(
  y,
  h = 10L,
  mean_model = forecast::auto.arima,
  model_residuals = forecast::meanf,
  fit_func = ahead::ridge,
  predict_func = predict,
  lags_vol = 10L,
  type_pi = c("surrogate", "bootstrap", "kde"),
  type_sim_conformalize = c("surrogate", "block-bootstrap", "bootstrap", "kde",
    "fitdistr"),
  ml_method = NULL,
  level = 95,
  B = 250L,
  ml = TRUE,
  stat_model = NULL,
  n_clusters = 0,
  clustering_dist = "euclidean",
  clustering_method = "kmeans",
  ...
)

Arguments

y: A numeric vector or time series of class ts
h: Forecasting horizon
mean_model: Function to fit the mean model (default: forecast::auto.arima)
model_residuals: Function to model the residuals (default: forecast::thetaf)
fit_func: Fitting function for the variance model (default: ahead::ridge)
predict_func: Prediction function for the variance model (default: predict)
lags_vol: Number of lags for squared residuals regression
type_pi: Type of prediction interval ("kde", "surrogate", or "bootstrap") for volatility modeling
type_sim_conformalize: Type of simulation for conformalization of standardized residuals ("block-bootstrap", "surrogate", "kde", "bootstrap", or "fitdistr")
ml_method: caret package Machine learning method to use, if fit_func and predict_func aren't provided. If NULL, uses fit_func and predict_func. See unique(caret::modelLookup()$model).
level: Confidence level for prediction intervals
B: Number of bootstrap replications or simulations
ml: If TRUE, fit_func and predict_func are used, otherwise a statistical model in stat_model
stat_model: A statistical model, e.g forecast::thetaf or forecast::auto.arima
n_clusters: Number of clusters for residuals (default is 0) for ml == TRUE and ml_method not NULL
clustering_dist: Only if n_clusters > 0. If "euclidean", then mean square error, if "manhattan ", the mean absolute error is used.
clustering_method: Only if n_clusters > 0. If "kmeans", then we have the kmeans clustering method, if "hardcl" we have the On-line Update (Hard Competitive learning) method, and if "neuralgas", we have the Neural Gas (Soft Competitive learning) method. Abbreviations of the method names are accepted.
...: Additional parameters to be passed to stat_model

Value

A forecast object containing predictions and prediction intervals

Examples


y <- fpp2::goog200

# Default model for volatility (Ridge regression for volatility)
(obj_ridge <- ahead::mlarchf(y, h=20L, B=500L))
#>     Point Forecast    Lo 95    Hi 95
#> 201       534.4214 513.0126 549.3933
#> 202       536.2434 516.2732 554.8150
#> 203       536.4303 521.4628 553.5272
#> 204       536.6375 516.4954 559.5297
#> 205       536.9744 513.3218 555.6087
#> 206       536.8669 518.1739 552.4974
#> 207       539.0674 519.1329 558.1132
#> 208       537.4618 501.6436 560.6395
#> 209       540.1076 518.9179 558.1104
#> 210       536.5789 509.1344 555.3092
#> 211       540.7064 517.4373 555.1692
#> 212       543.6837 528.5274 561.0367
#> 213       543.4614 524.3126 564.4024
#> 214       544.3046 526.8852 564.2636
#> 215       544.1797 522.0492 560.9355
#> 216       544.7224 525.9826 560.9314
#> 217       545.7186 524.2705 564.2966
#> 218       546.5558 524.4960 568.5002
#> 219       546.0634 527.0138 562.9881
#> 220       548.5014 525.1081 567.3897
plot(obj_ridge)