Autocorrelated multistep-ahead conformal prediction method

Compute prediction intervals and other information by applying the Autocorrelated Multistep-ahead Conformal Prediction (AcMCP) method. The method can only deal with asymmetric nonconformity scores, i.e., forecast errors.

Usage

acmcp(
  object,
  alpha = 1 - 0.01 * object$level,
  ncal = 10,
  rolling = FALSE,
  integrate = TRUE,
  scorecast = TRUE,
  lr = 0.1,
  Tg = NROW(object$ERROR),
  delta = 0.01,
  Csat = NULL,
  KI = max(abs(object$ERROR), na.rm = TRUE),
  update = FALSE,
  ma_method = c("CSS-ML", "CSS"),
  ...
)

Arguments

object: An object of class "cvforecast". It must have an argument x for the original univariate time series, an argument MEAN for the point forecasts and ERROR for the forecast errors on the validation set. See the results of a call to cvforecast.
alpha: A numeric vector of significance levels to achieve a desired coverage level $1-\alpha$. Defaults to 1 - 0.01 * object$level, the levels used in object.
ncal: Length of the burn-in period for training the scorecaster. If rolling = TRUE, it is also used as the length of the trailing windows for the learning rate calculation and of the windows for the calibration set. If rolling = FALSE, it is used as the initial period of the calibration sets and of the trailing windows for the learning rate calculation. Defaults to 10.
rolling: If TRUE, a rolling window strategy will be adopted to form the trailing window for the learning rate calculation and the calibration set for the scorecaster if applicable. Otherwise, an expanding window strategy will be used. Defaults to FALSE.
integrate: If TRUE, error integration will be included in the update process. Defaults to TRUE.
scorecast: If TRUE, scorecasting will be included in the update process. Defaults to TRUE.
lr: A positive initial learning rate used for quantile tracking. Defaults to 0.1.
Tg: The time that is set to achieve the target absolute coverage guarantee before this. It must be greater than 1 when Csat is not supplied. Defaults to NROW(object$ERROR), the number of cross-validation periods in object.
delta: A number in $(0, 1)$. The target absolute coverage guarantee is set to $1-\alpha-\delta$. Defaults to 0.01.
Csat: A positive constant ensuring that by time Tg, an absolute guarantee is of at least $1-\alpha-\delta$ coverage. Defaults to NULL, in which case it is derived from Tg and delta as 2 / pi * (ceiling(log(Tg) * delta) - 1 / log(Tg)).
KI: A non-negative constant to place the integrator on the same scale as the scores. Defaults to max(abs(object$ERROR), na.rm = TRUE), the largest absolute forecast error in object.
update: If TRUE, object already holds the results of a previous call and only the newly added time steps are computed; the prediction intervals produced earlier are carried over unchanged. Set by update.cpforecast and not normally set by hand. Defaults to FALSE.
ma_method: Estimation method for the MA$(h-1)$ scorecaster. "CSS-ML" uses conditional sum of squares for starting values followed by maximum likelihood. "CSS" uses conditional sum of squares only and may be faster, especially for longer forecast horizons, but can produce different estimates. Defaults to "CSS-ML".
...: Not used.

Value

A list of class c("acmcp", "cpforecast", "cvforecast", "forecast") with the following components:

x: The original time series.
series: The name of the series x.
xreg: Exogenous predictor variables used, if applicable.
method: A character string "acmcp".
cp_times: An integer vector giving the number of conformal predictions performed in cross-validation for each forecast horizon.
scorecast_times: An integer vector giving the number of successful scorecasts for each forecast horizon. Returned when scorecast = TRUE.
MEAN: Point forecasts as a multivariate time series, where the $h$th column holds the point forecasts for forecast horizon $h$. The time index corresponds to the period for which the forecast is produced.
ERROR: Forecast errors given by $e_{t+h|t} = y_{t+h}-\hat{y}_{t+h|t}$.
LOWER: A list containing lower bounds for prediction intervals for each level. Each element within the list will be a multivariate time series with the same dimensional characteristics as MEAN.
UPPER: A list containing upper bounds for prediction intervals for each level. Each element within the list will be a multivariate time series with the same dimensional characteristics as MEAN.
level: The confidence values associated with the prediction intervals.
call: The matched call.
model: A list containing information about the conformal prediction model: the resolved arguments in model$args, the call and the arguments of the underlying cross-validation in model$cvforecast, the learning rates actually used in model$lr_update, and the recursion state needed to extend the results in model$state and model$t_last. The integrator and scorecaster series are returned in model$integrator and model$scorecaster when integrate and scorecast are TRUE; each holds a lower and an upper element.

If mean is included in the object, the components mean, lower, and upper will also be returned, showing the information about the test set forecasts generated using all available observations.

Details

Similar to the PID method, the AcMCP method also integrates three modules (P, I, and D) to form the final iteration. However, instead of performing conformal prediction for each individual forecast horizon h separately, AcMCP employs a combination of an MA$(h-1)$ model and a linear regression model of $e_{t+h|t}$ on $e_{t+h-1|t},\dots,e_{t+1|t}$ as the scorecaster. This allows the AcMCP method to capture the relationship between the $h$-step-ahead forecast error and the past errors.

Scorecasts are constructed recursively, so longer forecast horizons require more history before all scorecaster inputs are available. For horizon $h$, the first scorecast can be computed at cross-validation error index ncal + $h(h-1)/2$. The number of successful scorecasts is reported in scorecast_times.

References

Wang, X., and Hyndman, R. J. (2024). "Online conformal inference for multi-step time series forecasting", arXiv preprint arXiv:2410.13115.

Examples

# Simulate time series from an AR(2) model
library(forecast)
set.seed(1)
series <- arima.sim(n = 200, list(ar = c(0.8, -0.5)), sd = sqrt(1))

# Cross-validation forecasting
far2 <- function(x, h, level) {
  Arima(x, order = c(2, 0, 0)) |>
    forecast(h = h, level)
}
fc <- cvforecast(series, forecastfun = far2, h = 3, level = 95, window = 50)

# AcMCP setup
Tg <- 200; delta <- 0.01
Csat <- 2 / pi * (ceiling(log(Tg) * delta) - 1 / log(Tg))
KI <- 2

# AcMCP with integrator and scorecaster
acmcpfc <- acmcp(fc, ncal = 50, rolling = TRUE, KI = KI, Csat = Csat)
print(acmcpfc)
#> ACMCP 
#> 
#> Call:
#>  acmcp(object = fc, ncal = 50, rolling = TRUE, Csat = Csat, KI = KI) 
#> 
#>  cp_times (the forward step included): 101 (h=1), 100 (h=2), 99 (h=3)
#> 
#> Forecasts of the forward step:
#>     Point Forecast     Lo 95    Hi 95
#> 201      0.6271538 -1.013545 2.673790
#> 202      0.8607034 -2.090441 4.044708
#> 203      0.4935805 -2.055964 3.526965
summary(acmcpfc)
#> ACMCP 
#> 
#> Call:
#>  acmcp(object = fc, ncal = 50, rolling = TRUE, Csat = Csat, KI = KI) 
#> 
#>  cp_times (the forward step included): 101 (h=1), 100 (h=2), 99 (h=3)
#> 
#> Forecasts of the forward step:
#>     Point Forecast     Lo 95    Hi 95
#> 201      0.6271538 -1.013545 2.673790
#> 202      0.8607034 -2.090441 4.044708
#> 203      0.4935805 -2.055964 3.526965
#> 
#> Cross-validation error measures:
#>       ME   MAE   MSE RMSE    MPE    MAPE  MASE RMSSE Winkler_95 MSIS_95
#> CV 0.007 0.946 1.415 1.06 -3.933 269.763 0.992 0.882      6.618   7.132
acmcpfc$scorecast_times
#> [1] 101 100  98