Size-Power Tradeoff Visualization for Equal Predictive Ability of Two Forecasts.

ForeComp

This repository contains the ForeComp R package for forecasting comparison and equal predictive accuracy testing. The package implements Diebold-Mariano and related procedures, including fixed-smoothing (fixed-b and fixed-m) variants, along with tools for size-power tradeoff analysis.

Contents

• R/ package source code implementing forecast evaluation functions
• data/ packaged datasets used by the package
• man/ function documentation (.Rd files)
• tests/testthat/ automated tests
• NEWS.md release notes and version history

The package code and datasets support forecasting-method comparison and robustness analysis for equal predictive accuracy testing.

Installation

ForeComp requires R (>= 3.5.0).

Install the stable release from CRAN:

install.packages("ForeComp")

Install the development version from GitHub:

remotes::install_github("mcmcs/ForeComp")

Main functions

Plot_Tradeoff(data, f1, f2, y, ...)

• Computes loss differentials between two forecast series and evaluates how the truncation choice (M) affects size distortion and maximum power loss.
• Fits an ARIMA approximation to the loss differential process and uses simulation (n_sim) to construct size-power tradeoff results across m_set.
• Returns a list where the first element is a ggplot2 tradeoff figure and the second element is the underlying computed table.

For direct equal-predictive-accuracy testing with specific bandwidth rules:

ForeComp supports both fixed-smoothing defaults and baseline alternatives for Bartlett-kernel DM tests:

• dm.test.bt(d, M = NA, Mopt = ...) (normal approximation, default Mopt = 2)
• dm.test.bt.fb(d, M = NA, Mopt = ...) (fixed-b approximation, default Mopt = 1)

For both functions, Mopt has the same meaning:

• Mopt = 1 (LLSW): M = ceiling(1.3 * sqrt(T))
• Mopt = 2 (NW 1994): M = ceiling(4 * (T / 100)^(2/9))
• Mopt = 3 (textbook NW / Andrews): M = ceiling(0.75 * T^(1/3))
• Mopt = 4 (CI baseline): M = floor(sqrt(T))

where T = length(d).

For EWC fixed-smoothing:

• dm.test.ewc.fb(d, B = NA, Bopt = 1) uses B = floor(0.4 * T^(2/3)) with a lower bound of 1.

Introductory Paper

Coming soon!