---
title: "Replicating Giacomini & White (2006)"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Replicating Giacomini & White (2006)}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment  = "#>",
  fig.width  = 7,
  fig.height = 4
)
```

This article applies `gw_test()`, the Giacomini and White (2006) test of conditional equal predictive ability (CEPA), to the Survey of Professional Forecasters' mean CPI inflation forecasts. The benchmark is a naive random walk in inflation (next quarter's forecast equals this quarter's realised inflation). The alternative is the SPF mean at horizons $h \in \{0, 1, 2, 3, 4\}$ quarters. The bundled `gw2006` dataset extends GW's quarterly sample to the present using the same Philadelphia Fed SPF source.

```{r setup, message = FALSE}
library(forecastdom)
data(gw2006)
str(gw2006)
```

## CEPA across forecast horizons

```{r table-horizons}
run_h <- function(h) {

  spf_col <- paste0("spf_h", h)
  ok <- complete.cases(gw2006[, c("infl", spf_col, "infl_lag")])
  d  <- gw2006[ok, ]
  e_spf <- d$infl - d[[spf_col]] # SPF errors
  e_rw  <- d$infl - d$infl_lag   # random-walk errors
  r <- gw_test(e_spf, e_rw)      # default: constant + lagged loss diff

  data.frame(h = h, n = nrow(d),
             mse_spf = mean(e_spf^2),
             mse_rw  = mean(e_rw^2),
             wald    = unname(r$statistic),
             pvalue  = unname(r$pvalue),
             reject  = unname(r$pvalue) < 0.05)

}

tab <- do.call(rbind, lapply(0:4, run_h))

knitr::kable(
  tab, digits = 3, row.names = FALSE,
  col.names = c("$h$", "$n$",
                "$MSE_{SPF}$", "$MSE_{RW}$",
                "Wald", "$p$-value", "Reject"))
```

At every horizon SPF has a lower mean squared error than the random-walk benchmark, yet the GW test fails to reject conditional equal predictive ability at the 5% level. The pattern matches Atkeson and Ohanian (2001): once the autocorrelation of the loss differential is taken into account, sophisticated inflation forecasts are hard to tell apart from a "no-change" benchmark in a formal test.

## Choice of instruments

`gw_test()` uses two instruments by default: a constant and the lagged loss differential. Different instruments target different features of the conditioning information set. Below we replace the lag with the lagged inflation level (a regime indicator that captures high- and low-inflation environments) at horizon $h = 1$.

```{r instrument-comparison}
h <- 1L
ok <- complete.cases(gw2006[, c("infl", paste0("spf_h", h), "infl_lag")])
d  <- gw2006[ok, ]
e_spf <- d$infl - d[[paste0("spf_h", h)]]
e_rw  <- d$infl - d$infl_lag

# Default: constant + lagged loss differential
r_default <- gw_test(e_spf, e_rw)

# Custom: constant + lagged inflation level (padded with NA at t = 1
# so that nrow(W) matches length(e1)).
W_infl <- cbind(1, c(NA, head(d$infl_lag, -1)))
keep   <- !is.na(W_infl[, 2])
r_infl <- gw_test(e_spf[keep], e_rw[keep],
                  instruments = W_infl[keep, , drop = FALSE])

# Absolute-error loss instead of squared-error
r_ae <- gw_test(e_spf, e_rw, loss = "AE")

tab2 <- data.frame(
  spec = c("Default (const + lagged Δloss)",
           "Const + lagged inflation",
           "Absolute-error loss"),
  wald   = c(r_default$statistic, r_infl$statistic, r_ae$statistic),
  df     = c(r_default$df,        r_infl$df,        r_ae$df),
  pvalue = c(r_default$pvalue,    r_infl$pvalue,    r_ae$pvalue)
)

knitr::kable(
  tab2, digits = 3, row.names = FALSE,
  col.names = c("Specification", "Wald", "df", "$p$-value"))
```

The decision is robust across all three specifications: SPF and
the random walk cannot be statistically distinguished at the 5%
level regardless of which conditioning instruments or loss function
we choose.

## Note on the original paper

Giacomini and White (2006, Section 4) compare SPF nowcasts to the *Greenbook* (the Federal Reserve staff's internal inflation forecast) using a richer instrument set. They find that the test *does* reject equal conditional predictive ability between the two *sophisticated* forecasts in some conditioning states. Replicating that result needs Greenbook data with its five-year embargo. This article instead demonstrates the test mechanics against the simpler random-walk benchmark, which is freely available.

## References

- Atkeson, A. and Ohanian, L. E. (2001). Are Phillips curves useful
  for forecasting inflation? *Federal Reserve Bank of Minneapolis
  Quarterly Review*, 25(1), 2-11.
- Giacomini, R. and White, H. (2006). Tests of conditional
  predictive ability. *Econometrica*, 74(6), 1545-1578.