Skip to contents

Emilio González Coya

Getting started with fgls

Source: vignettes/fgls.Rmd
fgls.Rmd

fgls estimates linear regressions by feasible generalized least squares (FGLS). It implements two recent methods:

  • Serial correlation (Perron and González-Coya, 2026): the errors are approximated by an autoregression whose order is chosen by BIC and whose coefficients come from a Durbin regression. The resulting estimator is consistent even when the regressors are only pre-determined, a situation where OLS and Cochrane-Orcutt are not. A variance estimator accounts for the estimated filter.
  • Heteroskedasticity of unknown form (González-Coya and Perron, 2024): the log of the squared residuals is modeled by an adaptive Lasso over a dictionary of candidate covariates, and the weighted regression is paired with a bias-corrected sandwich variance that stays valid when that model is wrong.

The two corrections can be combined (Supplement S-4 of the 2026 paper). The R package is a port of the Python package pyfgls, and the two agree to machine precision on deterministic quantities (see vignette("validation")).

Three examples

Serial correlation

simulate_serial() generates the design of Table 1 of the 2026 paper: AR(1) errors with coefficient 0.8, and a regressor that responds to past innovations, \(x_t = v_t + e_{t-1}\), so it is pre-determined but not exogenous.

sim <- simulate_serial(500, seed = 1)
fit <- serial_fgls(y ~ x, data = sim)
summary(fit)
#> 
#> Call:
#> serial_fgls(formula = y ~ x, data = sim)
#> 
#> Method:               FGLS (serial)
#> Covariance:           robust
#> Observations:         500 (effective 499)
#> Inference:            normal
#> AR order k*:          1 [BIC, kmax = 15]
#> AR coefficients:      0.7800
#> Non-exogeneity share: 0.011
#> 
#> Coefficients:
#>             Estimate Std. Error z value Pr(>|z|)    
#> (Intercept)  1.11877    0.20552   5.444 5.22e-08 ***
#> x            0.98062    0.02517  38.967  < 2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> Notes:
#> [1] Covariance: PG26 eq. (29), robust to non-exogenous (pre-determined)
#>     regressors.
#> [2] Collinear lagged columns dropped from the Durbin regression:
#>     L1.(Intercept).

The true slope is 1. OLS is badly biased in this design; FGLS is not:

compare_fgls(OLS = lm(y ~ x, data = sim), FGLS = fit)
#>          term statistic        OLS       FGLS
#> 1 (Intercept)  estimate 1.12860678 1.11876951
#> 2 (Intercept) std.error 0.09034505 0.20551635
#> 3 (Intercept)  ci.width 0.35414609 0.80560929
#> 4           x  estimate 1.32962748 0.98062019
#> 5           x std.error 0.05108277 0.02516513
#> 6           x  ci.width 0.20024076 0.09864549

fit$ar_order is the order chosen by BIC, fit$rho the AR coefficients and fit$nonexogeneity_share the relative size of the correction that makes the variance robust to non-exogenous regressors.

Heteroskedasticity of unknown form 

het <- simulate_hetero(200, skedastic = "log2", seed = 2)
fit_h <- lasso_fgls(y ~ x, data = het, seed = 1)
fit_h$skedastic$selected
#> [1] "log(x)" "cos(x)"
compare_fgls(OLS = lm(y ~ x, data = het), FGLS = fit_h)
#>          term statistic        OLS       FGLS
#> 1 (Intercept)  estimate 0.98861426 0.98253777
#> 2 (Intercept) std.error 0.18311445 0.08973843
#> 3 (Intercept)  ci.width 0.71779546 0.35176818
#> 4           x  estimate 1.02852565 1.03190100
#> 5           x std.error 0.09523622 0.06032162
#> 6           x  ci.width 0.37331913 0.23645640

The candidate covariates default to \(x, x^2, \log x, (\log x)^2, \cos x, \cos 2x\) for every non-constant regressor (make_dictionary()); supply your own with z. The penalty and the exponent of the adaptive weights are chosen by 10-fold cross-validation, so set seed for reproducible results.

Both 

both <- simulate_combined(200, ar = 0.5, skedastic = "x2", seed = 3)
fit_b <- fgls(y ~ x, data = both, seed = 1)
fit_b
#> 
#> Call:
#> fgls(formula = y ~ x, data = both, seed = 1)
#> 
#> F-CH (serial+hetero); AR order k* = 1 (BIC), rho = 0.4976; skedastic fit: lasso (selected: x, log(x)); covariance: robust
#> 
#> Coefficients:
#> (Intercept)            x  
#>      1.5934       0.8101
compare_fgls(`F-C` = fit_b$stages$serial, `F-CH` = fit_b)
#>          term statistic       F-C      F-CH
#> 1 (Intercept)  estimate 1.3723275 1.5934094
#> 2 (Intercept) std.error 0.5413481 0.4430785
#> 3 (Intercept)  ci.width 2.1220455 1.7368357
#> 4           x  estimate 0.9130718 0.8100948
#> 5           x std.error 0.2611294 0.2123379
#> 6           x  ci.width 1.0236086 0.8323494

fgls() is the general function: serial = "none" gives the heteroskedasticity-only estimator and hetero = "none" the serial-only one.

Working with fits

Fits behave like lm objects:

coef(fit)
#> (Intercept)           x 
#>   1.1187695   0.9806202
confint(fit, level = 0.9)
#>                   5 %     95 %
#> (Intercept) 0.7807252 1.456814
#> x           0.9392272 1.022013
predict(fit, newdata = data.frame(x = c(-1, 0, 1)), interval = "confidence")
#>         fit        lwr      upr
#> 1 0.1381493 -0.2675824 0.543881
#> 2 1.1187695  0.7159649 1.521574
#> 3 2.0993897  1.6934951 2.505284
vcov(fit, type = "naive") # PG26 eq. (27), valid only for exogenous regressors
#>              (Intercept)            x
#> (Intercept) 4.223673e-02 6.788463e-06
#> x           6.788463e-06 6.197699e-04

They work with lmtest::coeftest(), lmtest::coefci() and the broom generics tidy(), glance() and augment(). plot(fit) draws diagnostics: the autocorrelation function of the innovations (residuals of the final filtered regression) should look like white noise.

Time series inputs

Rows must be in time order. Besides data frames, fgls() accepts ts, zoo and xts objects; a data frame can carry a time index through time =, which is checked for sorting and gaps:

dates <- seq(as.Date("2000-01-01"), by = "month", length.out = 500)
df <- data.frame(date = dates, sim)
serial_fgls(y ~ x, data = df, time = "date")$ar_order
#> [1] 1

Choosing options

  • With several stochastic regressors, the BIC that penalizes only the AR order can select the largest order; use control = fgls_control(ic_penalty = "params").
  • Declare lagged dependent variables among the regressors in lagged_dep_vars.
  • For errors known to be MA(1), such as multi-step forecast errors, use serial = "ma1".
  • exog = "exogenous" never changes the estimates. In the serial-only estimator it only changes the notes, and the robust variance is valid either way. In the combined estimator it also drops the first-stage correction from the variance by default (first_stage_correction = FALSE); that correction is needed when the regressors are only pre-determined, so keep the default exog = "predetermined" unless they are known to be exogenous.

See vignette("serial-correlation"), vignette("heteroskedasticity") and vignette("combined") for the methods in more detail.

References

Perron, P. and González-Coya, E. (2026). Feasible GLS for time series regression. Journal of Econometrics, 258, 106346. https://doi.org/10.1016/j.jeconom.2026.106346

González-Coya, E. and Perron, P. (2024). Estimation in the presence of heteroskedasticity of unknown form: a Lasso-based approach. Journal of Econometric Methods, 13(1), 29–48. https://doi.org/10.1515/jem-2023-0007