`pi_correction`

Back to L4 | Browse all axes | Browse all options

Axis pi_correction on sub-layer L4_E_predict (layer l4).

Sub-layer

L4_E_predict

Axis metadata

Default: 'none'
Sweepable: True
Status: operational

Operational status summary

Operational: 2 option(s)
Future: 0 option(s)

Options

`none` – operational

No PI correction; standard Gaussian-residual sigma.

Default predict-op behaviour: prediction-interval bands derive from the fitted family’s residual variance σ²_ε (Gaussian approximation around the point forecast). This treats factor regressors and parameter estimates as if they were observed exactly. Appropriate for non-factor-augmented families (OLS, ridge, AR_p, etc.) or when factor estimation noise is negligible relative to residual variance.

When to use

Default for any family that does not estimate latent factors as regressors – the residual-variance band is honest in that case.

When NOT to use

Factor-augmented forecasts where estimated factors enter the regression – use bai_ng to inflate the band for the factor-estimation noise.

References

macroforecast design Part 2, L4: ‘forecasting model is the layer where every authoring iteration ends – pick family, tune, repeat.’
Bai & Ng (2006) ‘Confidence Intervals for Diffusion Index Forecasts and Inference for Factor-Augmented Regressions’, Econometrica 74(4): 1133-1150.

Related options: bai_ng

Last reviewed 2026-05-04 by macroforecast author.

`bai_ng` – operational

Bai-Ng (2006) generated-regressor PI correction.

Activates the Bai-Ng (2006) Theorem 3 + Corollary 1 correction to the prediction-interval sigma. The corrected sigma reflects (a) factor-estimation noise V₂/N where V₂ = β̂_F^T (Λ̂ diag(Σ̂_e) Λ̂^T / N) β̂_F, (b) parameter-estimation noise V₁/T from the OLS coefficient covariance evaluated at the last training factor row, and (c) the residual variance σ²_ε. Active only when the upstream fitted family is factor_augmented_ar; for any other family the predict op falls through to the uncorrected Gaussian-residual sigma.

When to use

Factor-augmented forecasts (FAR / FAVAR-style) where the band should be honest about factor-estimation noise on top of the usual parameter and residual uncertainty.

When NOT to use

Non-factor families – the correction is a no-op there. Use none to keep the predict op’s default behaviour.

References

macroforecast design Part 2, L4: ‘forecasting model is the layer where every authoring iteration ends – pick family, tune, repeat.’
Bai & Ng (2006) ‘Confidence Intervals for Diffusion Index Forecasts and Inference for Factor-Augmented Regressions’, Econometrica 74(4): 1133-1150.

Related options: none