正在学习

10.7 CONCLUSION

Many methods have been developed for reducing the dimension of a set of conditioning predictor variables deemed too large to be subject to a comprehensive model search or even viable for parameter estimation. First and foremost among these are factor approaches which use a much smaller set of common factors, extracted from a large-dimensional list of variables, to predict the outcome variable of interest. Additional variables can be added to such factors, including own lags, which has given rise to factor-augmented vector autoregressions. Such models allow the simultaneous inclusion of ARMA terms and common factors and so can be used to combine the strength of two very different types of models.

In practice, forecasters are confronted with the decision whether to adopt model selection (regularization) methods that attempt to identify a small set of variables included in the preferred model, while the remaining variables get dropped, or, instead, to first aggregate the information in a large set of potential predictor variables in the form of common factors and then use these in the prediction model. Factoraugmented models represent a middle way in that they facilitate the inclusion of a few key predictors along with aggregate information extracted in some factors.

Which method works best will surely depend on the underlying data-generating process and how this relates to the variables in the forecaster’s information set. If the predicted variable is driven by a small number of variables contained in the forecaster’s information set, then regularization methods are likely to perform better than factor methods provided that these variables can be identified. This conclusion could be hampered, of course, if the predictors are strongly correlated, but in this case, the specific identity of the selected variables may not matter too much to the predictive accuracy since the predictor variables are close substitutes. Conversely, if the predicted variable depends on common factors, as opposed to individual predictors, then the true coefficients of relatively many predictors will be nonzero, and methods based on identifying a scarce set of predictors will not work well. For example, in the case with a large set of predictor variables, many of which are strongly correlated, factor methods that dilute the “noise” in the individual predictors can be expected to perform better than methods that attempt to identify a small set of individual predictors.

练习题

Which of the following best describes factor approaches for dimension reduction?

A. They use a large set of variables to predict the outcome variable directly.

B. They use a smaller set of common factors extracted from a large-dimensional list of variables to predict the outcome variable.

C. They ignore the correlation between variables and focus on individual predictors.

D. They are only useful when the number of predictors is very small.

What is the key feature of factor-augmented vector autoregressions (FAVAR)?

A. They exclude ARMA terms and focus solely on common factors.

B. They include only individual predictors without any aggregation.

C. They allow the simultaneous inclusion of ARMA terms and common factors.

D. They are limited to univariate time series analysis.

Which of the following are true about the decision between model selection (regularization) methods and factor aggregation?

A. Model selection methods attempt to identify a small set of variables for the preferred model.

B. Factor aggregation methods first aggregate information in a large set of predictors into common factors.

C. Model selection methods always perform better than factor aggregation methods.

D. Factor aggregation methods are only useful when predictors are uncorrelated.

Factor-augmented models can include both a few key predictors and aggregate information extracted from factors.

Regularization methods are likely to perform better than factor methods when the predicted variable depends on common factors rather than individual predictors.

If the predicted variable is driven by a small number of variables, regularization methods are likely to perform better than factor methods, provided that these variables can be ___.

In cases with a large set of predictor variables, many of which are strongly correlated, factor methods that dilute the “noise” in the individual predictors can be expected to perform better than methods that attempt to identify a small set of individual ___.

Explain why the performance of regularization methods might be hampered if predictors are strongly correlated.

What is the advantage of factor-augmented models over pure factor models or pure regularization methods?

Which of the following statements are true about the performance of factor methods?

A. Factor methods perform better when the predicted variable depends on common factors.

B. Factor methods are ineffective when predictors are strongly correlated.

C. Factor methods dilute the “noise” in individual predictors when the predictor set is large and correlated.

D. Factor methods always outperform regularization methods regardless of the data-generating process.

When the predicted variable is primarily influenced by a small number of predictors in the dataset, which method is likely to perform better, assuming these variables can be identified and are not highly correlated?

A. Factor methods that aggregate information into common factors

B. Regularization methods that select a small set of variables

C. Both methods will perform equally well

D. Neither method is suitable

Which of the following statements are true regarding the use of factor-augmented models in forecasting?

A. They allow the simultaneous inclusion of ARMA terms and common factors.

B. They are only effective when predictors are uncorrelated.

C. They can dilute the 'noise' in individual predictors when many are strongly correlated.

D. They are a middle way between model selection methods and factor aggregation.

E. They perform poorly when the predicted variable depends on common factors.

Factor methods are generally preferred over regularization methods when the predicted variable depends on common factors rather than individual predictors.

In forecasting, when dealing with a large set of predictor variables that are strongly correlated, ___ methods can be expected to perform better by diluting the 'noise' in the individual predictors.

登录后解锁笔记、知识点解析、AI 问答

立即登录