You are on page 1of 3

*1.

Introduction

- VAR contain less than 10 variables
- good alternative: few predictors which pool the inforamtion from all candidate predictors
- diffusion indexes formally introduced by sargent and Sims (1977)
- modern dynamic general equilibrium macroeconomic models postulate that a small set of driving
variables is responsible for variation in macro time series -> these variables can be viewed as a set of
common factors.
- 2 step procedure: factors are estimated through PCA using Xt and these factors are then used to
forecast yt+1 (no need to interpret the factors themselves beacuse they are only used for forecasting).
- improvement over benchmark AR model is dramatic (MSE is 33% less for diffusion index model).

*2.1 An Approximate Dynamic Factor Model ->MODELUL

yt and xt have mean 0.

- acelasi, doar ca mai adauga si laguri ale lui yt ca variabila independenta
- facute doua modificari fata de varianta anterioara
-> lag polinomials have finite orders of at most q
-> empirical aplication focuses on an h-step-ahead forecast
=> the multi-step forecast is linear in Ft and yt and their lags and thus use a h-step-
ahead projection to costruct the forecasts directly

*2.2 Estimation and Forecasting

- 2 step procedure formally explained:
1) sample data Xt are used to estimate a time series of factors (the diffusion indexes)
2) we regress yt+1 on constant, Ft, and lags of yt -> the obtained parameter estimators are used
to forecast yt+h
- the feasible forecast is first-order efficient -> its mean square forecast error approaches the MSE of the
optimal infeasible forecast when N,T -> inf.
=> feasible forecasts are likely to be nearly optimal when N,T are large, regardless of the
ratio on N and T.

*2.3 Data Irregularities and Computational Issues

- EM algorithm for unbalanced data panel

*3.1 Forecasting Models and Data

- 1959:1 -> 1998:12 with 8 macrovariables to be forecasted
=> 4 variables of real economic activity: total industrial output, real personal income less
transfers, real manufacturing and trade sales, no of employees on nonagricultural payrolls
=> 4 price indexes:CPI, Personal consumption price deflator, CPI less food and energy, Producer
price index for finished goods

6M, 12M 24M horizons-> DI forecasts based on estimated factors, a benchmark univariate
autoregression and benchmark multivariate models
real variables modeled as I(1) in logarithms

DI FORECASTS

Y(T+h) = a+L(F(T)) + L(y(T))

k -> no of factors
m -> no of factor lags
p -> no of y lags

3 model types:
1) DI-AR LAG => lags of factors, lags of y
**1<=k<=4 ; 1<=m<=3 ; 0<=p<=6
2) DI-AR
**1<=k<=12 ; m=1 ; 0<=p<=6
3) DI
**1<=k<=12 ; m=1; p=0

3 preliminary steps towards analysing data:
-> possible to take logarithms (for all nonnegative series which were not rates or
percentages)
-> first differences: most series were first differeced
-> screening for outliers
-> stardardized
-> another outlier screening
using this data, three sets of empirical factors were constructed
1) PCA from 149 balanced variables
2) Unbalaced panel data -> 215 series
3) stacked balced variables
also VAR forecast with

You might also like