forecast

Forecast vector autoregression (VAR) model responses

collapse all in page

Syntax

Y = forecast(Mdl,numperiods,Y0)
Y = forecast(Mdl,numperiods,Y0,Name,Value)
[Y,YMSE] = forecast(___)

Description

example

Y = forecast(Mdl,numperiods,Y0) returns a path of minimum mean squared error (MMSE) forecasts (Y) over the length numperiods forecast horizon using the fully specified VAR(p) model Mdl. The forecasted responses represent the continuation of the presample data Y0.

example

Y = forecast(Mdl,numperiods,Y0,Name,Value) uses additional options specified by one or more name-value pair arguments. For example, you can specify future exogenous predictor data or include future responses for conditional forecasting.

example

[Y,YMSE] = forecast(___) returns the corresponding mean squared error (MSE) of each forecasted response using any of the input arguments in the previous syntaxes.

Examples

collapse all

Open Live Script

Fit a VAR(4) model to the consumer price index (CPI) and unemployment rate data. Then, forecast unconditional MMSE responses from the estimated model.

Load the Data_USEconModel data set.

load Data_USEconModel

Plot the two series on separate plots.

figure;
plot(DataTable.Time,DataTable.CPIAUCSL);
title('Consumer Price Index');
ylabel('Index');
xlabel('Date');

figure;
plot(DataTable.Time,DataTable.UNRATE);
title('Unemployment Rate');
ylabel('Percent');
xlabel('Date');

Stabilize the CPI by converting it to a series of growth rates. Synchronize the two series by removing the first observation from the unemployment rate series.

rcpi = price2ret(DataTable.CPIAUCSL);
unrate = DataTable.UNRATE(2:end);
Data = array2timetable([rcpi unrate],'RowTimes',DataTable.Time(2:end),...
    'VariableNames',{'rcpi','unrate'});

Create a default VAR(4) model using the shorthand syntax.

Mdl = varm(2,4);

Estimate the model using the entire data set.

EstMdl = estimate(Mdl,Data.Variables);

EstMdl is a fully specified, estimated varm model object.

Forecast responses from the estimated model over a three-year horizon. Specify the entire data set as presample observations.

numperiods = 12;
Y0 = Data.Variables;
Y = forecast(EstMdl,numperiods,Y0);

Y is a 12-by-2 matrix of forecasted responses. The first and second columns contain the simulated CPI growth rate and unemployment rate, respectively.

Plot the forecasted responses and the last 50 true responses.

fh = dateshift(Data.Time(end),'end','quarter',1:12);

figure;
h1 = plot(Data.Time((end-49):end),Data.rcpi((end-49):end));
hold on;
h2 = plot(fh,Y(:,1));
title('CPI Growth Rate');
ylabel('Growth rate');
xlabel('Date');
h = gca;
fill([Data.Time(end) fh([end end]) Data.Time(end)],h.YLim([1 1 2 2]),'k',...
    'FaceAlpha',0.1,'EdgeColor','none');
legend([h1 h2],'True','Forecast')
hold off;

figure;
h1 = plot(Data.Time((end-49):end),Data.unrate((end-49):end));
hold on;
h2 = plot(fh,Y(:,2));
title('Unemployment Rate');
ylabel('Percent');
xlabel('Date');
h = gca;
fill([Data.Time(end) fh([end end]) Data.Time(end)],h.YLim([1 1 2 2]),'k',...
    'FaceAlpha',0.1,'EdgeColor','none');
legend([h1 h2],'True','Forecast','Location','northwest')
hold off;

Open Live Script

Estimate a four-degree vector autoregression model including exogenous predictors (VARX(4)) of the consumer price index (CPI), the unemployment rate, and the gross domestic product (GDP). Include a linear regression component containing the current quarter and the last four quarters of government consumption expenditures and investment (GCE). Forecast a response path from the estimated model.

Load the Data_USEconModel data set. Compute the real GDP.

load Data_USEconModel
DataTable.RGDP = DataTable.GDP./DataTable.GDPDEF*100;

Plot all variables on separate plots.

figure;
subplot(2,2,1)
plot(DataTable.Time,DataTable.CPIAUCSL);
ylabel('Index');
title('Consumer Price Index');
subplot(2,2,2)
plot(DataTable.Time,DataTable.UNRATE);
ylabel('Percent');
title('Unemployment Rate');
subplot(2,2,3)
plot(DataTable.Time,DataTable.RGDP);
ylabel('Output');
title('Real Gross Domestic Product');
subplot(2,2,4)
plot(DataTable.Time,DataTable.GCE);
ylabel('Billions of $');
title('Government Expenditures');

Stabilize the CPI, GDP, and GCE by converting each to a series of growth rates. Synchronize the unemployment rate series with the others by removing its first observation.

inputVariables = {'CPIAUCSL' 'RGDP' 'GCE'};
Data = varfun(@price2ret,DataTable,'InputVariables',inputVariables);
Data.Properties.VariableNames = inputVariables;
Data.UNRATE = DataTable.UNRATE(2:end);

Expand the GCE rate series to a matrix that includes its current value and up through four lagged values. Remove the GCE variable from Data.

rgcelag4 = lagmatrix(Data.GCE,0:4);
Data.GCE = [];

Create a default VAR(4) model using the shorthand syntax.

Mdl = varm(3,4);
Mdl.SeriesNames = ["rcpi" "unrate" "rgdpg"];

Estimate the model using all but the last three years of data. Specify the GCE matrix as data for the regression component.

bfh = Data.Time(end) - years(3);
estIdx = Data.Time < bfh;
EstMdl = estimate(Mdl,Data(estIdx,:).Variables,'X',rgcelag4(estIdx,:));

Forecast a path of quarterly responses three years into the future.

Y0 = Data(estIdx,:).Variables;
Y = forecast(EstMdl,12,Data(estIdx,:).Variables,'X',rgcelag4(~estIdx,:));

Y is a 12-by-3 matrix of simulated responses. The columns correspond to the CPI growth rate, unemployment rate, and real GDP growth rate, respectively.

Plot the forecasted responses and the last 50 true responses.

figure;
for j = 1:Mdl.NumSeries
    subplot(2,2,j)
    h1 = plot(Data.Time((end-49):end),Data{(end-49):end,j});
    hold on;
    h2 = plot(Data.Time(~estIdx),Y(:,j));
    title(Mdl.SeriesNames{j});
    h = gca;
    fill([bfh h.XLim([2 2]) bfh],h.YLim([1 1 2 2]),'k',...
        'FaceAlpha',0.1,'EdgeColor','none');
    hold off;
end

hl = legend([h1 h2],'Data','Forecast');
hl.Position = [0.6 0.25 hl.Position(3:4)];

Open Live Script

Analyze forecast accuracy using forecast intervals over a three-year horizon. This example follows from Forecast Unconditional Response Series from VAR(4) Model.

Load the Data_USEconModel data set and preprocess the data.

load Data_USEconModel

rcpi = price2ret(DataTable.CPIAUCSL);
unrate = DataTable.UNRATE(2:end);
Data = array2timetable([rcpi unrate],'RowTimes',DataTable.Time(2:end),...
    'VariableNames',{'rcpi','unrate'});

Estimate a VAR(4) model of the two response series. Reserve the last three years of data.

bfh = Data.Time(end) - years(3);
estIdx = Data.Time < bfh;
Mdl = varm(2,4);
EstMdl = estimate(Mdl,Data(estIdx,:).Variables);

Forecast responses from the estimated model over a three-year horizon. Specify the entire data set as presample observations. Return the MSE of the forecasts.

numperiods = 12;
Y0 = Data(estIdx,:).Variables;
[Y,YMSE] = forecast(EstMdl,numperiods,Y0);

Y is a 12-by-2 matrix of forecasted responses. YMSE is a 12-by-1 cell vector of corresponding MSE matrices.

Extract the main diagonal elements from the matrices in each cell of YMSE. Apply the square root of the result to obtain standard errors.

extractMSE = @(x)diag(x)';
MSE = cellfun(extractMSE,YMSE,'UniformOutput',false);
SE = sqrt(cell2mat(MSE));

Estimate approximate 95% forecast intervals for each response series.

YFI = zeros(numperiods,Mdl.NumSeries,2);

YFI(:,:,1) = Y - 2*SE;
YFI(:,:,2) = Y + 2*SE;

Plot the forecasted responses and the last 50 true responses.

figure;
h1 = plot(Data.Time((end-49):end),Data.rcpi((end-49):end));
hold on;
h2 = plot(Data.Time(~estIdx),Y(:,1));
h3 = plot(Data.Time(~estIdx),YFI(:,1,1),'k--');
plot(Data.Time(~estIdx),YFI(:,1,2),'k--');
title('CPI Growth Rate');
ylabel('Growth rate');
xlabel('Date');
h = gca;
fill([bfh h.XLim([2 2]) bfh],h.YLim([1 1 2 2]),'k',...
    'FaceAlpha',0.1,'EdgeColor','none');
legend([h1 h2 h3],'True','Forecast','95% Forecast interval',...
    'Location','northwest')
hold off;

figure;
h1 = plot(Data.Time((end-49):end),Data.unrate((end-49):end));
hold on;
h2 = plot(Data.Time(~estIdx),Y(:,2));
h3 = plot(Data.Time(~estIdx),YFI(:,2,1),'k--');
plot(Data.Time(~estIdx),YFI(:,2,2),'k--');
title('Unemployment Rate');
ylabel('Percent');
xlabel('Date');
h = gca;
fill([bfh h.XLim([2 2]) bfh],h.YLim([1 1 2 2]),'k',...
    'FaceAlpha',0.1,'EdgeColor','none');
legend([h1 h2 h3],'True','Forecast','95% Forecast interval',...
    'Location','northwest')
hold off;

Input Arguments

collapse all

VAR model, specified as a varm model object created by varm or estimate. Mdl must be fully specified.

Forecast horizon, or the number of time points in the forecast period, specified as a positive integer.

Data Types: double

Presample responses that provide initial values for the forecasts, specified as a numpreobs-by-numseries numeric matrix or a numpreobs-by-numseries-by-numprepaths numeric array.

numpreobs is the number of presample observations. numseries is the number of response series (Mdl.NumSeries). numprepaths is the number of presample response paths.

Rows correspond to presample observations, and the last row contains the latest observation. Y0 must contain at least Mdl.P rows. If you supply more rows than necessary, forecast uses only the latest Mdl.P observations.

Columns must correspond to the response series names in Mdl.SeriesNames.

Pages correspond to separate, independent paths.

  • If you do not specify the YF name-value pair argument, then forecast initializes each forecasted path (page) using the corresponding page of Y0. Therefore, the output argument Y has numprepaths pages.

  • If you specify the YF name-value pair argument, then forecast takes one of these actions.

    • If Y0 is a matrix, then forecast initializes each forecast path (page) in YF using Y0. Therefore, all paths in the output argument Y derive from common initial conditions.

    • Otherwise, forecast applies Y0(:,:,j) to initialize forecasting path j. Y0 must have at least numpaths pages, and forecast uses only the first numpaths pages.

    Among all pages, observations in a particular row occur at the same time.

Data Types: double

Name-Value Pair Arguments

Specify optional comma-separated pairs of Name,Value arguments. Name is the argument name and Value is the corresponding value. Name must appear inside quotes. You can specify several name and value pair arguments in any order as Name1,Value1,...,NameN,ValueN.

Example: 'X',X,'YF',YF uses the matrix X as predictor data in the regression component, and the matrix YF as partially known future responses for conditional forecasting.

Forecasted time series of predictors to include in the model regression component, specified as the comma-separated pair consisting of 'X' and a numeric matrix containing numpreds columns.

numpreds is the number of predictor variables (size(Mdl.Beta,2)).

Rows correspond to observations. Row j contains the j-step-ahead forecast. X must have at least numperiods rows. If you supply more rows than necessary, forecast uses only the earliest numperiods observations. The first row contains the earliest observation.

Columns correspond to individual predictor variables. All predictor variables are present in the regression component of each response equation.

forecast applies X to each path (page); that is, X represents one path of observed predictors.

To maintain model consistency into the forecast horizon, it is a good practice to specify forecasted predictors when Mdl has a regression component.

By default, forecast excludes the regression component, regardless of its presence in Mdl.

Data Types: double

Future multivariate response series for conditional forecasting, specified as the comma-separated pair consisting of 'YF' and a numeric matrix or 3-D array containing numseries columns.

Rows correspond to observations in the forecast horizon, and the first row is the earliest observation. Specifically, row j in sample path k (YF(j,:,k)) contains the responses j periods into the future. YF must have at least numperiods rows to cover the forecast horizon. If you supply more rows than necessary, forecast uses only the first numperiods rows.

Columns correspond to the response variables in Y0.

Pages correspond to sample paths. Specifically, path k (YF(:,:,k)) captures the state, or knowledge, of the response series as they evolve from the presample past (Y0) into the future.

  • If YF is a matrix, then forecast applies YF to each of the numpaths output paths (see Y0).

  • Otherwise, YF must have at least numpaths pages. If you supply more pages than necessary, forecast uses only the first numpaths pages.

Elements of YF can be numeric scalars or missing values (indicated by NaN values). forecast treats numeric scalars as deterministic future responses that are known in advance, for example, set by policy. forecast forecasts responses for corresponding NaN values conditional on the known values.

By default, YF is an array composed of NaN values indicating a complete lack of knowledge of the future state of all responses in the forecast horizon. In this case, forecast estimates conventional MMSE forecasts.

For more details, see Algorithms.

Example: Consider forecasting one path of a VAR model composed of four response series three periods into the future. Suppose that you have prior knowledge about some of the future values of the responses, and you want to forecast the unknown responses conditional on your knowledge. Specify YF as a matrix containing the values that you know, and use NaN for values you do not know but want to forecast. For example, 'YF',[NaN 2 5 NaN; NaN NaN 0.1 NaN; NaN NaN NaN NaN] specifies that you have no knowledge of the future values of the first and fourth response series; you know the value for period 1 in the second response series, but no other value; and you know the values for periods 1 and 2 in the third response series, but not the value for period 3.

Data Types: double

Note

NaN values in Y0 and X indicate missing values. forecast removes missing values from the data by list-wise deletion. If Y0 is a 3-D array, then forecast performs these steps.

  1. Horizontally concatenate pages to form a numpreobs-by-numpaths*numseries matrix.

  2. Remove any row that contains at least one NaN from the concatenated data.

In the case of missing observations, the results obtained from multiple paths of Y0 can differ from the results obtained from each path individually.

For missing values in X, forecast removes the corresponding row from each page of YF. After row removal from X and YF, if the number of rows is less than numperiods, then forecast throws an error.

Output Arguments

collapse all

MMSE forecasts of the multivariate response series, returned as a numobs-by-numseries numeric matrix or a numobs-by-numseries-by-numpaths numeric array. Y represents the continuation of the presample responses in Y0. Rows correspond to observations, columns correspond to response variables, and pages correspond to sample paths. Row j is the j-period-ahead forecast.

If you specify future responses for conditional forecasting using the YF name-value pair argument, then the known values in YF appear in the same positions in Y. However, Y contains forecasted values for the missing observations in YF.

MSE matrices of the forecasted responses in Y, returned as a numperiods-by-1 cell vector of numseries-by-numseries numeric matrices. Cells of YMSE compose a time series of forecast error covariance matrices. Cell j contains the j-period-ahead MSE matrix.

YMSE is identical for all paths.

Because forecast treats predictor variables in X as exogenous and non-stochastic, YMSE reflects the error covariance associated with the autoregressive component of the input model Mdl only.

Algorithms

  • forecast estimates unconditional forecasts using the equation

    y^t=Φ^1y^t−1+...+Φ^py^t−p+c^+δ^t+β^xt,

    where t = 1,...,numperiods. forecast filters a numperiods-by-numseries matrix of zero-valued innovations through Mdl. forecast uses specified presample innovations (Y0) wherever necessary.

  • forecast estimates conditional forecasts using the Kalman filter.

    1. forecast represents the VAR model Mdl as a state-space model (ssm model object) without observation error.

    2. forecast filters the forecast data YF through the state-space model. At period t in the forecast horizon, any unknown response is

      y^t=Φ^1y^t−1+...+Φ^py^t−p+c^+δ^t+β^xt,

      where y^s, s < t, is the filtered estimate of y from period s in the forecast horizon. forecast uses specified presample values in Y0 for periods before the forecast horizon.

    For more details, see filter and [4], pp. 612 and 615.

  • The way forecast determines numpaths, the number of pages in the output argument Y, depends on the forecast type.

    • If you estimate unconditional forecasts, which means you do not specify the name-value pair argument YF, then numpaths is the number of pages in the input argument Y0.

    • If you estimate conditional forecasts and Y0 and YF have more than one page, then numpaths is the number of pages in the array with fewer pages. If the number of pages in Y0 or YF exceeds numpaths, then forecast uses only the first numpaths pages.

    • If you estimate conditional forecasts and either Y0 or YF has one page, then numpaths is the number of pages in the array with the most pages. forecast uses the array with one page for each path.

  • forecast sets the time origin of models that include linear time trends (t0) to size(Y0,1) – Mdl.P (after removing missing values). Therefore, the times in the trend component are t = t0 + 1, t0 + 2,..., t0 + numobs. This convention is consistent with the default behavior of model estimation in which estimate removes the first Mdl.P responses, reducing the effective sample size. Although forecast explicitly uses the first Mdl.P presample responses in Y0 to initialize the model, the total number of observations (excluding missing values) determines t0.

References

[1] Hamilton, James. D. Time Series Analysis. Princeton, NJ: Princeton University Press, 1994.

[2] Johansen, S. Likelihood-Based Inference in Cointegrated Vector Autoregressive Models. Oxford: Oxford University Press, 1995.

[3] Juselius, K. The Cointegrated VAR Model. Oxford: Oxford University Press, 2006.

[4] Lütkepohl, H. New Introduction to Multiple Time Series Analysis. Berlin: Springer, 2005.

See Also

Objects

Functions

Topics

Introduced in R2017a

Econometrics Toolbox Documentation

Support