Transfer function models

ST 434/534

Donald E.K. Martin

North Carolina St. U.

November 14, 2022

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 1 / 56
 Outline

1 Introduction

2 Some preliminaries

3 Relationship between the cross-correlation function and the

transfer function

4 Procedure for obtaining the transfer function model

5 An example: fitting a transfer function model to sales and

advertising data

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 2 / 56
 Introduction
A transfer function model is a model of the relationship between
an output series and one or more related input predictor series.

Here we consider the case where there is only one predictor

series.

A procedure is given to build a transfer function model. The

procedure involves (i) identifying a model to describe the input
series, (ii) identifying a preliminary model describing the output
series, and (iii) using the residuals from the preliminary model
for the output series to identify a model describing the errors of
the preliminary structure.

A more detailed description of the procedure for determining the

transfer function model will given as we proceed.
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 3 / 56
 Examples of applications

Sales as a function of advertising expenditures

Electricity consumption as a function of daily maximum

temperature

Percent CO2 emission as a function of input gas rate

Residual building costs as a function of housing price

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 4 / 56
 The model
Let Xt and Yt be stationary series, related by
Yt = ν(B)Xt + Nt ,
P∞
where ν(B) = j=−∞ νj B j is referred to as the transfer
function, and {Nt } is a noise series that is independent of the
input series {Xt } and that follows an ARMA model.
The coefficients νj are often called the impulse response weights.
P
The transfer function model is stable if |νj | < ∞, and causal
if νj = 0 for j < 0.
In a stable system, a bounded input always produces a bounded
output. In a causal system, the present output is only affected
by the current and past inputs.
We seek to identify the transfer function and an ARMA model
for the noise Nt in stable and causal systems.
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 5 / 56
 Representation of the transfer function
The transfer function ν(B) may contain an infinite number of
coefficients, which is a problem for estimation since we have only
a finite number of data points.
The problem is alleviated by representing the transfer function as
a rational function
ωs (B)B b
ν(B) = ,
δr (B)
where ωs (B) = ω0 − ω1 B − . . . − ωs B s ,
δr (B) = 1 − δ1 B − . . . − δr B r , and b is the delay that elapses
before the input Xt affects the output Yt .
ωs (B) and δr (B) have only a finite number of coefficients that
need to be estimated.
For a stable system, we assume that the roots of δr (B) lie
outside the unit circle.
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 6 / 56
 Nature of weights νj

Once ωs (B) and δr (B) are determined, the coefficients νj may be

obtained by equating the coefficients of B j on both sides of the
equation
δr (B)ν(B) = ωs (B)B b ,

or
(1 − δ1 B − . . . − δr B r )(ν0 + ν1 B + ν2 B 2 + . . .)
= (ω0 − ω1 B − . . . − ωs B s )B b .

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 7 / 56
 Thus we have


 0 if j < b,
 1 νj−1 + δ2 νj−2 + . . . δr νj−r + ω0
δ if j = b,


νj = δ1 νj−1 + δ2 νj−2 + . . . δr νj−r − ωj−b if j = b + 1, . . . , b + s,
δ ν + δ2 νj−2 + . . . δr νj−r if j > b + s.

 1 j−1




(1)
Then the weights νj consist of

b zero weights ν0 , ν1 , . . . , νb−1 ,

s − r + 1 weights νb , νb+1 , . . . , νb+s−r that do not follow a fixed
pattern (there are none of these if s < r ),
r weights νb+s−r +1 , νb+s−r +2 , . . . , νb+s that serve as starting
values for the difference equation δ(B)νj = 0, j > b + s.
Weights νj , j > b + s, that follow the pattern δ(B)νj = 0.
In practice, the values of r and s rarely exceed 2.
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 8 / 56
 An example of obtaining the ν weights from δ and
ω
If r = s = 1 and b = 2 we have

(1 − δ1 B)(ν0 + ν1 B + ν2 B 2 + . . .) = ω0 B 2 − ω1 B 3

so that 

 0 if j < 2,
 ω0 if j = 2,


νj = δ1 ω0 − ω1 if j = 3,
δν if j > 3.

 1 j−1




Thus the first b = 2 weights are 0, weight ν2 doesn’t follow a

pattern, and weight ν3 is the start-up value needed for the recursion
νj = δ1 νj−1 .
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 9 / 56
 The cross-correlation function
Time series Xt and Yt , t = 0, ±1, ±2, . . . are jointly stationary if
they are both individually stationary and the cross-covariance
function between Xt and Ys is a function of the time difference
t − s only, not on where we are in the sequences.

If that is the case, define

γXY (k) = E [(Xt − µX )(Yt+k − µY )]
for k = 0, ±1, ±2, . . . .

Dividing by the respective standard deviations gives the

cross-correlation function:
γXY (k)
ρXY (k) =
σX σY
for k = 0, ±1, ±2, . . . .

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 10 / 56
 Note that γXX (k) = γX (k) and ρXX (k) = ρX (k). However the
cross-covariance and cross-correlation functions are not
symmetric, i.e. γXY (−k) ̸= γXY (k). Instead,

γXY (k) = E [(Xt − µX )(Yt+k − µY )]

= E [(Yt+k − µY )(Xt − µX )] = γYX (−k).

Thus the cross-correlation function measures not only the

strength of the association between the two variables but also
the direction.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 11 / 56
 An example
Consider the AR(1) model (1 − ϕ1 B)Z̃t = at , where |ϕ1 | < 1.
We have
 
1
Z̃t+k = at+k = at+k + ϕ1 at+k−1 + ϕ21 at+k−2 + . . .
1 − ϕ1 B

The cross-covariance function between at and Zt is

 k 2
ϕ1 σa , if k ≥ 0,
γaz (k) = E [at Z̃t+k ] =
0 if k < 0

Recalling that Var [Zt ] = σa2 [1/(1 − ϕ21 )], the cross-correlation
function is
 kp
ϕ1 1 − ϕ21 , if k ≥ 0,
ρaz (k) =
0 if k < 0.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 12 / 56
 An ARMA model is a transfer function model

In general, an ARMA model ϕ(B)Z̃t = θ(B)at , which may be

written as
θ(B)
Z̃t = at ,
ϕ(B)

is a transfer function model without the noise term.

The input is at , the output Z̃t , and the transfer function is

θ(B)
ν(B) = ϕ(B) .

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 13 / 56
 Relationship between the cross-correlation function
and the transfer function
The transfer function model is

Yt+k = ν0 Xt+k + ν1 Xt+k−1 + ν2 Xt+k−2 + . . . + Nt+k .

Without loss of generality assume that µX = µY = 0. Multiplying

through by Xt on the left and taking expectations gives

γXY (k) = ν0 γXX (k) + ν1 γXX (k − 1) + ν2 γXX (k − 2) + . . .

Recall that γXN (k) = 0 for all k. Dividing through by σX σY ,

σX
ρXY (k) = [ν0 ρX (k) + ν1 ρX (k − 1) + ν2 ρX (k − 2) + . . .]. (2)
σY

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 14 / 56
 Untangling the contaminated relationship
Thus the relationship between the cross-correlation function and
the impulse response function {νj } is contaminated by the
autocorrelation structure of the input series {Xt }.

This is true even if r = 0 so that ν(B) contains only a finite

number of weights.

However, if the input series is white noise, i.e. if ρX (k) = 0 for

k ̸= 0, (2) reduces to
σY
νk = ρXY (k).
σX

In that case then νk is directly proportional to the

cross-correlation function ρXY (k).
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 15 / 56
 Pre-whitened input series
In the transfer function model

Yt = ν(B)Xt + Nt (3)

we assume that Xt follows an ARMA model

ϕx (B)Xt = θx (B)αt ,

where αt is white noise.

αt satisfies
ϕx (B)
αt = Xt ,
θx (B)
and is often called the pre-whitened input series.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 16 / 56
 Applying the same transformation to the output series Yt , we
obtain a filtered output series

ϕx (B)
βt = Yt .
θx (B)

ϕx (B) ϕx (B)
Let ϵt = N.
θx (B) t
Applying the same filter θx (B)
across (3) gives

ϕx (B) ϕx (B) ϕx (B)

Yt = ν(B) Xt + Nt
θx (B) θx (B) θx (B)
or
βt = ν(B)αt + ϵt . (4)

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 17 / 56
 The impulse response weights νk of the transfer function may
then be found using
σβ
νk = ραβ (k). (5)
σα

(5) is due to the fact that the input series αt in transfer function
model (4) is white noise.

This then leads to a general procedure for estimating the

transfer function model that will be considered next.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 18 / 56
 Procedure for obtaining a transfer function model

Pre-whiten the input series to obtain

ϕx (B)
αt = Xt ,
θx (B)

where αt is mean zero white noise with variance σα2 . Apply the
same filter and calculate the filtered output series

ϕx (B)
βt = Yt .
θx (B)

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 19 / 56
 Step 2
Calculate the sample cross-correlation between αt and βt to
estimate νk ,
σ̂β
ν̂k = ρ̂αβ (k).
σ̂α

Identify b, r , and s by matching the pattern of ν̂k with known

theoretical patterns of the νk . (Recall the discussion of the
nature of the weights νk , and that those weights are proportional
to ραβ (k).)

Once b, r and s are chosen, estimates of

δr (B) = 1 − δ1 B − . . . − δr B r and
ωs (B) = (ω0 − ω1 B − . . . − ωs B s )B b may be obtained.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 20 / 56
 Determining b, r , and s
The value of b indicates the period when the input series begins
to affect the output series. It is determined by the fact that
νb = 0 for j < b and νb ̸= 0.

Practice has shown that the first spike in the sample

cross-correlation is followed by a clear dying down pattern
determined by
δr (B)νj = 0, j > b + s.
(The last equation comes from equation (1), as it is the same as

νj = δ1 νj−1 + δ2 νj−2 + . . . δr νj−r if j > b + s.

δr (B) operates on coefficients νj , with j considered as a time

index.)

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 21 / 56
 Determining b, r , and s (continued)

The value s is then set to the number of lags that reside between
the first spike and the beginning of the dying down pattern
(similar to identifying the value of p in an ARMA model.)

If there were no noise, r would indicate the number of past Y

values used to predict the current value. The value of r is
typically set to 1 if the dying down pattern is as a damped
exponential, 2 if the decay is as a damped sinosoid, and 0 if
there is more of a ’cutting off’ pattern.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 22 / 56
 Obtaining estimates of the transfer function model

The polynomials δ and ω may then be obtained by plugging in

estimated values of νj in

(1 − δ1 B − . . . − δr B r )(ν0 + ν1 B + ν2 B 2 + . . .)
= (ω0 − ω1 B − . . . − ωs B s )B b .

(Solve for the δ’s and ω’s in (1).)

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 23 / 56
 Step 3

From step 2 we have a fitted preliminary model

ω̂s (B)B b
Yt = Xt .
δ̂r (B)

Next, determine an ARMA model for the residuals from the

fitted preliminary transfer function model.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 24 / 56
 Some remarks

In the construction of the model, it is assumed that the variables

Xt and Yt are stationary. For nonstationary time series, some
variance stabilization and differencing transformations should be
used to achieve stationarity before proceeding.

The procedures depend on a causal model (the crosscorrelation

function should be zero at negative lags.) If that is not the case,
then Yt is influenced by previous Xt ’s and also Xt is influenced
by previous Yt ’s. There is feedback, a much more complicated
situation where both the input and output series should be
pre-whitened. We will not deal with that situation here.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 25 / 56
 The variance of the sample crosscorrelation function at lag k is
of order n1 , where n is the data length. One can then compare
estimated values to their standard deviation to determine which
values are significantly different from zero.
When the effect of noise is considerable in a transfer function
model, the model should be kept as simple as possible. In other
words, the variation of input Xt should be large relative to that
of the noise to get good parameter estimates for the transfer
function.
Whereas αt and βt are used to identify the transfer function, the
original series is used for parameter estimation, forecasts, etc.
In the transfer function model, it is assumed that the white noise
from the noise series Nt is independent from the input series xt
and thus also from the pre-whitened series αt . Diagnostics
should be used to check these assumptions. We also need to
check that the noise model is adequate.
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 26 / 56
 An example: fitting a transfer function model to
sales and advertising data

In the data set (see the class moodle page)

Xt (second column) is advertising expenditures in month t (in

thousands of dollars);

Yt (third column) is total sales in month t (in thousands of

cases);

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 27 / 56
 SAS code for fitting transfer function model to
sales data
data sales;
input month $ x y;
cards;

1 116.44 202.66
2 119.58 232.91
...
; run;

proc arima data=sales;

identify var=x stationarity=(adf=2); run;
identify var=x(1); run;
estimate p=1 q=1 noconstant printall plot; run;
identify var=y(1) crosscor=(x(1)) nlag=10; run;
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 28 / 56
 SAS code (continued)

estimate input=(2$(1,2)/(1,2)x) noconstant

altparm maxit=30 backlim=-3 plot; run;

estimate p=2 input=(2$(1,2)/(1,2)x) noconstant

altparm maxit=30 backlim=-3 plot; run;
forecast lead=12; run;
quit;

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 29 / 56
 Modeling x

identify var=x stationarity=(adf=2); run;

identify var=x(1); run;
estimate p=1 q=1 noconstant printall plot; run;

First we look at the series x and its autocorrelations.

We test for a unit root using an AR(3) for X . (AR(6) is also a

possibility).

After differencing, we look at diagnostics for x(1) and fit an

ARMA(1,1) model to x(1).

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 30 / 56
 Figure: 1. Trend and correlation analysis for input advertising data

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 31 / 56
 Table: 1

Type Lags ρ Pr < ρ τ Pr < τ F Pr > F

Zero mean 0 -0.0884 0.6609 -0.20 0.6110
1 -0.1872 0.6385 -0.31 0.5707
2 -0.1508 0.6467 -0.36 0.5529
Single mean 0 -10.2864 0.1164 -2.31 0.1715 2.66 0.4004
1 -22.6356 0.0041 -3.31 0.0172 5.47 0.0279
2 -12.3884 0.0674 -2.37 0.1533 2.82 0.3618
Trend 0 -14.0267 0.1984 -2.78 0.2096 3.88 0.4114
1 -30.8908 0.0039 -3.87 0.0171 7.47 0.0244
2 -17.5154 0.0958 -2.75 0.2179 3.80 0.4276

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 32 / 56
 Figure: 2. Trend and correlation analysis for differenced advertising data

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 33 / 56
 Figure: 3. Residual correlation analysis for differenced advertising data
with arma(1,1) model fit

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 34 / 56
 Autocorrelation check for white noise after fitting model to x(1)
Table: 2

Lag χ2 DF Pr ρk
6 8.68 4 0.0697 -0.008 -0.066 -0.184 -0.143 0.141 0.062
12 12.22 10 0.2709 -0.157 -0.036 -0.023 -0.009 0.044 -0.061
18 17.86 16 0.3323 0.016 0.094 -0.095 0.038 -0.160 0.040
24 21.87 22 0.4677 0.134 -0.008 -0.066 0.058 0.039 -0.065

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 35 / 56
 identify var=y; run; identify var=y(1) crosscor=(x(1)) nlag=10; run;

We next difference y and compute crosscorrelations with x at

lags -10 to 10.

var=y(1) crosscor=(x(1)) nlag=10; run;

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 36 / 56
 Figure: 4. Trend and correlation analysis for output y

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 37 / 56
 Figure: 5. Trend and correlation analysis for differenced output y

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 38 / 56
 Figure: 6. Estimated crosscorrelations

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 39 / 56
 Examining the crosscorrelation plot

From the crosscorrelation plot we identify that b = 2, as the

first non-zero spike is at lag 2.

Notice that there are no spikes at negative lags. (Any such

spikes would strongly suggest that the model is not causal).

It is unclear how to set s. It depends on whether 0, 1 or 2 bars in

the crosscorrelation function at positions 2 and 3 are part of the
dying down period or not. We choose to say not, setting s = 2.

r = 2 seems to be the best choice, since the decay appears to be

sinusoidal.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 40 / 56
 estimate input=(2$(1,2)/(1,2)x) noconstant
altparm maxit=30 backlim=-3 plot; run;

input=(2$(1,2)/(1,2)x) means that b = 2, and the transfer

function has s = 2 (numerator) and r = 2 (denominator).

noconstant asks that there be no constant term in the model.

backlim=-3 computes back forecasts for improved estimation.

altparm says that an alternative parameterization is needed to

interpret the form of the model specified.

maxit=30 is needed to ensure that the estimation procedure

carries out sufficient iterations.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 41 / 56
 Figure: 7. Residual plot after transfer function fit

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 42 / 56
 estimate p=2 input=(2$(1,2)/(1,2)x) noconstant
altparm maxit=30 backlim=-3 plot; run;

estimate p=2 says that for the noise process, fit an AR(2),
which we chose after looking at plots.

All of the model coefficients were statistically significant.

White noise was not rejected for the residuals.

The residuals were not crosscorrelated with x.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 43 / 56
 Autocorrelation Check of Residuals after fitting AR(2) model to Nt
Table: 3

To Lag χ2 DF Pr ρk
6 7.69 4 0.1037 0.003 -0.088 0.061 -0.162 -0.195 -0.047
12 11.53 10 0.3181 0.022 -0.047 -0.025 0.143 0.034 0.107
18 23.52 16 0.1006 -0.086 0.088 -0.213 -0.082 0.192 -0.042
24 25.97 22 0.2528 0.049 -0.016 -0.122 0.005 -0.050 -0.013

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 44 / 56
 Correlation check of residuals with input x
Table: 4

To Lag χ2 DF Pr ρk
5 1.41 1 0.2346 -0.017 -0.104 0.032 0.019 0.009 -0.055
11 2.29 7 0.9422 0.031 -0.039 -0.052 -0.063 0.014 -0.020
17 6.35 13 0.9323 -0.062 0.151 -0.025 0.034 0.104 -0.077
23 8.80 19 0.9767 0.047 -0.015 -0.019 0.016 -0.011 -0.155

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 45 / 56
 forecast lead=12 requests forecasts 12 units ahead.

Note that the fitted ARMA(1,1) model for ∆xt is used to

forecast future x values.

If we want to specify values of future x, the values can be added

at the end of the data set, with empty y values.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 46 / 56
 Figure: 8. Forecasts of 12 values of Y

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 47 / 56
 Fitting a transfer function model to gas data

A gas furnace was employed in which air and methane were

combined to produce a gas that contained CO2 .

The air feed was kept constant, while the methane gas feed rate
was varied. Let X denotes methane gas feed rate, Y the CO2
content of the resulting gas.

The data set consists of 296 observations on the two variables.

(See Series J in the back of the class text, and the discussion in
Section 12.2).

We seek a transfer function model for these data.

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 48 / 56
 Figure: 9. Trend and correlation analysis for input gas data

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 49 / 56
 Autocorrelation check for white noise after fitting AR(3) model to x
Table: 5

Lag χ2 DF Pr ρk
6 10.03 3 0.0183 -0.043 0.070 0.053 -0.143 -0.008 0.055
12 19.48 9 0.0214 0.015 0.006 -0.060 0.041 0.140 -0.074

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 50 / 56
 Figure: 10. Residual correlation diagnostics after fitting an AR(3) model
to X

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 51 / 56
 Autocorrelation check for white noise after fitting p=3 q=(4) model
to X
Table: 6

Lag χ2 DF Pr ρk
6 4.22 2 0.1213 -0.024 0.056 -0.020 0.010 -0.090 0.040
12 12.91 8 0.1149 0.014 -0.027 -0.067 0.057 0.123 -0.069

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 52 / 56
 Figure: 11. Residual correlation diagnostics after fitting an AR3 MA(4)
model to X (one MA term at lag 4)

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 53 / 56
 Figure: 11. Cross correlation after fitting an AR3 q=(4) model to X (one
MA term at lag 4)

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 54 / 56
 Analysis
We first tried r = 2, taking it that the pattern of dying down is
sinusoidal, and s = 2, assuming that the dying down begins after
lag 5, which is 2 lags after the first spike. Clearly, we should take
b = 3, the location of the first spike.

Though the trend and correlation plot suggests an AR(3) model

for X, that model is not adequate. We also add in an MA factor
with a coefficient at lag 4, giving a model that fits better.

Not all coefficients of δ and ω are needed. We ended up fitting a

model with only a coefficient at lag 2 in the numerator (ω), and
only a coefficient at lag 1 in the denominator (δ).

In the end, the resulting model fits fairly well, though there is
still crosscorrelation of the noise series with X, an indication of a
transfer function model that isn’t quite adequate.
Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 55 / 56
 Figure: 12. Forecasts of CO2 content

Donald E.K. Martin (North Carolina St. U.) Transfer function models November 14, 2022 56 / 56