Você está na página 1de 7

(1) (1 mark) Suppose that you have daily data and would like to remove the weekly seasonal

pattern through decomposition. In the first step of the decomposition, how many terms would
you include in the moving average to get a rough estimate of the trend-cycle component?
Why?
Answer: We should use an MA7. A moving average with the same number of terms as the
length of the seasonal pattern will filter out the seasonality. If it was monthly data an MA12
would be appropriate.
(2) (3 marks) Explain the mechanics of using training and test sets (also referred to as insample and out-of-sample data respectively) to evaluate the predictive accuracy of different
models. Why is this an appropriate way of evaluating predictive methods in contrast to
looking at how well a model fits all the available data? In the context of time series forecasting
models, how does computing real time forecasts (also referred to as rolling forecasting
origin) relate to using training and test sets for evaluating forecasting accuracy?
Answer: We split our available data into two sets: the training set and test set. We use the
training set to estimate the parameters of the model. We then predict the observations in the
test set using these estimated parameters. This procedure generates genuine predictions for
the test set since we use different data (the training set) to estimate the model. We evaluate
predictive accuracy only based on the test set.
Evaluating genuine predictions prevents overfitting: we can always improve the in-sample
fit by making the model more complicated. However achieving better in-sample fit does
not necessarily lead to better predictive accuracy as we may be fitting noise rather than
underlying patterns.
In a time series context the observations are ordered. Real time forecasting means that we
increase the training set with each observation. This mimics what we do in real time: using
the past to predict the future. Genuine forecasts.
(3) (2 marks) It is a well established empirical fact that stock returns display pronounced
time-varying (conditional) variance. This finding motivated the development of several methods to forecast the variance of stock returns, which is a critical task for asset pricing and risk
management in financial institutions. Exponentially weighted moving averages are often used
for this purpose by practitioners due to their simplicity and ease of computation. Consider
1

the following version of simple exponential smoothing for variance forecasting:


p
yt = ht1 t
ht = yt2 + (1 )ht1 ,
where yt is the stock return at period t and t is an i.i.d. error such that E(t ) = 0 and
E(2t ) = 1. Derive the one and two period ahead forecasts for the variance of yt based on
this model. That is, compute the formulas for Var(yt+1 |y1 , . . . , yt ) and Var(yt+2 |y1 , . . . , yt ).

Answer:

This question is a bit different from the others as it evaluates whether you can apply the
reasoning in a slightly modified context (which demonstrates that you actually understand
it). All the steps below are from basic stats. Not as much detail needed for full marks.

Note that because the error t has mean zero, yt has mean zero and Var(yt ) = E(yt2 ). First,
we can check that given that we know ht1 ,
Var(yt |ht1 ) = Var(

p
ht1 t ) = ht1 Var(t ) = ht1 E(2t ) = ht1

So for the first part,


Var(yt+1 |y1 , . . . , yt ) = Var(

p
ht t+1 |y1 , . . . , yt ) = E(ht 2t+1 |y1 , . . . , yt ) = ht

Next,

Var(yt+2 |y1 , . . . , yt ) = Var(

p
ht+1 t+2 |y1 , . . . , yt )

= E(ht+1 |y1 , . . . , yt )E(2t+1 )


= E(ht+1 |y1 , . . . , yt )
2
= E(yt+1
+ (1 )ht |y1 , . . . , yt )
2
= E(yt+1
|y1 , . . . , yt ) + (1 )ht

= E(ht 2t+1 |y1 , . . . , yt ) + (1 )ht


= ht + (1 )ht
= ht .

(4) (1 mark) Long term asset allocation also requires forecasts of the stock returns themselves.
There is a large literature on this topic and the degree of forecastability in stock returns is
a hotly debated topic as the size of the unpredictable component of returns swamps the size
of the predictable component. We can define the return of an asset at period t as follows:
yt =

Pt
1,
Pt1

where Pt is the price of the asset at day t.


The mean-squared error (MSE) is a common metric for evaluating the accuracy of stock return
forecasting methods. Why should we prefer the MSE over metrics based on percentage errors
(such as the mean absolute percentage error) in this case?
Answer: Percentage errors are unstable when the forecasted variable can be close to zero
(division by a small number) and invalid when the forecasted variable can be zero (since
we cannot divide by zero). An asset return can be zero, hence percentage errors are not
appropriate.

(5) Consider the simple exponential smoothing model:

yt = `t1 + t
`t = yt + (1 )`t1 ,
where t N (0, 2 ) and 0 < 1.
(i) (2 marks) Define what is a stationary time series process and explain whether the above
process stationary or non-stationary. (Hint: think about the random walk example in class).
Answer: A time series process is stationary when the joint distribution of Yt , Yt+1 , . . . , Yt+k
does not depend on t. In the case of the simple exponential smoothing model, the error
correction formulation
`t = `t+1 + t
shows that the variance of `t , and consequently the variance of yt , increases by 2 2 at every
period. Therefore the variance of yt is a function of t. The process is non-stationary.
(ii) (2 marks) Derive the one and two step ahead 95% forecasting intervals for the simple
exponential smoothing model above.

Answer: Then 95% forecasting interval for a prediction h steps ahead is

p
ybt+h 1.96 Var(yt+h |y1 , . . . , yt ).
From the lectures, we have
ybt+1 = E(yt+1 |y1 , . . . , yt ) = `t
Var(yt+1 |y1 , . . . , yt ) = 2
E(yt+2 |y1 , . . . , yt ) = E(`t+1 + t+2 |y1 , . . . , yt ) = E(`t+1 |y1 , . . . , yt ) = `t

Var(yt+2 |y1 , . . . , yt ) = Var(`t+1 + t+2 |y1 , . . . , yt )


= Var(`t+1 |y1 , . . . , yt ) + 2
= Var(`t + t1 |y1 , . . . , yt ) + 2
= 2 2 + 2
= 2 (1 + 2 )
(iii) (1 mark) Why are we often interested in constructing forecasting intervals in addition to
point forecasts?
Answer: Because we may want to quantify the uncertainty in our forecasts or report a range
of plausible values for the random quantity that we are trying to forecast.
(iv) (1 mark) How would you compute the interval forecasts in practice? Note that when you
start your empirically analysis the parameters and 2 are not given. Provide details such
as the relevant expressions.
Answer: We estimate by minimising the sum of the squares of the residuals (the in-sample
prediction errors):

SSE =

n
X
(yt lt1 )2
t=1

Let et be the residuals based on the estimated . We estimate 2 as

b2 =

n
X
t=1

t
,
n1

where we divide by n 1 because we estimated one parameter () before.


We then compute the forecasting intervals based on the estimated parameters.

(6) Consider the dampened trend exponential smoothing model:

yt+1 = `t + bt + t+1 ,
`t = yt + (1 )(`t1 + bt1 ),
bt = (`t `t1 ) + (1 )bt1 ,
where t N (0, 2 ) , 0 1, 0 1, 0 1.
(i) (1 mark) What are the interpretations of bt and ?
Answer: bt is the drift component (or the linear growth rate), which is the increase (or
decrease) in level in each period (a series with a positive trend will have a positive bt ). is
a drift dampening parameter that slows down the trend as we project it to the future.
(ii) (1 mark) What is possible advantage of using the dampened trend exponential smoothing
method for forecasting over the trend corrected (Holt) exponential smoothing model?
Answer: The Hold method extrapolates the current trend indefinitely into the future when
forecasting, which may judge not to be realistic. The dampened method exponentially slows
down the trend when forecasting. Eventually the drift goes to zero.
(iii) (2 marks) Derive the error correction formulation for the `t and bt recursions.
Answer: This is identical to a lecture slide but with a before bt1 .

lt = yt + (1 )(lt1 + bt1 )
= lt1 + bt1 + (yt lt1 bt1 )
= lt1 + bt1 + t

bt = (lt lt1 ) + (1 )bt1


= bt1 + (lt lt1 bt1 )
= bt1 + (yt lt1 bt1 )
= bt1 + t
(iv) (1 mark) How does deriving the error correction formulation as in (iii) helps with understanding the forecasting method?
Answer: The error correction form allows us to write the future values of yt a linear function
of future errors, therefore a linear function of independent variables. That makes it simple
to derive the forecasts and variances for any number of periods ahead.

(7) (1 mark) List some of the most useful residual diagnostics for estimated time series models
and explain how they can assist with the modelling process.
Answer: There is a range of answers here but the ones that we emphasised the most were the
residual plot, the ACF (autocorrelation function plot) and the histogram of residuals. Our
models assume that the errors are independent and additionally (to construct forecasting
intervals) that they are normal. The residual diagnostics allows us to verify whether these
assumptions hold. For example, the presence of significant autocorrelations would suggest
that there are still patterns in the residuals and that we should improve the model accordingly.

(8) (1 mark) Identify and give a brief description of the problem that arises when regressing
non-stationary time series on each other (for example, regressing a time series with a trend
on another time series with a trend).

Answer: Regressing non-stationary time series is problematic as it leads to spurious regressions: we may observe significant regression coefficients and high R2 even if the variables are
totally unrelated. This problem does not go away with more observations. One example is
if two series have a trend: there is positive correlation between the population of Australia
and the number of cat videos uploaded in the US in the past 10 years.

Você também pode gostar