INDIAN

 INSTITUTE  OF  SCIENCE  

STOCHASTIC HYDROLOGY
Lecture -17
Course Instructor : Prof. P. P. MUJUMDAR
Department of Civil Engg., IISc.
Summary  of  the  previous  lecture  

• Behavior of AR and MA process

• Parameter estimation
– Matlab function armax
• Model selection
– Maximum likelihood rule

2  
 ARIMA Models
Mean square error, MSE (Prediction approach):

• Using a portion of available data (N/2) estimate the

parameters of different models
• Forecast the series one step ahead by using the
candidate models
• Estimate the MSE corresponding to each model
• The model with least value of MSE is selected for
prediction

3  
 ARIMA Models
The one step ahead forecast for ARMA(p, q) is
p q
Xˆ t +1 = ∑ φ j X t − j + ∑ φ j et − j
j =1 j =1

The error for one step ahead forecast is

et +1 = X t +1 − Xˆ t +1

If the series consists on N observations, the first N/2

observations are used for parameter estimation and N/
2+1 to N are used for error series calculation.

4  
 ARIMA Models
The MSE for model is
N

∑ ei2
N
i= +1
2
MSE =
N 2

5  
 ARIMA Models
3. Model testing / Validation:

Calibration data Test data

X1 XT-2 XT-1 X XT+1 XN

T

First T values are used to build the model (say 50%

of the available data) and the rest of data is used to
validate the model.

All the tests are carried out on the residual series only.

6  
 ARIMA Models
The tests are performed to examine whether the
following assumptions used in building the model are
valid for the model selection
• The residual series has zero mean
• No significant periodicities are present in the
residual series
• The residual series is uncorrelated
⎛ m1 m2
⎞
et = X t − ⎜ ∑ φ j X t − j + ∑θ j et − j ⎟
Residual   ⎝ j =1 j =1 ⎠
Data  
Simulated from the model  

7  
 ARIMA Models
Validation tests are listed here

• Significance of residual mean

• Significance of periodicities
• Cumulative periodogram test or Bartlett s test
• White noise test
• Whittle s test
• Portmanteau test

8  
 Significance of residual mean
Significance of residual mean:

• This test examines the validity of the assumption

that the error series e(t) has zero mean
• A statistic η(e) is defined as
1/2
N e
η ( e ) = 1/2
ρˆ
Where
e is the estimate of the residual mean
ρ̂ is the estimate of the residual variance
Ref: Kashyap R.L. and Ramachandra Rao.A, Dynamic stochastic models from empirical
data , Academic press, New York , 1976   9  
Significance of residual mean
• The statistic η(e) is approximately distributed as t
(α, N–1), where α is the significance level at
which the test is being carried out.
• If the value of η(e) < t(α, N–1), then the mean of
the residual series is not significantly different
from zero – series passes the test.

10  
 Significance of periodicities
Significance of periodicities:

• This test ensures that no significant periodicities

are present in the residual series
• The test is conducted for different periodicities
and the significance of each of the periodicities is
tested.
• A statistic η(e) is defined as
γ k2 ( N − 2 )
η (e) =
4 ρˆ1
γk corresponds to the periodicity being tested  
11  
 Significance of periodicities
Where γk2= αk2 + βk2
2
1 ⎡ N ⎤
ρˆ1 = ⎢∑ et − αˆ cos (ωk t ) − βˆ sin (ωk t )
{ } ⎥
N ⎢⎣ t =1 ⎥⎦
2 n
α k = ∑ et cos (ωk t )
N t =1
2 n
β k = ∑ et sin (ωk t )
N t =1
2π/ωk is the periodicity for which test is being carried
out.

12  
 Significance of periodicities
• The statistic η(e) is approximately distributed as
Fα(2, N–2 ), where α is the significance level at
which the test is being carried out.
• If the value of η(e) < Fα(2, N–2 ), then the
periodicity is not significant.

13  
 Bartlett s test
Cumulative periodogram test or Bartlett s test :

• This test is also carried out to examine significant

periodicities in the residual series

• This test is more convenient computationally and

is preferred because of its ability to test all the
periodicities at a time.

14  
 Bartlett s test
N 2 N 2
2 ⎧ 2 ⎫ ⎧ 2 ⎫
γ = ⎨ ∑ et cos (ωk t )⎬ + ⎨ ∑ et sin (ωk t )⎬
k
⎩ N t =1 ⎭ ⎩ N t =1 ⎭
k = 1,2,……N/2
k
2
∑ j
γ
j =1
gk = N /2 0 < gk < 1  
2
∑ k
γ
k =1

The plot of gk vs k is called as cumulative periodogram

Ref: Kashyap R.L. and Ramachandra Rao.A, Dynamic stochastic models from empirical
data , Academic press, New York , 1976   15  
 Bartlett s test
• On the cumulative periodogram two confidence
limits (λ/(N/2)1/2) are drawn on either side of line
joining (0, 0) and (N/2, 1)
• The value of λ prescribed for 95% confidence limits
is 1.35 and for 99% confidence limits is 1.65
• If all the values of gk lie within the significance band,
there is no significant periodicities in the series.
• If a value of gk lies outside the significance band, the
periodicity corresponding to that value of gk is
significant.

16  
 Bartlett s test
1  

gk  
λ
N
2
λ
N
2

0     N/2    
k    

17  
 Whittle s test for white noise
White noise test (Whittle s test):

• This test is carried out to test the absence of

correlation in the series.
• The covariance rk at lag k of the error series e(t)
N
1
rk = ∑ee j j −k k = 0, 1, 2,…….kmax  
N −k j = k +1

• The value of kmax is normally chosen as 0.15N

Ref: Kashyap R.L. and Ramachandra Rao.A, Dynamic stochastic models from empirical
data , Academic press, New York   18  
Whittle s test for white noise
• The covariance matrix is

⎡ r0 r1 r2 . . rkmax ⎤
⎢ ⎥
⎢ r1 r0 r1 . . rkmax −1 ⎥
⎢ r ⎥
Γ n1 = ⎢ 2 ⎥
⎢ . ⎥
⎢ ⎥
⎢ . ⎥
⎢ rk rkmax −1 r0 ⎥⎦ k x k  
⎣ max max max

19  
Whittle s test for white noise
• A statistic η(e) is defined as
N ⎛ ρˆ0 ⎞
η (e) = ⎜ − 1⎟ n1 = kmax  
n1 − 1 ⎝ ρˆ1 ⎠
Where ρ̂ 0 is the lag zero correlation =1, and

det Γ n1
ρˆ1 =
det Γ n1−1
The matrix Γn1-1 is constructed by eliminating the
last row and the last column from the Γn1 matrix.

20  
Whittle s test for white noise
• The statistic η(e) is approximately distributed as
Fα(n1, N–n1 ), where α is the significance level at
which the test is being carried out.
• If the value of η(e) < Fα(n1, N–n1 ), then the
residual series is uncorrelated.

21  
Portmanteau test for white noise
White noise test (Portmanteau test):

• This test is also carried out to test the absence of

correlation in the series.
• This test also uses the covariance rk defined
earlier.
• A statistic η(e) is defined as
2
n1
⎛ rk ⎞
η ( e ) = ( N − n1) ∑ ⎜ ⎟
k =1 ⎝ r0 ⎠

Ref: Kashyap R.L. and Ramachandra Rao.A, Dynamic stochastic models from empirical
data , Academic press, New York   22  
Portmanteau test for white noise
• The statistic η(e) is approximately distributed as
χ2α(n1), where α is the significance level at which
the test is being carried out.
• The value of n1 is normally chosen as 0.15N
• If the value of η(e) < χ2α(n1), then the residual
series is uncorrelated.
• Kashyap & Rao(1976) have proved that the
Portmanteau test is uniformly inferior to
Whittle s test and recommended the latter for
applications.

Ref: Kashyap R.L. and Ramachandra Rao.A, Dynamic stochastic models from empirical
data , Academic press, New York   23  
 ARIMA Models
Data Generation:
Consider AR(1) model,
Xt = φ1Xt-1 + et

e.g., φ1 = 0.5 : AR(1) model is

Xt = 0.5Xt-1 + et Choose et terms with zero

mean and uncorrelated  

24  
 ARIMA Models
Say X1 = 3.0

X2 = 0.5*3.0 + 0.335
= 1.835

X3 = 0.5*1.835 + 1.226
= 2.14

And so on…

25  
 ARIMA Models
Consider ARMA(1, 1) model,
Xt = φ1Xt-1 + θ1et-1 + et

e.g., φ1 = 0.5, θ1 = 0.4 : ARMA(1, 1) model is written as

with zero mean and

uncorrelated  
Xt = 0.5Xt-1 + 0.4et-1 + et

Choose et-1 terms as

previous et and set initial
value as zero  
26  
 ARIMA Models
Say X1 = 3.0

X2 = 0.5*3.0 + 0.4*0 + 0.667

= 2.167

X3 = 0.5*2.167 + 0.4*0.667 + 1.04

= 2.39

X4 = 0.5*2.39 + 0.4*1.04 + 2.156

= 3.767 and so
on...
27  
 ARIMA Models
Data Forecasting:
Consider AR(1) model,
Xt = φ1Xt-1 + et

Expected value is considered.

E [ X t ] = φ1E [ X t −1 ] + E [et ]
Xˆ t = φ1 X t −1
Expected value of et is zero  

28  
 ARIMA Models
Consider ARMA(1, 1) model,
Xt = φ1Xt-1 + θ1et-1 + et

E[Xt] = φ1Xt-1 + θ1et-1 + 0

Error in forecast in the
previous period  

e.g., φ1 = 0.5, θ1 = 0.4: Forecast model is written as

Xt = 0.5Xt-1 + 0.4et-1

29  
 ARIMA Models
Say X1 = 3.0 Initial error assumed to
be zero  
Xˆ 2 = 0.5 × 3.0 + 0.4 × 0
= 1.5
X2 = 2.8
Error e2 = 2.8 – 1.5 = 1.3

Xˆ 3 = 0.5 × 2.8 + 0.4 ×1.3

= 1.92
Actual value to be used  

30  
 ARIMA Models
X3 = 1.8
Error e3 = 1.8 – 1.92 = -0.12

Xˆ 4 = 0.5 ×1.8 + 0.4 × ( −0.12 )

= 0.852

and so on.

31  
 Markov Chains
Markov Chains:
• Markov chain is a stochastic process with the
property that value of process Xt at time t depends
on its value at time t-1 and not on the sequence of
other values (Xt-2 , Xt-3,……. X0) that the process
passed through in arriving at Xt-1.

P [ X t X t −1 , X t −2 ,..... X 0 ] = P [ X t X t −1 ]
Single step Markov
chain  

32  
 Markov Chains

P ⎡⎣ X t = a j X t −1 = ai ⎤⎦

• The conditional probability gives the probability at

time t will be in state j , given that the process
was in state i at time t-1.
• The conditional probability is independent of the
states occupied prior to t-1.
• For example, if Xt-1 is a dry day, what is the
probability that Xt is a dry day or a wet day.
• This probability is commonly called as transitional
probability
33  
 Markov Chains
t
P ⎣ X t = a j X t −1 = ai ⎦ = Pij
⎡ ⎤

• Usually written as Pijt indicating the probability of a

step from ai to aj at time ‘t’.
• If Pij is independent of time, then the Markov chain
is said to be homogeneous.

i.e., Pijt = Pijt +τ v t and τ

the transitional probabilities remain same across

time
34  
 Markov Chains
Transition Probability Matrix(TPM):
t+1 1 2 3 . . m
t
1 ⎡ P11 P12 P13 . . P1m ⎤
⎢ P P22 P23 . . P2 m ⎥⎥
2
⎢ 21
3 ⎢ P31 ⎥
P= ⎢ ⎥
. ⎢ . ⎥
. ⎢ . ⎥
⎢ ⎥
m ⎢⎣ Pm1 Pm 2 Pmm ⎥⎦
m x m  
35  
 Markov Chains
m

∑P
j =1
ij =1 v j

• Elements in any row of TPM sum to unity

(stochastic matrix)
• TPM can be estimated from observed data by
tabulating the number of times the observed data
went from state i to j
• Pj (n) is the probability of being in state j in the
time step n .

36  
 Markov Chains
• pj(0) is the probability of being in state j in period
t = 0.

p(0) = ⎡⎣ p1(0) p2(0) . . pm(0) ⎤⎦ …. Probability

1× m vector at time 0

(n) ( n) (n) ( n ) ⎤
p = ⎣ p1
⎡ p2 . . pm ⎦ …. Probability
1× m vector at time
n
• Let p(0) is given and TPM is given
(1) (0)
p = p ×P

37  
 Markov Chains
⎡ P11 P12 P13 . . P1m ⎤
⎢ P P22 P23 . . P2 m ⎥⎥
⎢ 21
p (1) = ⎡⎣ p1(0) p2(0) . . pm(0) ⎤⎦ ⎢ P31 ⎥
⎢ ⎥
⎢ . ⎥
⎢⎣ Pm1 Pm 2 Pmm ⎥⎦

= p1(0) P11 + p2(0) P21 + .... + pm(0) Pm1 …. Probability of

going to state 1

( 0) ( 0) ( 0)
= p1 P12 + p2 P21 + .... + pm Pm2 …. Probability of
going to state 2
And so on…

38  
 Markov Chains
Therefore

(1) (1) (1) (1) ⎤

p = ⎣ p1
⎡ p2 . . pm ⎦
1× m

p( 2) = p(1) × P
= p(0) × P × P
(0)
= p × P2
In general,
(n) (0) n
p = p ×P
39  
 Markov Chains
• As the process advances in time, pj(n) becomes less
dependent on p(0)
• The probability of being in state j after a large
number of time steps becomes independent of the
initial state of the process.
• The process reaches a steady state ay very large n
n
p = p× P
• As the process reach steady state, TPM remains
constant

40  
 Example – 1
Consider the TPM for a 2-state (state 1 is non-rainfall day
and state 2 is rainfall day) first order homogeneous
Markov chain as
⎡0.7 0.3⎤
TPM = ⎢ ⎥
⎣ 0.4 0.6 ⎦
Obtain the
1. probability of day 1 is non-rainfall day / day 0 is
rainfall day
2. probability of day 2 is rainfall day / day 0 is non-
rainfall day
3. probability of day 100 is rainfall day / day 0 is non-
rainfall day
41  
 Example – 1 (contd.)
1. probability of day 1 is non-rainfall day / day 0 is
rainfall day
No rain rain
No rain ⎡0.7 0.3⎤
TPM =
rain ⎢ 0.4 0.6⎥⎦
⎣
The probability is 0.4

2. probability of day 2 is rainfall day / day 0 is non-

rainfall day

p( 2) = p (0) × P 2

42  
 Example – 1 (contd.)
( 2) ⎡0.7 0.3⎤
p = [0.7 0.3] ⎢ ⎥
⎣ 0.4 0.6 ⎦
= [0.61 0.39]
The probability is 0.39

3. probability of day 100 is rainfall day / day 0 is non-

rainfall day

(n) (0)
p = p × Pn

43  
 Example – 1 (contd.)
P2 = P × P
⎡0.7 0.3⎤ ⎡0.7 0.3⎤ ⎡ 0.61 0.39 ⎤
= ⎢ ⎥ ⎢ ⎥ = ⎢ ⎥
⎣ 0.4 0.6 ⎦ ⎣ 0.4 0.6 ⎦ ⎣ 0.52 0.48 ⎦
4 2 2 ⎡0.5749 0.4251⎤
P = P × P = ⎢ ⎥
⎣ 0.5668 0.4332 ⎦
8 4 ⎡ 0.5715
4 0.4285 ⎤
P = P × P = ⎢
⎣0.5714 0.4286 ⎥⎦
16 8 8 ⎡0.5714 0.4286 ⎤
P = P × P = ⎢
⎣0.5714 0.4286 ⎥⎦

44  
 Example – 1 (contd.)
Steady state probability
p = [0.5714 0.4286]

For steady state,

p = p × Pn
⎡0.5714 0.4286 ⎤
= [0.5714 0.4286] ⎢ ⎥
⎣ 0.5714 0.4286 ⎦
= [0.5714 0.4286]

45