Time Series Analysis
Project Report
Presented By:
Rim Moalla
Nour Fourati
Rayda Mallek
May Boulaares Elharzi
(Senior Finance/BA)
Submission Date: 20/01/2023
1
I. Introduction
The report contains the implementation of Box-Jenkins Method on the “Brent Spot
Price” dataset which includes a four-stage process as follows:
1. Model Specification
2. Parameter Estimation
3. Model Checking
4. Forecasting
II. Model Identification
1. Checking Stationarity and Seasonality
◆ The initial data’s plot suggests non-constant mean with time-dependent
variance:
◆ The graph displays low prices during the first period but starting from
2005 it shows a significant increase with high fluctuations.
➔ Our initial data is not stationary.
➔ It is essential to transform it into a stationary form prior to analysis.
2
● The first difference’s plot shows a significant transformation of our initial data:
○ The absence of the previous clear trend.
○ Stabilized mean and variance; Yet, we can observe a slight drop in the
mean in the second half of our period with the presence of few outliers.
➔ Visually, it is safe to conclude that our data is stationary.
● Implementing the necessary stationarity tests leads to a more sure conclusion
about our data’s stationarity:
○ Unit Root Tests:
For both tests, ADF and PP we have:
H0: One unit root (i.e. Non Stationarity)
3
H1: Zero unit root (i.e. Stationarity)
➢ The results show the two p-values = 0.01 < 0.05
➢ We reject H0 => Stationary data.
○ Stationarity Test:
For KPSS, we have:
H0: Stationary
H1: Not Stationary
➢ The results show a p-value = 0.1 > 0.05
➢ We fail to reject H0 => Stationary data.
⇒ According to the three tests coupled with the first difference’s plot, it is
safe to assume that our data is stationary.
2. Choosing model specification
Confidence interval = [-2/√396;+2/√396]
● The ACF plot cuts
off after lag 1, the remaining
autocorrelations fall within the
interval (dashed lines).
=> We can assume that the
model MA(1) is appropriate.
● The PACF plot shows
only one significant spike at lag 1
and all the other spikes fall within
the interval (dashed lines)
=> We can assume that the
model AR(1) is appropriate.
4
III. Parameter Estimation
AR(1) MA(1) ARMA(1,1) ARMA (1,7) ARMA(2,1) ARMA (2,7)
a0 0.09 0.09 0.09 0.10 0.09 0.10
(0.36) (0.30) (0.34) (0.17) (0.19) (0.18)
ɸ1 0.37 0.27 0.91 1.28 -0.09
(0.04) (0.11) (0.13) (0.10) (0.19)
ɸ2 -0.39 0.78
(0.04) (0.16)
𝛉1 0.35 0.12 -0.52 -0.89 0.49
(0.04) (0.11) (0.14) (0.10) (0.20)
𝛉2 -0.24 -0.63
(0.07) (0.24)
𝛉3 -0.15 -0.34
(0.06) (0.10)
𝛉4 -0.06 -0.20
(0.06) (0.06)
𝛉5 0.06 -0.005
(0.05) (0.08)
𝛉6 -0.059 -0.01
(0.06) (0.08)
𝛉7 0.05 -0.02
(0.063) (0.07)
AIC 2308.23 2312.03 2309.04 2310.62 2304.04 2309.81
Q(1) 0.25 0.65 0.002 0.00 0.00 0.00
Q(12) 12.19 16.06 11.09 3.05 6.60 4.56
● Based on the models’ AIC, we can consider ARMA(2,1) and AR(1) as our
candidate models as they demonstrate the minimum AICs, 2304.04 and 2308.23
respectively.
5
IV. Model Checking
In order to examine the independence of the different models estimated we will base our
conclusions on the Ljung-Box test statistic which includes:
H0: The residuals are independently distributed.
H1: The residuals are not independently distributed (i.e. they exhibit serial
correlation)
AR(1) ARMA(2,1)
P-value (lag =1) 0.6169 0.9976
P-value (lag=12) 0.4303 0.8828
Conclusion Both are greater than 0.05: Both are greater than 0.05:
➢ Fail to reject H0. ➢ Fail to reject H0.
➔ Both models’ residuals are independent
➔ The candidate models conform to the specifications of a stationary univariate
process.
V. Forecasting
1. Forecasts Characteristics
We based our forecasts on 50 observations out of 396 variables (our total sample).
Forecasted AR(1) Forecasted ARMA(2,1)
Mean -0.095 -0.033
Variance 5.095 5.697
2. Forecasts’ Accuracy
6
● Almost all the forecast’s accuracy measures (except for the RMSE
measure) present lower values for the model ARMA(2,1) than for the
AR(1) model.
➔ It is safe to conclude that the ARMA(2,1) model best fits our data.
