Time Series Project forecasting

Project Report
 Page. 1

INTRODUCTION
This report consists of Time Series analysis and forecasting of 2 datasets -

• DATASET 1 - Sales data of Rose Wine

• DATASET 2 - Sales data of Sparkling Wine
Please find the Jupyter Code Notebook here. Analysis code is in Python.

PROBLEM - Wine Sales

For this particular assignment, the data of different types of wine sales in the 20th century is to be
analysed.

Both of these data are from the same company but of different wines. As an analyst in the ABC
Estate Wines, you are tasked to analyse and forecast Wine Sales in the 20th century.

Datasets used - Sales of Rose Wine and Sparkling Wine

SYNOPSIS
1. Total No. Of Rose Data Entries = 187

Total No. Of Sparkling Data Entries = 187

2. No. Of Missing Values in Rose data = 2

- We use INTERPOLATION to impute these missing values

No. Of Missing Values in Sparkling data = 0

3. No. Of Duplicate entries in Rose data = 0

No. Of Duplicate entries in Sparkling data = 0

4. Both datasets are split in Train : Test at year 1991 - Test data starts at 1991

5. Various forecasting models applied are -

1. Linear Regression
2. Naive Bayes
3. Simple Average
4. 2-pt Moving Average
 Time Series Project

5. 4-pt Moving Average

6. 6-pt Moving Average
7. 9-pt Moving Average
8. Single Exponential Smoothing
9. Double Exponential Smoothing (Holt’s Model)
10. Triple Exponential Smoothing (Holt-Winter Model)
11. ARIMA / SARIMA (Auto fitted)
12. ARIMA / SARIMA (Manually fitted)

[ Q 1 ] Read the data as an appropriate Time Series data and plot the data.

• Both Datasets are read and stored as Pandas Data Frames for analysis

• Datasets are read as Time Series data using parse_dates=True &

index_col=‘YearMonth’

• First 5 rows of both the data are given below -

YearMonth Rose YearMonth Sparkling

1980-01-01 112.0 1980-01-01 1686

1980-02-01 118.0 1980-02-01 1591

1980-03-01 129.0 1980-03-01 2304

1980-04-01 99.0 1980-04-01 1712

1980-05-01 116.0 1980-05-01 1471

Rose Data Header Sparkling Data Header
• Rose Data plot -

Chetan Dudhane Page 3

 Page. 1

• Sparkling Data plot -

[ Q 2 ] Perform appropriate Exploratory Data Analysis to understand the data and

also perform decomposition.

Exploratory Data Analysis -

count mean std min 25% 50% 75% max

 Time Series Project

Rose 187.0 89.914 39.238 28.0 62.5 85.0 111.0 267.0

1070.
Sparkling 187.0 2402.417 1295.112 1605.0 1874.0 2549.0 7242.0
0
Descriptive Stats of Rose and Sparkling datasets
<class 'pandas.core.frame.DataFrame'> • Month-wise Boxplot of Rose -
DatetimeIndex: 187 entries,

1980-01-01 to 1995-07-01 Data

columns (total 1 columns): <class 'pandas.core.frame.DataFrame'>
# Column Non-Null Count Dtype DatetimeIndex: 187 entries,
--- ------ -------------- ----- 1980-01-01 to 1995-07-01 Data
0 Rose 187 non-null float64 columns (total 1 columns):
dtypes: float64(1) memory usage: 2.9 # Column Non-Null Count Dtype
KB --- ------ -------------- -----
0 Sparkling 187 non-null int64
dtypes: int64(1) memory
Info - Rose data usage: 2.9 KB

Info - Sparkling data

Chetan Dudhane Page 5

 Page. 1
• Month-wise Boxplot of Sparkling -

• Sales of both - Rose and Sparkling, show a spike in the last quarter of Oct to Dec

• Spike is much more accentuated in Sparkling sales

• This spike may be due to the Holiday season starting in Oct

YearMonth trend YearMonth seasonal YearMonth resid

1980-01-01 1980-01-01 -27.91 1980-01-01

Additive
1980-02-01 Decomposition of 1980-02-01
Rose - -17.44 1980-02-01

1980-03-01 1980-03-01 -9.29 1980-03-01

1980-04-01 1980-04-01 -15.10 1980-04-01

1980-05-01 1980-05-01 -10.20 1980-05-01

1980-06-01 1980-06-01 -7.68 1980-06-01

1980-07-01 147.08 1980-07-01 4.90 1980-07-01 -33.98

1980-08-01 148.13 1980-08-01 5.50 1980-08-01 -24.62

1980-09-01 148.38 1980-09-01 2.77 1980-09-01 53.85

1980-10-01 148.08 1980-10-01 1.87 1980-10-01 -2.96

1980-11-01 147.42 1980-11-01 16.85 1980-11-01 -14.26

1980-12-01 145.13 1980-12-01 55.71 1980-12-01 66.16

 Time Series Project

Rose Trend Rose Seasonality Rose Residual

Multiplicative Decomposition of Rose -

YearMonth trend YearMonth seasonal YearMonth resid

1980-01-01 1980-01-01 0.67 1980-01-01

1980-02-01 1980-02-01 0.81 1980-02-01

Chetan Dudhane Page 7

 Page. 1
1980-03-01 1980-03-01 0.90 1980-03-01

1980-04-01 1980-04-01 0.85 1980-04-01

1980-05-01 1980-05-01 0.89 1980-05-01

1980-06-01 1980-06-01 0.92 1980-06-01

1980-07-01 147.0 1980-07-01 1.06 1980-07-01 0.76

8
1980-08-01 148.1 1980-08-01 1.04 1980-08-01 0.84
3
1980-09-01 148.3 1980-09-01 1.02 1980-09-01 1.36
8
1980-10-01 148.0 1980-10-01 1.02 1980-10-01 0.97
8
1980-11-01 147.4 1980-11-01 1.19 1980-11-01 0.85
2
1980-12-01 145.1 1980-12-01 1.63 1980-12-01 1.13
3
Rose Trend Rose Seasonality Rose Residual

Additive Decomposition of Sparkling -

YearMonth trend YearMonth seasonal YearMonth resid

1980-01-01 1980-01-01 -854.26 1980-01-01

 Time Series Project

1980-02-01 1980-02-01 -830.35 1980-02-01

1980-03-01 1980-03-01 -592.36 1980-03-01

1980-04-01 1980-04-01 -658.49 1980-04-01

1980-05-01 1980-05-01 -824.42 1980-05-01

1980-06-01 1980-06-01 -967.43 1980-06-01

1980-07-01 2360.67 1980-07-01 -465.50 1980-07-01 70.84

1980-08-01 2351.33 1980-08-01 -214.33 1980-08-01 316.00

1980-09-01 2320.54 1980-09-01 -254.68 1980-09-01 -81.86

1980-10-01 2303.58 1980-10-01 599.77 1980-10-01 -307.35

1980-11-01 2302.04 1980-11-01 1675.07 1980-11-01 109.89

1980-12-01 2293.79 1980-12-01 3386.98 1980-12-01 -501.78

Sparkling Trend Sparkling Seasonality Sparkling Residual

Multiplicative Decomposition of Sparkling -

Chetan Dudhane Page 9

 Time Series Project

Sparkling Trend Sparkling Seasonality Sparkling Residual

• Additive Models -

• The seasonality is relatively constant over time

• yt = Trend + Seasonality + Residual

• Multiplicative Models -

• The seasonality increases or decreases over time. It is proportionate to the trend

YearMonth trend YearMonth seasonal YearMonth resid

1980-01-01 1980-01-01 0.65 1980-01-01

1980-02-01 1980-02-01 0.66 1980-02-01

1980-03-01 1980-03-01 0.76 1980-03-01

1980-04-01 1980-04-01 0.73 1980-04-01

1980-05-01 1980-05-01 0.66 1980-05-01

1980-06-01 1980-06-01 0.60 1980-06-01

1980-07-01 2360.6 1980-07-01 0.81 1980-07-01 1.03

7
1980-08-01 2351.3 1980-08-01 0.92 1980-08-01 1.14
3
1980-09-01 2320.5 1980-09-01 0.89 1980-09-01 0.96
4
1980-10-01 2303.5 1980-10-01 1.24 1980-10-01 0.91
8
1980-11-01 2302.0 1980-11-01 1.69 1980-11-01 1.05
4
1980-12-01 2293.7 1980-12-01 2.38 1980-12-01 0.95
9
• yt = Trend * Seasonality * Residual

• Here by just observing the Residual patterns of Additive and Multiplicative models of Rose
and Sparkling datasets. It seems that -

• Rose is Multiplicative

• Sparkling is Additive

Chetan Dudhane Page 10

 Time Series Project

[ Q 3 ] Split the data into training and test. The test data should start in 1991.

• Both datasets of Rose and Sparkling are split at the year 1991

• Test datasets start at 1991

• Rose dataset - TRAIN

Rose Train - First 5 rows

• Rose dataset - TEST

Chetan Dudhane Page 11

 Time Series Project

Rose Train - Last 5 rows

YearMonth Rose

1995-03-01 45.00

1995-04-01 52.00

1995-05-01 28.00

1995-06-01 40.00

1995-07-01 62.00

1980-05-01 1471.00
Sparkling Train - First 5

• Sparkling dataset - TEST

Chetan Dudhane Page 12

 Time Series Project

YearMonth Sparkling

1991-01-01 1902.00

1991-02-01 2049.00

1991-03-01 1874.00

1991-04-01 1279.00

1991-05-01 1432.00

YearMonth Sparkling

1990-08-01 1605.00

1990-09-01 2424.00

1990-10-01 3116.00

1990-11-01 4286.00

1990-12-01 6047.00
Sparkling Train - Last 5

YearMonth Sparkling

1995-03-01 1897.00

1995-04-01 1862.00

1995-05-01 1670.00

1995-06-01 1688.00

1995-07-01 2031.00

Sparkling Test - First 5 Sparkling Test - Last 5

[ Q 4 ] Build various exponential smoothing models on the training data and
evaluate the model using RMSE on the test data. Other models such as regression,
naïve forecast models, simple average models etc. should also be built on the
training data and check the performance on the test data using RMSE.

Chetan Dudhane Page 13

 Time Series Project

Model 1 - Linear Regression

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

Model 2 - Naive Bayes

Chetan Dudhane Page 14

 Time Series Project

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

Chetan Dudhane Page 15

 Time Series Project

Model 3 - Simple Average

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

SimpleAverageModel 53.46 1275.08

Model 4.A - Moving Average (Rose)

Chetan Dudhane Page 16

 Time Series Project

Chetan Dudhane Page 17

 Time Series Project

Test RMSE Rose

2pointTrailingMovingAverage 11.53

4pointTrailingMovingAverage 14.45

6pointTrailingMovingAverage 14.57

Chetan Dudhane Page 18

 Time Series Project

9pointTrailingMovingAverage 14.73

Model 4.B - Moving Average (Sparkling)

Chetan Dudhane Page 19

 Time Series Project

Test RMSE Sparkling

2pointTrailingMovingAverage 813.40

4pointTrailingMovingAverage 1156.59

6pointTrailingMovingAverage 1283.93

Chetan Dudhane Page 20

 Time Series Project

9pointTrailingMovingAverage 1346.28

Consolidated Moving Average Forecasts (Rose & Sparkling)

NOTE -

Chetan Dudhane Page 21

 Time Series Project

• We have built 4 models till now for both Rose and Sparkling Wine datasets

• We fitted various models to the Train split and Tested it on Test split. Accuracy metrics used is
Root Mean Squared Error (RMSE) on Test data

• Model 1 - Linear Regression (yt = β0 + β1Xt + ϵt)

• We regressed variables ‘Rose’ and ‘Sparkling’ against their individual time instances

• We modified the datasets and tagged individual sales to their time instances

• TEST RMSE ROSE = 15.27 | TEST RMSE SPARKLING = 1389.14

• Model 2 - Naive Approach (yt̂+1 = yt)

• Naive approach says that prediction for tomorrow is same as today

• And, prediction for day-after is same as tomorrow

• So, effectively all future predictions are going to be same as today

• TEST RMSE ROSE = 79.72 | TEST RMSE SPARKLING = 3864.28

(
• Model 3 - Simple Average yt̂+1 = yt̂+2 = ... = yt̂+n = Mean(y1,y2,...,yt))

• All future predictions are the same as the simple average of all data till today

• TEST RMSE ROSE = 53.46 | TEST RMSE SPARKLING = 1275.08

• Model 4 - Moving Average (MA)

• We calculate rolling means (Moving averages) over different intervals for the whole train data

• 2 Pt MA ====> means, we find average of 1st and 2nd to predict 3rd

similarly, average of 2nd and 3rd to predict 4th and so on

• 4 Pt MA ====> means, we find average of 1st, 2nd, 3rd & 4th to predict 5th

also, average of 2nd, 3rd, 4th & 5th to predict 6th and so on

• 2 PT MA ====>

TEST RMSE ROSE = 11.53 | TEST RMSE SPARKLING = 813.40

Chetan Dudhane Page 22

 Time Series Project

• 4 PT MA ====>

TEST RMSE ROSE = 14.45 | TEST RMSE SPARKLING = 1156.59

• 6 PT MA ====>

TEST RMSE ROSE = 14.57 | TEST RMSE SPARKLING = 1283.93

• 9 PT MA ====>

TEST RMSE ROSE = 14.73 | TEST RMSE SPARKLING = 1346.28

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

SimpleAverageModel 53.46 1275.08

2pointTrailingMovingAverage 11.53 813.40

4pointTrailingMovingAverage 14.45 1156.59

6pointTrailingMovingAverage 14.57 1283.93

9pointTrailingMovingAverage 14.73 1346.28

Consolidated Scores of Regression, Naive, Simple Average & Moving Average

• Till now, Best Model which gives lowest RMSE score for both Rose and Sparkling is ——> 2 Pt
Moving Average Model

• We’ll continue to forecast using Exponential Smoothing Models for both datasets of Rose and
Sparkling Wine Sales

• Exponential smoothing averages or exponentially weighted moving averages consist of

forecast based on previous periods data with exponentially declining influence on the older
observations

• Exponential smoothing methods consist of special case exponential moving with notation
ETS (Error, Trend, Seasonality) where each can be None(N), Additive (N), Additive damped
(Ad), Multiplicative (M) or Multiplicative damped (Md)

• One or more parameters control how fast the weights decay. The values of the parameters lie
between 0 and 1

• We’ll build following Exponential Smoothing Models -

• Single Exponential Smoothing with Additive Errors - ETS(A, N, N)

Chetan Dudhane Page 23

 Time Series Project

• Double Exponential Smoothing with Additive Errors, Additive Trends - ETS(A, A, N)

• Triple Exponential Smoothing with Additive Errors, Additive Trends, Additive Seasonality -
ETS(A, A, A)

• Triple Exponential Smoothing with Additive Errors, Additive Trends, Multiplicative Seasonality
- ETS(A, A, M)

• Triple Exponential Smoothing with Additive Errors, Additive DAMPED Trends, Additive
Seasonality - ETS(A, Ad, A)

• Triple Exponential Smoothing with Additive Errors, Additive DAMPED Trends, Multiplicative
Seasonality - ETS(A, Ad, M)
Single Exponential Smoothing with Additive Errors - ETS(A, N, N)

Chetan Dudhane Page 24

 Time Series Project

• For Rose - Level Parameter, Alpha = 0.09874

• For Sparkling - Level Parameter, Alpha = 0.07028

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

SimpleAverageModel 53.46 1275.08

2pointTrailingMovingAverage 11.53 813.40

4pointTrailingMovingAverage 14.45 1156.59

6pointTrailingMovingAverage 14.57 1283.93

9pointTrailingMovingAverage 14.73 1346.28

Simple Exponential Smoothing 36.80 1338.00

• Best Model till now for Rose and Sparkling ——- > 2 Pt Moving Average Model

Double Exponential Smoothing with Additive Errors, Additive Trends -

ETS(A, A, N)

Chetan Dudhane Page 25

 Time Series Project

• In Rose - DES has picked up the trend well. DES seems to perform better than SES here

• In Sparkling - DES shows a non-existent trend. DES does not perform well here

• Rose - Level parameter, Alpha = 0

Chetan Dudhane Page 26

 Time Series Project

Trend parameter, Beta = 0

• Sparkling - Level parameter, Alpha = 0.665

Trend parameter, Beta = 0.0001

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

SimpleAverageModel 53.46 1275.08

2pointTrailingMovingAverage 11.53 813.40

4pointTrailingMovingAverage 14.45 1156.59

6pointTrailingMovingAverage 14.57 1283.93

9pointTrailingMovingAverage 14.73 1346.28

Simple Exponential Smoothing 36.80 1338.00

Double Exponential Smoothing 15.27 5291.88

• Best Model till now for Rose and Sparkling ——- > 2 Pt Moving Average Model

Triple Exponential Smoothing with Additive Errors, Additive Trends, Additive Seasonality -
ETS(A, A, A)

Chetan Dudhane Page 27

 Time Series Project

• In Rose & Sparkling - TES has picked up the trend and seasonality very well

• Rose - Level parameter, Alpha = 0.0849

Trend parameter, Beta = 0.0

Seasonality parameter, Gamma = 0.00054

• Sparkling - Level parameter, Alpha = 0.11127

Trend parameter, Beta = 0.01236

Seasonality parameter, Gamma = 0.46071

Test RMSE Sparkling

Test RMSE Rose

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

SimpleAverageModel 53.46 1275.08

2pointTrailingMovingAverage 11.53 813.40

4pointTrailingMovingAverage 14.45 1156.59

6pointTrailingMovingAverage 14.57 1283.93

9pointTrailingMovingAverage 14.73 1346.28

Simple Exponential Smoothing 36.80 1338.00

Double Exponential Smoothing 15.27 5291.88

Chetan Dudhane Page 28

 Time Series Project

Triple Exponential Smoothing (Additive

Season) 14.24 378.63

• Till now, Best Model for Rose ——> 2 Pt Moving Average

Best Model for Sparkling ——> Holt-Winter - ETS (A, A, A)

Triple Exponential Smoothing with Additive Errors, Additive Trends,

Multiplicative Seasonality - ETS(A, A, M)

Chetan Dudhane Page 29

 Time Series Project

Chetan Dudhane Page 30

 Time Series Project

Rose - Level parameter, Alpha = 0.07736

Trend parameter, Beta = 0.03936

Seasonality parameter, Gamma = 0.00083

• Sparkling - Level parameter, Alpha = 0.07736

Trend parameter, Beta = 0.04943

Seasonality parameter, Gamma = 0.36205

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

SimpleAverageModel 53.46 1275.08

2pointTrailingMovingAverage 11.53 813.40

4pointTrailingMovingAverage 14.45 1156.59

6pointTrailingMovingAverage 14.57 1283.93

9pointTrailingMovingAverage 14.73 1346.28

Simple Exponential Smoothing 36.80 1338.00

Double Exponential Smoothing 15.27 5291.88

Triple Exponential Smoothing

14.24 378.63
(Additive Season)

Triple Exponential Smoothing

19.11 403.71
(Multiplicative Season)

• Till now, Best Model for Rose ——> 2 Pt Moving Average

Best Model for Sparkling ——> Holt-Winter - ETS (A, A, A)

Triple Exponential Smoothing with Additive Errors, Additive DAMPED

Chetan Dudhane Page 31
 Time Series Project

Trends, Additive Seasonality - ETS(A, Ad, A)

• Rose - Level parameter, Alpha = 0.07842

Trend parameter, Beta = 0.01153

Seasonality parameter, Gamma = 0.07738

Damping factor = 0.97503

Sparkling - Level parameter, Alpha = 0.10062

Chetan Dudhane Page 32
 Time Series Project

Trend parameter, Beta = 0.00018

Seasonality parameter, Gamma = 0.97025

Damping factor = 0.97025

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

SimpleAverageModel 53.46 1275.08

2pointTrailingMovingAverage 11.53 813.40

4pointTrailingMovingAverage 14.45 1156.59

6pointTrailingMovingAverage 14.57 1283.93

9pointTrailingMovingAverage 14.73 1346.28

Simple Exponential Smoothing 36.80 1338.00

Double Exponential Smoothing 15.27 5291.88

Triple Exponential Smoothing

14.24 378.63
(Additive Season)

Triple Exponential Smoothing

19.11 403.71
(Multiplicative Season)

Triple Exponential Smoothing (Additive

26.04 378.63
Season, Damped Trend)

• Till now, Best Model for Rose ——> 2 Pt Moving Average

Best Model for Sparkling ——> Holt-Winter - ETS (A, A, A)

Triple Exponential Smoothing with Additive Errors, Additive DAMPED Trends,

Multiplicative Seasonality - ETS(A, Ad, M)

Chetan Dudhane Page 33

 Time Series Project

Rose - Level parameter, Alpha = 0.05921

Trend parameter, Beta = 0.02025

Seasonality parameter, Gamma = 0.00405

Damping factor = 0.97723

Chetan Dudhane Page 34

 Time Series Project

• Sparkling - Level parameter, Alpha = 0.11107

Trend parameter, Beta = 0.03702

Seasonality parameter, Gamma = 0.39507

Damping factor = 0.98999

Test RMSE Rose Test RMSE Sparkling

RegressionOnTime 15.27 1389.14

NaiveModel 79.72 3864.28

SimpleAverageModel 53.46 1275.08

2pointTrailingMovingAverage 11.53 813.40

4pointTrailingMovingAverage 14.45 1156.59

6pointTrailingMovingAverage 14.57 1283.93

9pointTrailingMovingAverage 14.73 1346.28

Simple Exponential Smoothing 36.80 1338.00

Double Exponential Smoothing 15.27 5291.88

Triple Exponential Smoothing

14.24 378.63
(Additive Season)

Triple Exponential Smoothing

19.11 403.71
(Multiplicative Season)

Triple Exponential Smoothing

26.04 378.63
(Additive Season, Damped Trend)

Triple Exponential Smoothing

25.99 352.45
(Multiplicative Season, Damped Trend)

• We conclude that models with least RMSE,

Best Model for Rose ——> 2 Pt Moving Average

Best Model for Sparkling ——> Holt-Winter Damped Trend

ETS (A, Ad, M)

Best Models for Rose and Sparkling -

Chetan Dudhane Page 35

 Time Series Project

Chetan Dudhane Page 36

 Time Series Project

[ Q 5 ] Check for the stationarity of the data on which the model is being built on
using appropriate statistical tests and also mention the hypothesis for the
statistical test.

If the data is found to be non-stationary, take appropriate steps to make it

stationary.

Check the new data for stationarity and comment. Note: Stationarity should be
checked at alpha = 0.05.

To Check Stationarity of Data -

• We use Augmented Dicky - Fuller (ADF) Test to check the Stationarity of Data

• Hypotheses of ADF Test :

H0 : Time Series is not Stationary

Ha : Time Series is Stationary

• So for Industry standard (also given for this problem), the Confidence Interval is 95%

• Hence, alpha = 0.05

• So in ADF Test, if p-value < alpha ===> We reject the Null Hypothesis and hence
conclude that given Time Series is
Stationary

• So in ADF Test, if p-value > alpha ===> We fail to reject the Null Hypothesis and
hence conclude that given Time Series is
Not Stationary

• If Time Series is not Stationary then we apply one level of differencing and check for
Stationarity again.

• Again, if the Time Series is still not Stationary, we apply one more level of differencing and
check for Stationarity again

• Generally, with max 2 levels of differencing, Time Series becomes Stationary

• Once the Time Series is Stationary then we are ready to apply ARIMA / SARIMA models

Chetan Dudhane Page 37

 Time Series Project

Stationarity of Rose Wine Dataset -

• Augmented Dicky-Fuller Test was applied to the whole Rose dataset

• We found, p-value = 0.4671

• Here, p-value > alpha=0.05

• We fail to reject the Null Hypothesis and hence conclude that Rose Wine Time Series is Not
Stationary

• We take 1 level of differencing and check again for Stationarity

• Now, p-value = 3.0159e-11 ≈ 0.00

• Now, p-value < alpha=0.05

• Now, we reject the Null Hypothesis and conclude that Rose Time Series is Stationary with a
lag of 1

Stationarity of Sparkling Wine Dataset -

• Augmented Dicky-Fuller Test was applied to the whole Sparkling dataset

• We found, p-value = 0.70559

Chetan Dudhane Page 38

 Time Series Project

•
• Here, p-value > alpha=0.05

We fail to reject the Null Hypothesis and hence conclude that Sparkling Wine Time Series is
Not Stationary

• We take 1 level of differencing and check again for Stationarity

• Now, p-value = 0.00

• Now, p-value < alpha=0.05

• Now, we reject the Null Hypothesis and conclude that Sparkling Time Series is Stationary
with a lag of 1

[ Q 6 ] Build an automated version of the ARIMA/SARIMA model in which the

parameters are selected using the lowest Akaike
Information Criteria (AIC) on the training data and evaluate this model on the test
data using RMSE

ARIMA / SARIMA Models -

Chetan Dudhane Page 39

 Time Series Project

• ARIMA is an acronym for Auto-Regressive Integrated Moving Average

• SARIMA stands for Seasonal ARIMA, when the TS has seasonality

Chetan Dudhane Page 40

 Time Series Project

•
ARIMA / SARIMA are forecasting models on Stationary Time Series

ARIMA / SARIMA Modelling on Train Rose & Sparkling Data -

• We check for stationarity of Train Rose & Sparkling data by using Augmented Dicky Fuller
Test

• We take a difference of 1 and make both these datasets Stationary

• We apply the following iterations to both these datasets -

1. ARIMA Automated

2. SARIMA Automated

1. ARIMA Automated -

• We create a grid of all possible combinations of (p, d, q)

• Range of p = Range of q = 0 to 3, Constant d = 1

• Few Examples of the grid -

Model: (0, 1, 2)
Model: (0, 1, 3)
Model: (1, 1, 0)
Model: (1, 1, 1)
Model: (1, 1, 2)
Model: (1, 1, 3)
Model: (2, 1, 0)
Model: (2, 1, 1)
Model: (2, 1, 2)
Model: (2, 1, 3)
Model: (3, 1, 0)
Model: (3, 1, 1)

• We fit ARIMA models to each of these combinations for both datasets

• We choose the combination with the least Akaike Information Criteria (AIC)

• We fit ARIMA to this combination of (p, d, q) to the Train set and forecast on the Test set

• Finally, we check the accuracy of this model by checking RMSE of Test set

Test RMSE Rose Test MAPE Rose

Chetan Dudhane Page 41

 Time Series Project

•
ARIMA(2,1,3) 36.81 75.84
For Rose, Best Combination with Least AIC is - (p, d, q) —-> (2, 1, 3)

ARIMA (2, 1, 3) Diagnostic Plot - Rose

Test RMSE Sparkling Test MAPE Sparkling

Chetan Dudhane Page 42

 Time Series Project

•
ARIMA(2,1,2) 1299.98 47.10
For Sparkling, Best Combination with Least AIC is - (p, d, q) —-> (2, 1, 2)

ARIMA (2, 1, 2) Diagnostic Plot - Sparkling

Chetan Dudhane Page 43

 Time Series Project
2. SARIMA Automated -

• We create a grid of all possible combinations of (p, d, q) along with Seasonal

(P, D, Q) & Seasonality of 12 (for both datasets)

• Range of p = Range of q = 0 to 3, Constant d = 1

• Range of Seasonal P = Range of Seasonal Q = 0 to 3, Constant D = 1, Seasonality m = 12

• Few Examples of the grid (p, d, q) (P, D, Q, m) -

Model: (0, 1, 2)(0, 0, 2, 12)
Model: (0, 1, 3)(0, 0, 3, 12)
Model: (1, 1, 0)(1, 0, 0, 12)
Model: (1, 1, 1)(1, 0, 1, 12)
Model: (1, 1, 2)(1, 0, 2, 12)
Model: (1, 1, 3)(1, 0, 3, 12)
Model: (2, 1, 0)(2, 0, 0, 12)
Model: (2, 1, 1)(2, 0, 1, 12)
Model: (2, 1, 2)(2, 0, 2, 12)
Model: (2, 1, 3)(2, 0, 3, 12)
Model: (3, 1, 0)(3, 0, 0, 12)

• We fit SARIMA models to each of these combinations and select with least AIC

• We fit SARIMA to this best combination of (p, d, q) (P, D, Q, m) to the Train set and forecast
on the Test set. Then, we check accuracy using RMSE on Test set

• For Rose, Best Combination with Least AIC is - (3, 1, 1) (3, 0, 2, 12)

Test RMSE Rose Test MAPE Rose

Chetan Dudhane Page 44

 Time Series Project
ARIMA(2,1,3) 36.81 75.84
SARIMA (3, 1, 1) (3, 0, 2, 12) 18.88 36.38
SARIMA (3, 1, 1) (3, 0, 2, 12) Diagnostic Plot - ROSE

• For Sparkling, Best Combination with low AIC and low Test RMSE is - (3, 1, 1) (3, 0,
0, 12)

SARIMA (3, 1, 1) (3, 0, 0, 12) Diagnostic Plot - SPARKLING

Chetan Dudhane Page 45

 Time Series Project

Test RMSE Sparkling Test MAPE Sparkling

ARIMA(2,1,2) 1299.98 47.10

SARIMA (3, 1, 1) (3, 0, 0, 12) 601.24 25.87
• Till Now, Best Model for Rose with Least RMSE —-> SARIMA (3, 1, 1) (3, 0, 2, 12)

• Till Now, Best Model for Sparkling with Least RMSE ->SARIMA (3, 1, 1) (3, 0, 0, 12)

[ Q 7 ] Build ARIMA/SARIMA models based on the cut-off points of ACF and PACF
on the training data and evaluate this model on the test data using RMSE.

Auto-Correlation Function (ACF) -

• Autocorrelation refers to how correlated a time series is with its past values. e.g. yt with yt−1

also yt+1 with yt and so on.

• ‘Auto’ part of Autocorrelation refers to Correlation of any time instance with its previous time
instance in the SAME Time Series

• ACF is the plot used to see the correlation between the points, up to and including the lag
unit

Chetan Dudhane Page 46

 Time Series Project
• ACF indicates the value of ‘q’ - which is the Moving Average parameter in ARIMA / SARIMA
models

Partial Auto-Correlation Function (PACF) -

• Partial Autocorrelation refers to how correlated a time series is with its past lag values.

• For example, let lag=k, then Partial Autocorrelation is Correlation of yt with yt−k, ignoring the

effects of all the instances between yt and yt−k

• PACF is the plot used to see the correlation between the lag points

• PACF indicates the value of ‘p’ - which is the Auto-Regressive parameter in ARIMA / SARIMA
models

ACF & PACF of Rose -

• Observing the cutoffs in ACF and PACF plots for Rose dataset, we get -

FOR ARIMA —-> p = 2, q = 2 and difference d = 1

FOR SARIMA —-> p = 2, q = 2, d = 1 and P = 2, D = 1, Q = 2, Seasonality=12

Chetan Dudhane Page 47

 Time Series Project

ACF & PACF of Sparkling -

• Observing the cutoffs in ACF and PACF plots for Sparkling dataset, we get - FOR

ARIMA —-> p = 0, q = 0 and difference d = 1

FOR SARIMA —-> p = 0, q = 0, d = 1 and P = 0, 1, 2, 3 | D = 0, Q = 1, 2, 3

Chetan Dudhane Page 48

 Time Series Project

1. ARIMA Manual - Rose - (2, 1, 2)

ARIMA (2, 1, 2) Diagnostic Plot - ROSE

Chetan Dudhane Page 49

 Time Series Project

Test RMSE Rose Test MAPE Rose

ARIMA( 2, 1, 2) 36.87 76.06

2. SARIMA Manual - Rose - (2, 1, 2) (2, 1, 2, 12)

SARIMA (2, 1, 2) (2, 1, 2, 12) Diagnostic Plot - ROSE

Chetan Dudhane Page 50

 Time Series Project

Test RMSE Rose Test MAPE Rose

ARIMA( 2, 1, 2) 36.87 76.06

SARIMA (2, 1, 2) (2, 1, 2, 12) 16.55 25.48

3. SARIMA Manual - Rose - (2, 1, 2) (3, 1, 2, 12)

Test RMSE Rose Test MAPE Rose

ARIMA( 2, 1, 2) 36.87 76.06

SARIMA (2, 1, 2) (2, 1, 2, 12) 16.55 25.48

Chetan Dudhane Page 51

 Time Series Project
SARIMA (2, 1, 2) (3, 1, 2, 12) 15.36 22.96
SARIMA (2, 1, 2) (3, 1, 2, 12) Diagnostic Plot - ROSE

4. ARIMA Manual - Sparkling - (0, 1, 0)

ARIMA (0, 1, 0) Diagnostic Plot - SPARKLING

Chetan Dudhane Page 52

 Time Series Project

Test RMSE Sparkling Test MAPE Sparkling

ARIMA( 0, 1, 0) 3864.28 201.33

5. SARIMA Manual - Sparkling - (0, 1, 0) (1, 1, 1, 12)

Test RMSE Sparkling Test MAPE Sparkling

ARIMA( 0, 1, 0) 3864.28 201.33

Chetan Dudhane Page 53

 Time Series Project

SARIMA (0, 1, 0) (1, 1, 1, 12) 1189.84 54.87

SARIMA (0, 1, 0) (1, 1, 1, 12 Diagnostic Plot - SPARKLING

6. SARIMA Manual - Sparkling - (0, 1, 0) (2, 1, 2, 12)

Test RMSE Sparkling Test MAPE Sparkling

ARIMA( 0, 1, 0) 3864.28 201.33

SARIMA (0, 1, 0) (1, 1, 1, 12) 1189.84 54.87

Chetan Dudhane Page 54

 Time Series Project

SARIMA (0, 1, 0) (2, 1, 2, 12) 1757.73 81.79

SARIMA (0, 1, 0) (2, 1, 2, 12 Diagnostic Plot - SPARKLING

7. SARIMA Manual - Sparkling - (0, 1, 0) (3, 1, 2, 12)

SARIMA (0, 1, 0) (3, 1, 2, 12 Diagnostic Plot - SPARKLING

Chetan Dudhane Page 55

 Time Series Project

Test RMSE Sparkling Test MAPE Sparkling

ARIMA( 0, 1, 0) 3864.28 201.33

SARIMA (0, 1, 0) (1, 1, 1, 12) 1189.84 54.87

SARIMA (0, 1, 0) (2, 1, 2, 12) 1757.73 81.79

SARIMA (0, 1, 0) (3, 1, 2, 12) 1551.65 71.57

• In all Manual methods, Best Model for Rose with Least RMSE

—-> SARIMA (2, 1, 2) (3, 1, 2, 12)

• In all Manual methods, Best Model for Sparkling with Least RMSE

—-> SARIMA (0, 1, 0) (1, 1, 1, 12)

• Seasonal P and Q - it was difficult to gauge the correct values here as the data was not enough
and cutoffs were not visible

• Hence, we tried multiple combinations of Seasonal P and Q as given above

[ Q 8 ] Build a table (create a data frame) with all the models built along with their
corresponding parameters and the respective RMSE values on the test data.

All Models built with ROSE (sorted by RMSE) -

Chetan Dudhane Page 56

 Time Series Project

Model Parameters Test RMSE Rose

2pointTrailingMovingAverage 11.53

α = 0.0849 β =
Triple Exponential Smoothing
(Additive Season)
5.52e−6 ≈ 0.00 γ = 14.24
0.00054
4pointTrailingMovingAverage 14.45

6pointTrailingMovingAverage 14.57

9pointTrailingMovingAverage 14.73

RegressionOnTime 15.27

Double Exponential Smoothing 15.27

≈
SARIMA(2,1,2)(3,1,2,12) 15.36

SARIMA(3, 1, 1)(3, 0, 2, 12) 18.88

α = 0.07736
Triple Exponential Smoothing
(Multiplicative Season)
β = 0.03936 19.11
γ = 0.00083
α = 0.05921
Triple Exponential Smoothing
(Multiplicative Season, Damped Trend)
β = 0.0205 γ 25.99
= 0.00405
α = 0.07842
Triple Exponential Smoothing
(Additive Season, Damped Trend)
β = 0.01153 26.04
γ = 0.07738
Simple Exponential Smoothing α = 0.09874 36.80

ARIMA(2,1,3) 36.81

ARIMA(2,1,2) 36.87

SimpleAverageModel 53.46

NaiveModel 79.72

All Models built with SPARKLING (sorted by RMSE) -

Test RMSE
Model Parameters
Sparkling

Chetan Dudhane Page 57

 Time Series Project

α = 0.11107
Triple Exponential Smoothing
(Multiplicative Season, Damped Trend)
β = 0.03702 352.45
γ = 0.39507

α = 0.1112
Triple Exponential Smoothing
(Additive Season)
β = 0.01236 378.63
γ = 0.46071

α = 0.10062
Triple Exponential Smoothing
(Additive Season, Damped Trend)
β = 0.00018 378.63
γ = 0.51151

α = 0.11119
Triple Exponential Smoothing
(Multiplicative Season)
β = 0.04943 403.71
γ = 0.36205

SARIMA(3,1,1)(3,0,2,12) 601.24

2pointTrailingMovingAverage 813.40

4pointTrailingMovingAverage 1156.59

SARIMA(0,1,0)(3,1,2,12) 1189.84

SimpleAverageModel 1275.08

6pointTrailingMovingAverage 1283.93

ARIMA(2,1,2) 1299.98

Simple Exponential Smoothing α = 0.07028 1338.00

9pointTrailingMovingAverage 1346.28

RegressionOnTime 1389.14

SARIMA(0,1,0)(3,1,2,12) 1551.65

SARIMA(0,1,0)(2,1,2,12) 1757.73

NaiveModel 3864.28

ARIMA(0,1,0) 3864.28

α = 0.665 β
Double Exponential Smoothing 5291.88
= 0.0001

Chetan Dudhane Page 58

 Time Series Project
[ Q 9 ] Based on the model-building exercise, build the most optimum model(s) on
the complete data and predict 12 months into the future with appropriate
confidence intervals/bands.
• Best Models as per the Least RMSE on ROSE Test set ——>

- 2 Pt Trailing Moving Average

- Triple Exponential Smoothing

(Additive Seasonality)

• Best Model as per the Least RMSE on SPARKLING Test set ——>

Triple Exponential Smoothing

(Multiplicative Season, Damped Trend)

α = 0.11107, β = 0.03702, γ = 0.39507

Rose Forecast Next 12 months - 2 Pt Moving Average

• This model doesn’t seem to be predicting very well

• Hence, forecasting on the second best model - Triple Exponential Smoothing ETS(A, A, A) -
Additive Seasonality

Forecast

1995-08-01 42.50

Chetan Dudhane Page 59

 Time Series Project
1995-09-01 46.75

1995-10-01 44.63

1995-11-01 45.69

1995-12-01 45.16

1996-01-01 45.42

1996-02-01 45.29

1996-03-01 45.36

1996-04-01 45.32

1996-05-01 45.34

1996-06-01 45.33

1996-07-01 45.33
Rose Forecast Values - Next 12 Months - 2 Pt Moving Average

Rose Forecast Next 12 months - Triple Exponential Smoothing ETS (A, A, A)

Forecast

1995-08-01 50.02
1995-09-01 46.50

1995-10-01 45.94

1995-11-01 60.72

1995-12-01 97.66

Chetan Dudhane Page 60

 Time Series Project
1996-01-01 13.89

1996-02-01 24.29

1996-03-01 31.91

1996-04-01 24.73

1996-05-01 28.12

1996-06-01 33.56

1996-07-01 44.09
Rose Forecast Values - Next 12 Months - TES - ETS(A, A, A)

Sparkling Forecast Next 12 months - Triple Exponential Smoothing

ETS (A, Ad, M) - Damped Trend, Multiplicative Seasonality

Forecast

1995-08-01 1931.44
1995-09-01 2351.98

1995-10-01 3179.46

1995-11-01 3918.07

1995-12-01 5985.90

1996-01-01 1357.57

1996-02-01 1599.15

1996-03-01 1830.31

1996-04-01 1791.02

Chetan Dudhane Page 61

 Time Series Project
1996-05-01 1641.84

1996-06-01 1556.37

1996-07-01 1966.00
Sparkling Forecast Values - Next 12 Months -
TES - ETS(A, Ad, M)

[ Q 10 ] Comment on the model thus built and report your findings and
suggest the measures that the company should be taking for future sales.

Rose Wine Sales - Comments :

• Rose wine shows a clear trend of declining sales since 1980

- This shows decline in popularity of this variant of wine

• Also, there is a clear spike in sales seen in the last quarter of every year from Oct to Dec

- This might be due to the Holiday season in this period

- Highest peak in sales is seen in Dec every year

• There is also an instant crashing slump in sales in the first quarter of every year from Jan

- This might be due to the after effect or hangover of Holidays

Chetan Dudhane Page 62

 Time Series Project

•
Sales slowly pick up only after May-June

Rose Wine Sales - Forecast Models :

• Top 2 best models as per lowest Test RMSE were found to be - 2 Pt Moving Average and
Holt-Winters - Additive Seasonality & Trend

• 2 Pt Moving Average model, when used for forecasting do not seem to give good
predictions. Forecast values level out after a few iterations

• Holt-Winters seems to give a consistent forecast with respect to the data

• Hence, for final forecast of Rose Wine Sales - we choose Holt-Winters

Rose Wine Sales - Suggestions :

• Firstly, Holiday season is around the corner and forecast shows increasing sales and sharp
peak in Dec. Hence, Company should stock up

• But Declining sales of Rose Wine over the long period should be investigated with more
data crunching

• Company can rebrand its Rose variant along-with a new Wine-master

Chetan Dudhane Page 63
 Time Series Project

•
Company should take advantage of the oncoming spike from Aug-Oct by introducing
aggressive offers and Ad campaigns.

• This will entice first time Wine drinkers and fence sitters (who don’t have specific loyalties
to any particular brand)

• Still if there is no significant upward trend in sales by this Dec, then Company has 2 options
- invest in R&D or think of discontinuing this variant and come up with something
completely new

Sparkling Wine Sales - Comments :

• Sparkling wine sales don’t show any upward or downward trend - This

shows flat sales over long term range

• Also, there is very high spike in sales seen in the last quarter of every year from Oct to Dec

- This might be due to the Holiday season in this period

- Highest peak in sales is seen in Dec every year

- Dec sales are almost 3 times of Sep sales

Chetan Dudhane Page 64

 Time Series Project

•
Similar to Rose Wine, even in Sparkling sales, an instant crashing slump is seen in the first
quarter of every year from Jan
- This might be due to the after effect or hangover of Holidays

• Sales slowly pick up only from Jul-Aug

Sparkling Wine Sales - Forecast Models :

• Triple Exponential Smoothing - Holt-Winters Models perform the best on Sparkling

datasets, considering the least RMSE on Test data

• There has been incremental improvements in Test RMSE with each tuning of parameters

• Finally, for forecast of Sparkling Wine Sales - we choose Holt-Winters with Multiplicative
Seasonality and Additive Damped Trend

Sparkling Wine Sales - Suggestions :

• Even for Sparkling, Holiday season is around the corner and forecast shows increasing sales
and sharp peak in Dec. Hence, Company should stock up

• Sparkling wine has great holiday sales, so this shows popularity.

• So no need to introduce any offers here but hammering Ads are suggested in these times
of Oct-Dec. This will drive sales even further.

• Sparkling wines are generally associated with celebrations and mainly to burst open.

• A special designer bottle can be introduced at a cheaper price just for bursting. This will
maximise profits

• Year on Year sales do not show any significant increase or decrease

• Though, Holiday spikes are extreme, but general Year on Year sales need to be investigated
more. Early period from Jan should be used to do this deep dive

—————— END OF PROJECT ——————

Chetan Dudhane Page 65