Individual Assignment

Time Series Forecasting

Q1) A graph of the time series

Sales (₹1000s)
300
250
200
150
100
50
0
July

July

July
June

June

June
January
February

January
May

February

January
May

February

October
March

August

October
November

March

August

October
November

May
March

August

November
April

April

April
December

December

December
September

September

September
First year Second year Third year

Sales Linear (Sales)

From the sales graph, it is clear that the sales are seasonal (similar pattern across years) and we
can also notice a slightly decreasing trend.
Therefore a 12-month moving average method is suitable for forecasting sales data here.
Q2) Seasonal Index

Month Seasonal Index

Jan 1.442987341
Feb 1.29916017
Mar 1.343561546
Apr 1.040768871
May 1.048950468
Jun 0.800047992
Jul 0.827962384
Aug 0.852620041
Sep 0.627726178
Oct 0.700031676
Nov 0.852404864
Dec 1.15878992

Observations:
The seasonal indexes do make an intuitive sense, as observed below
❖ High sales at the end of the year (Dec-holiday season)
❖ High sales at the beginning of the year (Jan, Feb, Mar-Spring season abundance of fish)
❖ Relatively low sales during the middle months ( Monsoon & Rainy season-Low fishing)
Q3) Sales forecast for 4rth year (Dummy Variable Method)
Month Sales Forecast( ₹1000s)
Jan 299.0209
Feb 270.5439
Mar 281.163
Apr 218.8619
May 221.6542
Jun 169.8759
Jul 176.649
Aug 182.7809
Sep 135.2105
Oct 151.5002
Nov 185.3475
Dec 253.1521

Q4) Analyzing January Sales data

Actual Sale = ₹295,000
Forecasted sale = ₹299,020.9
----------------------------------------

Difference (Forecasted (-) Actual) = ₹4020.9

Error Percentage = (₹4020.9/₹295,000)*100 = 1.36%

The error is only ₹4020.9, a meager 1.36%, which is very good. In reality, the forecasted values
are rarely this close to actuals. Hence Andy should not worry about the uncertainty in this
forecasting.

But, to instill confidence in Andy, we can compare the January historical sales data.

Jan(Y2) (-) Jan(Y1) = ₹21000

Jan(Y3) (-) Jan(Y2) = ₹19000
Jan(Y4-Forecasted) (-) Jan(Y3) = ₹170209.2

From above, we can see that the sales are on a decreasing trend, and the difference between
forecasted Y4 and Y3 is well within limits.

Q5) Recommendation on data entry

The model is built on time-series data with a 12-month moving average method. Hence the
model becomes more accurate with the inclusion of each month's latest data.

For example, including Jan data impacts the seasonal index of Jul, and it also incorporates the
latest seasonality & trend in the forecasted values, thereby increasing the model’s prediction
accuracy.
(Refer to supporting excel)

Page | 2
Q6) Calculations

The Formulae Used:

Seasonality Index = Average(Sales/Moving Average) for each month

Deseasonal Sales = Sales/Seasonality Index

Trend(i) = Intercept + (Co-efficient of Month Series)*(i) where i -> Time series

Predicted Sales = Seasonal Index * Trend

Conjoint Analysis
Q1) Design Choice
The main attributes considered in the design are Battery Life, Processing Speed, Price. There
are 3 levels for each of these attributes, and each level adds a qualitative sense to the attributes,
which is considered a good aspect.
The choice cards are designed using the Pairwise approach, which is considered advantageous
over the Full-profile approach. There are now only 9 profiles created using Pairwise, which
will guarantee customer’s attention while giving preference instead of 3*3*3=27 profiles
created using Full-profile. Also, the designed choice cards have minimal correlation among
themselves, indicating exclusivity of the profiles intended to capture all main effects of interest.
The OLS regression model built using this data has an R-Square value of 0.9344262, which is
close to 1, indicating that the goodness of the model is very high.
Q2) Utility & Importance

Feature Level No Description Utility Importance

Battery Life 3 10 days -0.222
2 7 days -0.556
1 5 days 0.7778 0.2857143
Processor Speed 3 8 GHz -0.556
2 5 GHz 0.1111
1 2 GHz 0.4444 0.2142857
Price 3 ₹ 25,000 0.1111
2 ₹ 15,000 1.1111 0.5
1 ₹ 10,000 -1.222
Utility Plots

Battery Life Processor Speed

1 0.5
0.5 0
0
-0.5 -0.5
-1 -1
10 days 7 days 5 days 8 GHz 5 GHz 2 GHz
Battery Life Processor Speed
Series1-0.222222222-0.5555555560.777777778 Series1 -0.5555555560.1111111110.444444444

Page | 3
 Price
2
1
0
-1
-2
Rs 25000 Rs 15000 Rs 10000
Price
Series1 0.111111111 1.111111111 -1.222222222

Q3)
Observations
Utility:
❖ The utility is highest for 5 days (0.7778) for the Battery Life attribute
❖ The utility is highest for 2 GHz (0.4444) for the Processor Speed attribute
❖ The utility is highest for ₹ 15,000 (1.1111) for the Price attribute
Order of Importance:
❖ Price
❖ Battery Life
❖ Processor Speed
Inferences
Based on the preferences given, the customer seems to be a lite user who needs low
specifications at a moderate price level. Since the price has the highest importance, we can say
that the customer is price sensitive (May not like to seen as a user of a cheap product as we
observed a strong negative correlation for ₹ 10,000 level).
Recommendation
The price of the new product must be close to ₹ 15,000, and the product specifications must
include 5 days Battery Life with a 2 GHz Processor Speed.
Q4) Next steps
❖ Collect preferences from multiple customers across demographics
❖ Reliability and Validity Analysis:
o Though the build model has an R-Square value of 0.9344262, we have to perform Test-
Retest or divide our estimation sample into many subsets and can repeat the analysis on
each of the samples to check reliability and validity
o Have to confirm whether the conjoin analysis and MDS (Multidimensional scaling) are
complementing each other or not
❖ Clustering consumers based on their utility functions will let us identify our target
market(based on market size & growth potential) and target specific customers for effective
positioning of our product

Page | 4