Sales forecasting using walmart dataset

Instruction Will Be Attached ( Walmart Instruction Is An Example Of How Your Work Should Be)

Allstate Claim Prediction Challenge (AllState2)

A key part of insurance is charging each customer the appropriate price for the risk they represent.

Risk varies widely from customer to customer, and a deep understanding of different risk factors helps predict the likelihood and cost of insurance claims. The goal of this competition is to better predict Bodily Injury Liability Insurance claim payments based on the characteristics of the insured customer’s vehicle.

Many factors contribute to the frequency and severity of car accidents including how, where and under what conditions people drive, as well as what they are driving.

Bodily Injury Liability Insurance covers other people’s bodily injury or death for which the insured is responsible. The goal of this competition is to predict Bodily Injury Liability Insurance claim payments based on the characteristics of the insured’s vehicle.

Files

Train.cvs

Test.cvs

Data Description

Each row contains one year’s worth information for insured vehicles. Since the goal of this competition is to improve the ability to use vehicle characteristics to accurately predict insurance claim payments, the response variable (dollar amount of claims experienced for that vehicle in that year) has been adjusted to control for known non-vehicle effects. Some non-vehicle characteristics (labeled as such in the data dictionary) are included in the set of independent variables. It is expected that no “main effects” corresponding will be found for these non-vehicle variables, but there may be interesting interactions with the vehicle variables.

Calendar_Year is the year that the vehicle was insured. Household_ID is a household identification number that allows year-to-year tracking of each household. Since a customer may insure multiple vehicles in one household, there may be multiple vehicles associated with each household identification number. "Vehicle" identifies these vehicles (but the same "Vehicle" number may not apply to the same vehicle from year to year). You also have the vehicle’s model year and a coded form of make (manufacturer), model, and submodel. The remaining columns contain miscellaneous vehicle characteristics, as well as other characteristics associated with the insurance policy. See the "data dictionary" (data_dictionary.txt) for additional information.

Our dataset naturally contained some missing values. Records containing missing values have been removed from the test data set but not from the training dataset. You can make use of the records with missing values, or completely ignore them if you wish. They are coded as "?".

There are two datasets to download: training data and test data. You will use the training dataset to build your model, and will submit predictions for the test dataset. The training data has information from 2005-2007, while the test data has information from 2008 and 2009. Submissions should consist of a CSV file. Records from 2008 will be used to score the leaderboard, and records from 2009 will be used to determine the final winner.

Missing feature values have been kept as is, so that the competing teams can really use the maximum data available, implementing a strategy to fill the gaps if desired. Note that some variables may be categorical (e.g. f776 and f777).

The competition sponsor has worked to remove time-dimensionality from the data. However, the observations are still listed in order from old to new in the training set. In the test set they are in random order.

Walmart Sales Prediction Using Rapidminer Prepared by : Nagarjun Singharavelu

I. Introduction:

Wal-Mart Stores, Inc is an American Multinational retail corporation that

operates a chain of discount department stores and Warehouse Stores. Headquartered in

Bentonville, Arkansas, United States, the company was founded by Sam Walton in 1962 and

incorporated on October 31, 1969. It has over 11,000 stores in 27 countries, under a total 71

banners. Walmart is the world's largest company by revenue, according to the Fortune Global

500 list in 2014, as well as the biggest private employer in the world with 2.2 million employees.

Walmart is a family-owned business, as the company is controlled by the Walton family. Sam

Walton's heirs own over 50 percent of Walmart through their holding company, Walton

Enterprises, and through their individual holdings. The company was listed on the New York

Stock Exchange in 1972. In the late 1980s and early 1990s, the company rose from a regional to

a national giant. By 1988, Walmart was the most profitable retailer in the U.S. Walmart helps

individuals round the world economize and live better.

The main aim of our project is to identify the impact on sales throughout

numerous strategic selections taken by the corporate. The analysis is performed on historical

sales data across 45 Walmart stores located in different regions. The foremost necessary is

Walmart runs many promotional markdown events throughout the year and we have to check

the impact it creates on sales during that particular period. The markdowns precede prominent

holidays, the four largest of which are the Labor Day, Thanksgiving and Christmas. During these

weeks it is noted that there is a tremendous amount of change in the day-to-day sales. Hence

we tend to apply different algorithms which we learnt in class over this dataset to identify the

effect of markdowns on these holiday weeks.

II. Information about dataset:

We had taken four different datasets of Walmart from Kaggle.com

containing the information about the stores, departments, average temperature in that

particular region, CPI, day of the week, sales and mainly indicating if that week was a

holiday. Let us explain each dataset in detail.

Stores:

 The no. of attributes in this dataset is 3.  

 They are store number, type of store and the size of store.  

 Output attribute is the size of store.  

 There are 45 stores whose information is collected.  

 Stores are categorized into three such as A, B and C, which we assume it to be

superstores containing different types of products.  

 The store size would be calculated by the no. of products available in the particular

store ranging from 34,000 to 210,000. 

Train:

 This is the historical training data, which covers to 2010-02-05 to 2012-11-01.  

 It consists of the store and department number.  

 Date of the week.  

 Weekly sales in USD for the given department in the given store.  

 Also data denoting if the week is a special holiday week by displaying true or false.  

 The department no. ranges from 1-99.  

 The dataset consists of many records that are blank. Hence they were not considered

for analysis.  

 The output attribute is sales and the data denoting special holiday week. 

Test:

 This dataset is similar to train data, except that the sales column has been eliminated.  

 It is used to predict the sales for each triplet of store and department.  Features:

 This file contains regional activity for the given dates such as average temperature in

that region which is measured which ranges from -7 to 100 F.  

 Cost of fuel in that particular region is available in the dataset from which we can

identify that if the fuel is higher in that particular region, whether it affects the sales as

customer use automobiles to reach the store.  

 Customer price index (CPI) which is used to measure the changes in price level in that

particular store.  

 Un-employment rate in that particular region.  

 Data denoting if that particular week is a special holiday week.  

 5 Markdowns are available in this dataset which denotes the promotional markdowns

which Walmart is running. There are 5 markdowns in that particular duration in the

dataset.  

 The output attribute is Markdowns and data denoting if it’s a holiday week. 

Among these 5 files, Features and train files contain every necessary elements to do the predication. So we are going to take advantage of them.

III. Data transformation & Preprocessing:

Our aim is to find Walmart’s weekly sales. We have all the necessary data for

predicting the weekly sales in Feature data. But the previous weekly sales report is in Train data

set. So we decided to combine all the data sets into a single data set. After combining, we now

have the sales data of 45 stores with up to 99 departments which consists of totally 421,000

records. Now our final data set has the following attributes.

 Store   Date 

 Temperature   Fuel_Price   MarkDown1-5   CPI   Unemployment   IsHoliday   Weekly_Sales   Dept   Type 

Now we have 13 attributes, in which attributes such as department number (Dept), Date and Type are not required for our prediction, since it doesn’t have any impact on its weekly sales. Hence we are removing these attributes from our final data set. After removing all the unwanted

data and attributes in our dataset we now have 6,435 records on total.With our huge dataset, it is

difficult to get exact result about future sales from models because of the uncertainty in so many

disturbing elements. So we decided to group the data and then implementing with classifiers or

regression models. By using Discretize by Size operators, we grouped the data based on

Weekly_Sales into 5 categories and then exported the output as CSV file to desktop.

After processing, now check the data of weekly sales, we found that the ranges are as

follows; range1 [-? - 9144.495] range2 [9144.495 - 14213.040]

range3 [14213.040 - 22267.010]

range4 [22267.010 - 39499.755]

range5 [39499.755 - ?]

Now we manually label the whole 6,435 records into 5 groups each contain 1,287 records.

Groups Mark as Description

Group 1 A range1 [-? - 9144.495]

Group 2 B range2 [9144.495 - 14213.040]

Group 3 C range3 [14213.040 - 22267.010]

Group 4 D range4 [22267.010 - 39499.755]

Group 5 E range5 [39499.755 - ?]

IV. Mining models & Performance:

Our purpose is to build a model to predict Walmart’s weekly sales in future, so we would

like to select the most suitable model for predication. Look into the table, we find out the data is

structured. However, the parameters are complex and do not show its rules obviously. Even

more, we have no idea about whether it is linear or not. According to problems above, we determined to use Neural Network Model with our project

because it has the characteristics below:

•It is for complicated prediction problems.

•Visualization or understanding of the rules are not needed.

•Accuracy is very important.

The design for implementing the Neural Network model in rapid miner is as follows:

The output we obtained is as follows,

Accuracy = 77.61%, when Learning Rate = 2000 & Training Cycles = 0.05

true A

true B

true C

true D

true E class

precision

pred. A 939 49 83 36 95 78.12%

pred. B 17 1114 58 75 101 83.61%

pred. C 160 102 923 148 29 68.77%

pred. D 60 95 93 780 187 65.20%

pred. E 23 276 43 161 788 63.04%

67.67%

class recall 78.32% 68.09% 77.92% 65.00%

We tried to test our result by changing learning rate and training cycle values and noted down

the changes too. The result obtained in this process is as follows;

Accuracy = 72.69%, when Learning Rate = 1000 & Training Cycles = 0.05

true A

true B

true C

true D

true E class

precision

pred. A 1012 62 96 42 102 77.62%

pred. B 23 1193 58 76 120 84.61%

pred. C

184

114

1025

181

28

68.91%

pred. D 69 124 104 842 209 64.89%

pred. E 34 306 37 179 861 61.89%

class recall 76.72% 69.31% 78.65% 63.79% 67.23%

Accuracy = 75.35%, when Learning Rate = 1000 & Training Cycles = 0.3

class

true A

true B

true C

true D

true E

precision

pred. A 908 86 114 56 108 74.38%

pred. B 25 1033 50 64 83 86.31%

pred. C 164 123 895 176 40 64.02%

pred. D 73 117 94 732 189 62.75%

pred. E 29 277 47 172 780 63.07%

class recall 77.03% 65.24% 77.28% 64.42% 67.02%

Since weekly sales is the core for any industry. So it is necessary to get the more accurate

results, since it have direct impact on companies revenue. To achieve better performance and

accuracy results, we have now increased the hidden layer size to 10.

We now combined all the hidden layer results obtained from each node in a single table to check where the weekly sales have a greater impact.

Attributes Node 1 Node 2

Store -30.224 27.949

Temperature -0.192 0.68

Fuel_Price 1.054 1.459

MarkDown1 0.114 -2.679

MarkDown2 0.88 0.476

MarkDown3 9.209 -0.635

Node 3 Node 4 Node 5 Node 6 Node 7 Node 8 Node 9 Node 10

-20.9 -31.401 5.072 -17.01 40.862 27.088 18.266 16.821

-0.965 -0.601 0.038 -0.847 -0.436 -1.513 -1.094 0.126

-0.46 0.632 -1.242 -1.713 0.698 1.303 2.589 -0.007

-0.24 -0.636 -4.109 -1.322 1.004 0.648 -1.147 -2.402

-0.03 2.65 -0.387 -1.306 0.328 0.633 1.983 -0.821

0.631 7.522 -6.247 -2.194 -6.743 -0.075 1.157 6.285

MarkDown4 -0.613 -0.554 0.131 1.257 -2.967 -0.445 0.753 0.691 -0.115 0.599

MarkDown5 0.41 -3.728 -3.697 1.786 -14.643 -2.111 0.341 -0.999 -7.048 -0.908

CPI -6.316 29.004 -12.179 -12.975 0.568 11.217 -27.856 -15.442 -5.764 1.045

Unemployment -12.906 -17.876 4.064 -1.446 8.895 3.908 7.056 12.174 -31.068 1.142

IsHoliday 0.056 0 -0.191 -0.484 -0.068 -0.244 -0.151 -0.27 -0.036 -0.025

Bias -13.162 -10.766 -9.052 -8.572 -25.596 9.488 -10.875 20.103 -2.328 -9.733

The hidden layer network for this dataset will look like:

Other Models:

We tried to analyze data with more models other than Neural Network. But we find out that

Neural Network model is the most suitable one for Walmart dataset. Other models cannot

work on this example. Naïve Bayes: Design:

Output:

Accuracy = 23.42%

true A true B true C true D true D

class

precision

pred. A 8190 0 0 0 0 100.00%

pred. B 19795 23905 23910 23905 23909 20.71%

pred. C 15895 19128 19124 19128 19126 20.70%

pred. D 0 0 0 0 0 0.00%

pred. E 3917 4782 4781 4782 4780 20.74%

class

recall 17.13% 49.99% 40.00% 0.00% 10.00%

K-NN: Design:

Output:

When K = 1, Accuracy = 28.76%

true A true B true C true D true E

class

precision

pred. A 385 207 373 180 199 28.65%

pred. B 190 619 259 311 301 36.85%

pred. C 352 271 187 225 188 15.29%

pred. D 113 267 124 180 210 20.13%

pred. E 159 272 257 304 302 23.34%

class 32.11% 37.84% 15.58% 15.00% 25.17%

recall

When K = 10, Accuracy = 31.06%

true A true B true C true D true E

class

precision

pred. A 429 234 403 168 206 29.79%

pred. B 272 834 288 421 428 37.18%

pred. C 219 123 186 97 213 24.20%

pred. D 195 266 162 347 281 28.64%

pred. E 84 179 161 167 72 10.86%

class 35.78% 50.98% 15.50% 28.92% 6.00%

recall

Decision Tree:

No matter what the parameters are, the Accuracy is 25.42% and never change.

true D true S true C true B true A

class

precision

pred. D 0 0 0 0 0 0.00%

pred. S 1199 1636 1200 1200 1200 25.42%

pred. C 0 0 0 0 0 0.00%

pred. B 0 0 0 0 0 0.00%

pred. A 0 0 0 0 0 0.00%

class 0.00% 100.00% 0.00% 0.00% 0.00%

recall

As we show above, Naïve Bayes Model products the lowest Accuracy

(18.63%). Abstractly, the Naïve Bayes model is a conditional model like this:

P (Sales | Store, Data, Temperature … & IsHoliday)

=

P(Sales)*P(Store, Data, Temperature, … & IsHoliday | Sales)

P(Store, Data, Temperature, … & IsHoliday)

Because variable Store, Data to IsHoliday are independent on each other, so:

P (Store, Date, Temperature … & IsHoliday) =P (Store)*P (Date)*…..*P (IsHoliday)

Due to so many numbers in columns Store, Date … IsHoliday that do not repeat, the

probability of each Variables is too small. So P (Store)*P (Date)*…..*P (IsHoliday) will be far

lower than 1/6435. This means the probability of sales basing such a model is infeasible.

V. Interpretation:

As a result, we find out that how each parameter affects the weekly sales. MarkDown 1 to 5 has the highest weight as 9 to -16 which mean it really makes an enormous impact on the sales. Promotion will increase weekly sales remarkably. Fuel price and temperature also makes a positive impact, higher price and temperature makes higher sales. CPI and Unemployment rate having a heavy negative impact on the prospects of sales. The higher CPI and unemployment rate, the less weekly sales. Holidays affect weekly sales slightly. We think customers don’t care whether today is holiday or not, the only reason they buy items is promotion. Therefore, Promotion drives sales and revenue to increase. VI. Conclusion:

The main objective was to model the effects of markdown events throughout the year and it

was successfully measured using various algorithms. From the analysis it is noted that MarkDown 1

to 5 has the highest weight as 16, which means it really makes an enormous impact on the sales. Promotion will increase weekly sales remarkably. Fuel price and temperature also makes a positive impact, higher price makes higher sales. CPI and Unemployment rate has a heavy negative impact on the prospects of sales. The higher CPI and unemployment rate, the less is the weekly sales. Holidays affect weekly sales slightly. We felt customers didn’t care whether it was a holiday or not because the only reason they buy items is due to promotion. It is notable that they are attracted to various deals in departmental stores all over the world.

References: 1. https://www.kaggle.com/c/walmart-recruiting-store-sales-forecasting/data 2. http://www.walmart.com/ 3. http://en.wikipedia.org/wiki/Walmart

https://www.kaggle.com/c/walmart-recruiting-store-sales-forecasting/data
http://www.walmart.com/
http://en.wikipedia.org/wiki/Walmart