Overview / Usage

The data-set that is used for analysis is approximately 2GB and contains the following details:

1. Customer details

2. Store location / Store information

3. Purchase information

4. Billing information

5. Discount/Loyalty Details

6. Transaction information

The data-set that is used for forecasting sales is the sales and transaction data of a big retail store from India and is approximately 2GB. As per analysis, the data-set consists of not more than 7.3 million customers from 7 different branches of the store. The data available is for three years which starts from January 2015 to December 2017. The number of products that has been sold is all seven branches would count to 6000 distinct products of various categories which include grocery, beauty products and so on. As far sales prediction is considered, proper selection of features plays a major role. The input variables here include the lagged sales values for past time units, transaction date, product type and promotions on particular product and the output variable would be the sales related feature (sales of products on daily basis).

​

​

Methodology / Approach

Deep learning models comes into consideration because it out performs the machine learning and statistical model when the data size is large. As far prediction is done, the amount of data that we take is pretty large. But both the models has its pro and cons. Deep learning models takes time for training which in this case is at least 3.5 days but the time for prediction is much less when compared with linear models. The hardware that is used for training data has 4 GB RAM, NVIDIA GeForce 830M GPU and 1 TB hard drive.

CNN, LSTM and Attention(Cognition) based Deep Learning models are used.

​

Technologies Used

A sale of a product is the major key for any business and improving the sales is the goal for every business person for which the person keeps developing new strategies. Sales prediction is one such strategy which helps them make better financial decision. The data used has features affecting sales for three years which includes the holiday details, number of products sold on daily basis and number of products that had promotion on particular day. Having these features and sales value based on the feature, an attention based fused model is presented which includes an attention mechanism and a simple feed forward network. The model performs better on any sequential data and here the performance of the model is compared with commonly used models for forecasting which includes ARIMA model and various variants of LSTM model. The comparison is based on metrics such as mean squared error. The model trained faster when compared to other models and also the training and testing error is less when compared to other models. This model outperforms any kind ensemble models and provides a better prediction.

Other links

• Video