Summer Internship at UST software India Pvt. Ltd

About Company

UST India is a highly regarded Information Technology organization founded in Delhi. Since then UST India has been providing a full range of innovative and cost-effective IT services, and solutions that focus on satisfying client needs worldwide.

UST India is a subsidiary of Universal System Technologies, Inc. (USA) with its headquarters in Michigan USA.

UST India takes immense pride in its growth and successful performance to date, which can be directly attributed to our customer’s satisfaction and confidence in our services.

The company’s goal is to be a world class global provider of Information Technology.

Introduction

March of 2020 is the month with COVID 19 drastic spread all over the countries and WHO declared it as a global Pandemic. As the spread increased many countries implemented lockdown and the exports, imports, manufacturing were halted for a certain time and so the global economic crisis took place.

So we trained a machine learning model that could predict the GDP growth of India so that the government could invest in the right sectors and take smart and optimized decisions to overcome the economic crisis. This project aims on building a predictive model of the GDP growth rate of India after this pandemic COVID-19 and to suggest the optimal scenario for stimulating the growth rate.

We will be working with datasets on Agriculture, Corporate performance, Public expenditure, GST revenue, and Index of Industrial Production (IIP). We will be Implementing MLPRegressor and ARIMA time series model for GDP Prediction based on the previous year’s GDP growth. We also used different regression algorithm like Linear regression based on the accuracy.

Analysis

1.1 Regression on Neural Network (manually configured)

We implemented regression on a neural network which is different on the output layer compared to normal MLP’s because the output is not limited to 1 or 7 integers, we need a series of output so we to solve this issue we must remove the activation function of the last layer or set it to linear. We tried with different loss functions, optimizers for better results. In the end, the loss is measured with mean-squared logarithmic error, and optimizer is set to adam and metrics as a mean absolute percentage error for the best outcomes.

Fig 1 : Plotting was done with black dots as test data and blue dots as predicted data

1.2 Regression on Neural Network (using inbuilt function MLPREGRESSOR)

We implemented the MLP regression using the predefined function of sklearn with solver set to lbfgs, alpha set to 0.001, and epochs to 1000.

Fig 2 : Plotting was done with black dots as test data and blue dots as predicted data.

1.3 Linear Regression

Then we implemented linear regression with sklearn package with fit_intercept set to false and other parameters are left as default for best results.

Fig 3 : Plotting was done with black dots as test data and blue dots as predicted data

Learning Outcome

Fig 4 : represents the results of manually configured multi-layer perceptron regressor having mean absolute percentage error as 19.3342

Fig 5: MLPregressor outputs an accuracy of 85.44%

Fig 6: Linear Regression outputs an accuracy of 89.67%

Conclusion

We have finally trained a machine learning model that could predict the GDP growth of India so that the government could invest in the right sectors and take smart and optimized decisions to overcome the economic crisis. We have used a linear regression model as it is the best set model for prediction. We have also used the ARIMA time series model for better future prediction of GDP growth. We achieved an accuracy of 90%. So, we have completed a model for prediction of GDP growth rate of India after this pandemic COVID-19 and suggest the optimal scenario for stimulating the growth rate.

Challenges

There are so many models for predicting the GDP growth rate of India like LSTM, Random forest, etc, so we have to select the model which gives better prediction among all these models. We also need to learn the ARIMA time series model as it helps in the better future prediction of GDP. We also used Multi Linear Perceptron Regression so that we could calculate GDP using these values of Agriculture, Manufacturing, industry(including construction), Services, Imports of goods, and services and Exports of goods and services with respect to the year after the pandemic.

I would like to thank our mentor Dr A.K. Sinha for providing us with this oppurtunity and also my teammates Madhu Kiran and Avinash Boddeda for forming an excellent team.