Data Scientist/ Big Data Analyst/ Machine Learning Engineer

I worked with the solutions team of Group Data and Analytics of ABMCPL, and I was supposed to work on the demand forecasting and production planning using optimization techniques of Pulp and Fiber business of Aditya Birla Grasim industry.

DEMAND FORECASTING

• Worked on machine learning and time series techniques to get the 3 months ahead demand of viscose fiber and interacted with different stakeholders and business people to understand nuances and real time market change to implement on demand forecasting to get best accuracies.
• Pulp and Fiber Industry produces different fibers monthly based on the demand which we need to be forecasted so that they can avoid over production and under production. Different fibers have varying demand on market fluctuated due to different reasons that we have to consider in the model using different external variables and need to get the accuracies above 90%.
• We used different models like univariate and multivariate time series models and different data pre-processing techniques to reduce mean absolute error and thus improving accuracy.
• Instead of considering one type model we ensemble the different models to get single output based on weightage by minimizing the error. These reduces the fluctuations in monthly accuracy and keep it optimal.

OPTIMIZATION

• Why do we need optimization?
  Here in pulp and fiber industry customers have preference for fibers from plant and machines. They will pay more if we satisfy their preference, that is their willingness to pay depends on different factors so we need to optimize it to get the best combination.
• Based on the forecast and sales data, I worked on optimization project to plan the production plan for the next month by maximizing the profit of the business.
• We create a delivery plan monthly based on the forecast/sales data on customer level with from where they can take the products required. For this we use Forecast/Sales with pending demand as the demand, production & starting inventory as supply and revenue & costs as Margin
• Since the demand is always higher, we aim to optimize our production plan by maximizing the profit. That is, we consider total revenue, production cost and delivery cost to find best combination of demand satisfaction by maximizing the margin. Also separately worked on customer preference demand satisfaction.
• While working on the shade changeover was a great issue happened on plants. High changeover rate can impact low quality products with high manual effort for this I worked on an analysis where I am able to reduce the shade changeovers and hence it produced huge profit.

COMMODITY PRICE FORECASTING

• During my internship in ABMCPL GDNA, I was part of a team which deals with the Aluminum Price Forecasting for the HINDALCO business.
• My aim was to enhance existing tool kit for Risk Management by adding a machine Learning based Price Projection for Hedge Decision making process.
• This will complement existing tools of Fundamental & Technical (Charting) analysis.
• Hindalco have a share of Aluminum from LME exchange
• So based on our forecast they are making the decision when to sell and how long to hold the share that is basically the hedging process
• For this purpose, I have gone through different data preprocessing and model fitting to predict 1 month and 1 day ahead Aluminum price.
• I have tried RF feature selection and PCA for choosing best combinations of features for modelling and GridSearchCV & RandomSearchCV for hyper parameter tuning
• To perform multivariate analysis, I was supposed to add more features like technical indicators for modelling.
• I used univariate and multivariate time series models to predict the price and I acquired more than 90% accuracy from the models.
• Instead of using single model, ensembling different models using weightages gave me better results.