Forecasting Stock Values at a Tea Brokering Company for Reducing Risk in Granting Loans

Abstract

Asia Siyaka is a leading tea broker which has about 15% market share and annually sells tea with the worth of about 18 billion rupees. Tea factory owners and tea buyers are main clients of Asia Siyaka. Factory owners have to pay millions of rupees weekly for green leaves that are used for tea manufacturing. This money is borrowed from the broker as short term loans against pending tea stock. Asia Siyaka issues about 400 million rupees of loans weekly for this purpose. The broker keeps the pending tea stock as a security and grants these loans. The recovery of loan starts after selling the stock in the future auctions. Therefore the recovery of the loan is depending on the price which is achieved during the auction. The brokers’ main challenge is to estimate the stock which will be sold in the future auctions. The current system uses previous month’s each factory realized auction average price to estimate the each factory stock. This method is incorrect due to the variation of the grade mix of the production and the price variation of tea grades with time. To find the more accurate estimate for tea stock broker should do a detail tasting session. But it is not practical due to limited time. Therefore should discover a method to estimate the tea stock without tasting the tea. So the main objective of this project is to discover patterns, trends and rules that lie within the data and propose accurate stock estimation methodology for Asia Siyaka using data mining techniques. This project used initial descriptive statistics which helped to understand the data and the current system. Used time plots, stacked bar charts and cross tabulations to analyze the current system and research problem accurately. The project used more explorative approach because of the complexity of the problem and disclosed relationship of the data. Therefore selected cluster analysis which is an unsupervised learning technique. This project’s clustering is based on classical K-Means algorithm. The analysis shows that comparison should be done with the most recent data and also shows that factory, grade and package weight are the only visible attributes which represent the value of the tea. The proposed method is based on this result. The evaluation of the results shows this proposed method is more accurate than the current method.