GPU Implementation of Sales Forecasting with Linear Regression

Abstract

Forecasting of sales is very important in  any business as it helps managers to learn from  historical data and make informed decisions. This  generally involves intensive processes using  spreadsheets which require inputs from all levels in an  organization. This approach introduces bias and is  generally not accurate. There are several methods that  have been used in the past to forecast sales, such as  Exponential Smoothing, Moving Average, and  Autoregressive Moving Average (ARMA). Due to the  nature of the data, it usually takes more time for these  methods to analyze the sales data and make predictions.  In this paper, the sales data is analyzed and predictions  are made by using linear regression as implemented on  the GPU to make the process faster. Sales forecasting is  made possible by finding best fit line by linear  regression techniques (i.e. linear convolution). To  illustrate this process, simulated sales data was used.  The sales forecasting with linear regression  implementation using GPU was compared to the CPU  implementation and a speedup of up to 7.557x was  achieved.