Forecasting of sales by using fusion of machine learning techniques

Abstract

Forecasting is an integral part of any organization for their decision-making process so that they can predict their targets and modify their strategy in order to improve their sales or productivity in the coming future. This paper evaluates and compares various machine learning models, namely, ARIMA, Auto Regressive Neural Network(ARNN), XGBoost, SVM, Hy-brid Models like Hybrid ARIMA-ARNN, Hybrid ARIMA-XGBoost, Hybrid ARIMA-SVM and STL Decomposition (using ARIMA, Snaive, XGBoost) to forecast sales of a drug store company called Rossmann. Training data set contains past sales and supplemental information about drug stores. Accuracy of these models is measured by metrics such as MAE and RMSE. Initially, linear model such as ARIMA has been applied to forecast sales. ARIMA was not able to capture nonlinear patterns precisely, hence nonlinear models such as Neural Network, XGBoost and SVM were used. Nonlinear models performed better than ARIMA and gave low RMSE. Then, to further optimize the performance, composite models were designed using hybrid technique and decomposition technique. Hybrid ARIMA-ARNN, Hybrid ARIMA-XGBoost, Hybrid ARIMA-SVM were used and all of them performed better than their respective individual models. Then, the composite model was designed using STL Decomposition where the decomposed components namely seasonal, trend and remainder components were forecasted by Snaive, ARIMA and XGBoost. STL gave better results than individual and hybrid models. This paper evaluates and analyzes why composite models give better results than an individual model and state that decomposition technique is better than the hybrid technique for this application.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 57 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.