Price Forecasting with Deep Learning in Business to Consumer Markets

Abstract

Price forecasting is a challenging and essential problem studied in different markets. Many researchers and institutions, academically and professionally, develop future price forecasting techniques. This study proposes a data collection and processing pipeline to forecast the next day's price of a product in business to consumer (B2C) markets using the price data obtained from web crawlers, preprocessing steps, the deep features produced by the autoencoder, and the technical indicators. For this purpose, we use web crawlers to collect different airline companies' ticket prices daily and create a price index. We apply the discrete wavelet transform (DWT) preprocessing method to denoise the price index data, calculate some technical indicators analytically, and extract the deep features of the price data via three different autoencoders, linear, stacked linear, and long short term memory (LSTM). An LSTM forecaster generates forecasts using deep and calculated features. Finally, we measure the effects of autoencoder types, and mentioned features on the forecasting performance. Our study shows that using LSTM autoencoder on denoised time series price data with technical indicators in B2C markets yields promising results.