Deploying Gaussian Processes to Predict Avocado Price Fluctuations

Time series data often pose difficult, non-linear modeling problems and predictions tend to suffer from the inability to quantify uncertainty. We investigate alternatives to the traditional ARIMA methods of time series analysis by considering a Bayesian Gaussian process with a composite kernel function as the prior. The model fitted on historic avocado prices from 2015 to 2018 contained 94.7% of the sample/observed data and 94.1% of the out-of-sample/unobserved true data, despite having a fairly minimal feature space. This paper attempts to establish that for price prediction problems, Gaussian processes are an effective and interpretable model choice.