Prediction of peanut butter prices in the United States by tracking concept drift
Choi, Ho Suk
MetadataShow full item record
The objective of thesis is an application of model combining ensemble methods to peanut butter prices data from 2001 to 2012 in the United States. In the data, there exist gradual and sudden price changes over time. This phenomenon is known as concept drift, which is a challenge to build a predictive model in data streams, because a predictive model built on the past data might not be valid with the new data after the change. To find the best predictive model, we compare the performances of model combining ensemble methods, one-step ahead forecasting and a regression model controlling endogeneity in prices. From the comparison, we find that the model combining via ridge regression with constraints of sum-to-1 and non-negativity predicts the peanut butter prices most accurately and adapts to the concept drift quickly.