A0740
Title: Boosting time-series prediction performance for inflation indicators
Authors: Jeffrey Bohn - UC Berkeley (United States) [presenting]
Abstract: As a risk, trading, strategy, and decision-support systems have become more deeply integrated into financial services firms' workflows, predicting a collection of economic and market indicators becomes even more critical to support these systems than in the past. At the same time, the underlying processes that drive economies and markets have become increasingly dynamic, given they are more likely to be subject to rapid successions of regime changes. Conventional curve-fitting frameworks that assume linear/log-linear, stable relationships continue to exhibit degraded predictive performance. Fortunately, innovations are being found in machine-learning algorithmic frameworks that lead to a collection of promising techniques defined as boosted trees or gradient boosting. These boosting methods better capture underlying non-linear data relationships in ways that can materially improve predictive performance for economic and market indicators. Some of the newer boosting methods and report compelling results will be described for predicting inflation with a subset of these methods. An approach called entropy boosting will be introduced.