Parallel Processing Framework for Efficient Computation of Analyst Consensus Estimates and Measurement of Forecast Accuracy

Abstract

Forecasting of earnings is an integral component in the valuation of companies. Financial analysts provide such forecasts in the form of earnings estimates. Academic study has shown analyst forecasts to be more accurate than timeseries forecasts. Historically this has been based on a consensus forecast computed as the mean of analyst forecasts. In our research we consider alternative methods of aggregating consensus forecasts. We take inspiration from iterative filtering methods from Physics, as applied to other fields such as the aggregation of sensor readings and online reviews. In this paper we discuss the challenges of adapting iterative filtering algorithms to the aggregation of analyst earnings estimates. This encompasses modelling as well as technological challenges. We present our solution to the afore-mentioned challenges and develop a general framework for the systematic assessment of consensus aggregation algorithms. We show that a naïve implementation of this computation takes approximately 4 days to complete. Our framework performing the same computation takes a significantly reduced time of approximately 2 hours. We then apply this framework to the assessment of iterative filtering algorithms in the context of aggregating consensus earnings estimates. We present preliminary results of our study of the application of iterative filtering algorithms against a simple mean consensus.