Title: Looking forward, looking back: How machine learning predictions compare to expectations from option markets
Authors: Joachim Grammig - Eberhard Karls Universitaet Tuebingen (Germany) [presenting]
Constantin Hanenberg - University of Tuebingen (Germany)
Christian Schlag - Goethe University (Germany)
Jantje Soenksen - Eberhard Karls University Tuebingen (Germany)
Abstract: There are two opposing strategies to estimate the conditional expectation of a stock return, which is the optimal forecast using mean-squared error (MSE) loss. The first strategy is theory-based, parsimonious, and forward-looking. It utilizes the information contained in option prices, risk-free rates, and the price of the underlying to deliver conditional expected stock returns. The second strategy is theory-free, data science-driven, and backward-looking. It employs machine learning algorithms which are designed to find patterns in historical data to produce MSE-optimal return predictions. Although the two strategies are very different, they both pursue the same goal: Finding the best possible approximation of the conditional expected return. A level playing field is provided to assess the comparative advantages of the forward- and backward-looking approaches towards providing MSE-optimal return predictions. The analysis focuses on the S\&P 500 constituents with firm-level data ranging from 1972 to 2017 and volatility surface data ranging from 1996 to 2017. The results of the study suggest a tight head-to-head race of the two strategies since neither of them is able to substantially outperform the other.