Conducting, to our knowledge, the largest study ever of 5-min equity market returns using state-of-the-art machine learning models trained on the cross-section of lagged market index constituent returns, we show that regularized linear models and nonlinear tree-based models yield significant market return predictability. Ensemble models perform the best across time and their predictability translates into economically significant Sharpe ratios of 0.98 after transaction costs. These results provide strong evidence that intraday market returns are predictable during short time horizons, beyond what can be explained by transaction costs. Furthermore, we show that constituent returns hold significant predictive information that is not contained in market returns or in price trend and liquidity characteristics. Consistent with the hypothesis that predictability is driven by slow-moving trader capital, predictability decreased post-decimalization, and market returns are more predictable during the middle of the day, on days with high volatility or illiquidity, and in financial crisis periods.
This paper examines the steady state properties of the Threshold Vector Autoregressive model. Assuming that the trigger variable is exogenous and the regime process follows a Bernoulli distribution, necessary and sufficient conditions for the existence of stationary distribution are derived. A situation related to so-called “locally explosive models”, where the stationary distribution exists though the model is explosive in one regime, is analysed. Simulations show that locally explosive models can generate some of the key properties of financial and economic data. They also show that assessing the stationarity of threshold models based on simulations might well lead to wrong conclusions.
This paper proposes an innovative algorithm that significantly improves on the approximation of the optimal early exercise boundary obtained with simulation based methods for American option pricing. The method works by exploiting and leveraging the information in multiple cross-sectional regressions to the fullest by averaging the individually obtained estimates at each early exercise step, starting from just before maturity, in the backwards induction algorithm. With this method, less errors are accumulated, and as a result of this, the price estimate is essentially unbiased even for long maturity options. Numerical results demonstrate the improvements from our method and show that these are robust to the choice of simulation setup, the characteristics of the option, and the dimensionality of the problem. Finally, because our method naturally disassociates the estimation of the optimal early exercise boundary from the pricing of the option, significant efficiency gains can be obtained by using less simulated paths and repetitions to estimate the optimal early exercise boundary than with the regular method.