Анализ предсказуемости направления движения финансовых рынков

It is now widely believed that stock returns exhibit some predictability - a number of researches aiming at examining the predictive power of various indicators such as ratios (like dividend-price or earnings-price ratios) or bond yields and making use of that predictive power for the purpose of efficient assets allocation and pricing. As such, forecasting future stock returns indisputably is of a huge practical importance. Let us consider a basic return decomposition, in which excess returns on a stock at time $t$ are decomposed into a product of their absolute value and their sign: $r_t = |r_t|\cdot \text{sign}(r_t)$ (Anatolyev, Gospodinov, 2010, p. 232). Majority of studies on market movement predictions focus on stock volatility due to its widely known property of clustering and hence high dependence and predictability. As such, the focus is mostly on the first term of our decomposition, the $|r_t|$ (absolute value of returns) term. On the contrary, there are much less studies of returns sign predictability (which is represented by the term $\text{sign}(r_t)$), however they can contain important correlation patterns and can be used in short-term predictions of market movements. It is useful, because understanding whether a profit or a loss is expected to be incurred in the next period is sometimes more important that understanding the exact expected mean of returns. My hypothesis is that stock returns possess a characteristic of out-of-sample directional predictability (that is, that conditional expectation of the direction of market movement is not generally equal to the unconditional expectation of that) meaning that there should exist some interdependence in the signs and quantiles of returns in time. Conceptually, my work is based on a paper by Malakhov D.I. (2019) on sign predictability of returns. %Another hypothesis is that there are existing correlation patterns in signs of different S\&P 500 stock returns. With modern computational methods becoming available, there is also an increasing number of literature dedicated to examining predictability of asset returns via various methods, which require extensive computations. For example, Feng, He and Polson examine predictability of returns via multiple-layers deep learning methods and conclude that there exist some non-linear dependence patterns. Dixon and Polson arrive at a similar conclusion that there is some existing predictability, which is non-linear and could be captured by deep learning techniques, which also accou Chinco, Clark-Joseph and Ye have found some predictability in one-minute returns using a Lasso model, which allowed them to identify some short-lived non-intuitive factors that make their model perform better. In my research, I am going to investigate predictability of return signs from two different perspectives. The first approach is to inspect the signs indirectly, by investigating distribution quantiles of the series of returns. The second approach is to model them directly, by constructing a logistic regression model on the signs encoded as a dummy variable. As such, I am aiming at researching and concluding whether the sign of stock returns can be predicted from past historical data of different frequency returns of that stock.