Trading strategies based on estimates of conditional distribution of stock returns

• Thesis title in Czech: Trading strategies based on estimates of conditional distribution of stock returns
• Thesis title in English: Trading strategies based on estimates of conditional distribution of stock returns
• Key words: kvantilové regrese, obchodní strategie, odhady podmíněné distribuční funkce, GARCH
• English key words: quantile regression, trading strategy, conditional distribution function estimation, GARCH
• Academic year of topic announcement: 2016/2017
• Thesis type: diploma thesis
• Thesis language: angličtina
• Department: Institute of Economic Studies (23-IES)
• Supervisor: doc. PhDr. Jozef Baruník, Ph.D.
• Author: hidden - assigned by the advisor
• Date of registration: 27.04.2017
• Date of assignment: 27.04.2017
• Date and time of defence: 19.09.2018 08:30
• Venue of defence: Opletalova - Opletalova 26, O206, Opletalova - místn. č. 206
• Date of electronic submission: 30.07.2018
• Date of proceeded defence: 19.09.2018
• Opponents: prof. Ing. Miloslav Vošvrda, CSc.

References

Foresi, Silverio, and Franco Peracchi. The conditional distribution of excess returns: An empirical analysis. Journal of the American Statistical Association 90, no. 430 (1995): 451-466.
Friedman, Jerome, Trevor Hastie, and Robert Tibshirani. Additive logistic regression: a statistical view of boosting (with discussion and a rejoinder by the authors). The annals of statistics 28, no. 2 (2000): 337-407.
Hastie, Trevor J., and Robert J. Tibshirani. Generalized additive models. Vol. 43. CRC press, 1990.
Friedman, Jerome, Trevor Hastie, and Robert Tibshirani. The elements of statistical learning. Vol. 1. Springer, Berlin: Springer series in statistics, 2001.
Gallant, A. Ronald, Peter E. Rossi, and George Tauchen. Stock prices and volume. Review of Financial studies 5, no. 2 (1992): 199-242.
Fama, Eugene F., and G. William Schwert. Asset returns and inflation. Journal of financial economics 5, no. 2 (1977): 115-146.
JW Yang, J Parwada. Predicting stock price movements: an ordered probit analysis on the Australian Securities Exchange. Taylor & Francis Journals, vol. 12, no. 5 (2012): 791-804.
Thomas Q. Pedersen. Predictable Return Distributions. Journal of Forecasting 34, no. 2 (2015): 114-132.

Preliminary scope of work in English

Motivation:
In order to effectively invest or trade on financial markets it is needed to have a well-defined model and to stick to it while making investment decisions without an emotional impact. The model needs to be sufficiently simple but also include relevant available data in order to provide good results when it comes to real investment.
The majority of trading models in current literature estimate the expected stock price and its variance as a measure of risk. It is completely correct as long as the returns are normally distributed which is, however, one of the restrictive assumptions which are hardly met in reality. The distributions of returns have often higher kurtosis and heavy tails causing increase of probability of extreme losses. That was the reason why researchers and also practitioners diverted from the classical OLS model and developed more advanced techniques. They relaxed possibly all assumptions and created nonparametric models which take into account only the information contained in the data.
The best solution would be of course to know the future excess return but as this will most probably be never possible the next best solution is to know the distribution of returns. The knowledge of the distribution helps the investor assess the risk and behave accordingly. The goal of our thesis is first to predict conditional distribution function of excess returns based on various macroeconomic data, company accounting numbers and basic trading measures and second, based on the predictions, to develop a trading strategy which could systematically outperform the market returns – e.g. buy and hold strategy.

Hypotheses:
1. Hypothesis #1: The new trading strategy, based on estimates of future distribution of excess returns, systematically outperforms the market in very long run.
2. Hypothesis #2: The newly proposed trading strategy provides statistically better results than usual trading methods.
3. Hypothesis #3: Self-learning trading model provides better results than empirical or conceptual model.

Methodology:
The model used will be similar to the one used in Foresi and Peracchi (1995), particularly the logistic additive regression model. We will use a logit additive regression where instead of particular parameters s smooth functions will be estimated. For this approach kernel or spline regression can serve as a good tool. Instead of for expected value this regression will be repeated for sufficiently large number of quantiles. Each regression will estimate the probability of future excess return being below or above particular, previously defined, value. All these regressions together provide an estimate of distribution function of excess returns conditional on preselected variables. Based on the estimate of conditional distribution function of excess returns I will find out whether it is meaningful to use it for a development of trading strategy and whether this strategy outperforms at least buy and hold strategy. As other studies have shown (Foresi and Peracchi 1995; Thomas Q. Pedersen 2015) different quantiles of distribution are affected unequally by selected predictors thus opening space for more accurate trading strategy.

The analysis will be conducted mainly on daily returns of quoted companies and it will be also accounted for transaction cost associated with purchasing and selling of stocks.

As a possible extension we may try to create a model which is capable of learning itself, model which incorporates its past mistakes and improves the future estimates. For this purpose we try to incorporate existing learning algorithms.

Expected Contribution:
We expect to develop new trading strategy based on distribution estimation and possibly learning algorithms. The goal is to perform out-of-sample tests and conclude whether it is first possible to estimate future distribution function of returns to a level sufficient for outperforming other trading strategies and second create self-learning algorithm useful for everyday investing.