Palestrante
Descrição
This project is focused on using famous neural networks architectures in the forecast of stock prices, but the main twist of this idea is in the creation of a purely generated dataset, aimed in investigating it's effectiveness when applied in real stock prices of assets located in BOVESPA. To generate this simulated data, I simulated coupled stochastic processes that are commonly used in this field (1), that follows this type of differential equation:
\begin{equation}
dx = \mu dt + \sigma dW_1\
d\sigma = \alpha(x,\sigma)dt + \beta(x,\sigma)dW_2
\end{equation}
Where $x$ is the return and $\sigma$ is the volatility of an asset (which is also considered as a stochastic variable in this model). My goal is to create a dataset large enough, using those equations, to make those neural network architectures "learn" how to forecast the stock prices in short-term predictions. In the creation of the neural network, I am focused in using famous architectures that tackles time-series purposes, such as recurrent neural networks, such as the Long short-term memory (LSTM), which are hugely used in problems where past inputs have direct effect on the current input of the system. I am also interested in developing a transforming network, which is a recent network that became famous because of it's generation of long outputs that seems to understand the patterns of it's input. This network already showed amazing performance in generating very comprehensible texts (2) and songs. (3) After the development of a well structured network that can predict stock prices with a decent precision, I will aim in using this model on real market data to improve it's performance. To fulfill this goal, I will search for transfer learning techniques that are commonly used to improve a neural network that were trained to work in other tasks.
Referências
1 HESTON, S. L. A closed-form solution for options with stochastic volatility with applications to bond and currency options. Review of Financial Studies, v. 6, n. 2, p. 327-343, 1993.
2 RADFORD, A. et al. (2019). Language models are unsupervised multitask learners. Available from: https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf. Acessible at: 25 July 2019.
3 PAYNE, C. MuseNet. OpenAI, 25 Apr. 2019. Available from: https://openai.com/blog/musenet. Acessible at: 25 July 2019.
| Apresentação do trabalho acadêmico para o público geral | Sim |
|---|---|
| Subárea | Termodinâmica, Fluidodinâmica e Estatística |
