Granger causality is a phenomenon where there exists a causal relationship between
two or more different time series data. Complex data sets involving stock market
has higher probability of such problems since the stock price of one company
can be highly dependent on the rise and fall of its competitor.
In this project, we have delved deep to identify non-linear causality between such companies using a deep neural network. There are two main implementations done i) Feed
Forward network with Group Lasso and ii) LSTM network with Group Lasso.
Each of the above architecture is designed in such a way that the weights of
competitor stocked can be interpreted and checked if they are significant in the
final predicted value. We used the stocks of EBay, Amazon, and Apple for predicting the stocks of Facebook. We got non-zero weights values for EBay, Amazon, and Apple thus indicating Granger Causality.
Conclusion
Analyzing two different methods using Multi-Layered Perception (MLP) and
LSTM using group Lasso technique, it seen that each of them have pros and
cons. While MLP is simple and more interpretable, it needs the number of lag
variables pre-specified. The LSTM, although a bit complex learns the number
of lag variables required for the final prediction once it is provided with K - lag
values.
Thus, we would like to conclude saying that building the right model by selecting the best hyper parameters helped in identifying Granger causal relationship even in a constantly varying dataset like stock prices. This implementation
could be used in various other domains having more predictable outcomes to identify such relationship with a very high accuracy.
| Project Report: | See |