preview

Explanation Of A Neural Network

Decent Essays

First of all, we have encountered one major problem that is how to interpret a neural network given its black box characteristics. We really wanted to try ourselves giving interpretation to our results so that we dug into the existing literature and found out a very interesting research paper written by Garson in 1991. In « Illuminating the black box: a randomization approach for understanding variable contributions in artificial neural networks », Olden and al. describes Garson’s algorithm very concisely so that we were able to create a user-defined function on Python that replicates the method. The interpretation of the method is provided below. The outputs of the different algorithms in the context of our study are the following: …show more content…

We don’t know the influence direction of the variable (given the PCA black-box transformation and Garson’s output) but we can formulate hypotheses on the fact that value companies tend to beat the market more often than growth companies. Dividend yield and EVtoEBITDA that are our 2nd and 3rd most relevant variables for Neural Nework output also support this previous fact high dividend yield paying companies characterize generally value companies, while the EVtoEBITDA ratio characterize the relative price of a company since it returns how many times an investor is willing to buy the EBITDA (Earnings before interests, taxes, depreciation and amortization) of a company when he purchases a share of the company. An investor that invests in high EVtoEBITDA companies has generally good growth prospects views regarding the recurrent earnings of the company. Finally, the factor size is the 4th most relevant factor that affects our output and this fact is also supported by Fama and French literature: portfolios formed over small capitalization companies tend to out-perform portfolios formed over big capitalization companies over time. It is interesting to note that this feature is an output of the

Get Access