search

LouisChen1992 / Deep_Learning_in_Asset_Pricing

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3350138
120 stars 50 forks source link

Questions about Macro Hidden States #6

Open quantmacroeco opened 2 years ago

quantmacroeco commented 2 years ago

Would you mind provide little details about how to get the hidden states using lstm model? For example ,what loss function did you use?

ChadyRaach commented 2 years ago

@quantmacroeco I think that the two networks (SDF Network and Conditional network) are trained end-to-end. This means that the LSTM of each of the two models will be trained using the same backpropagation process as their corresponding network. I quote the paper:

The two LSTMs are based on the criteria function of the two networks, that is h_t are the hidden states that can minimize the pricing errors, while h^g_t generate the test assets with the largest mispricing