VGSF - WU Vienna - LC

Andreas Neuhierl, Washington University in St. Louis

Campus WU D4.4.008 11:00 - 12:00

Organizer WU Vienna

The Fin­ance Brown Bag Sem­inar is held by the In­sti­tute for Fin­ance, Bank­ing and In­sur­ance (WU Vi­enna) and the Vi­enna Gradu­ate School of Fin­ance (VGSF). It serves as a present­a­tion plat­form for PhD stu­dents, fac­ulty mem­bers, and vis­it­ors. An over­view of BBS on the web­site of the In­sti­tute for Fin­ance, Bank­ing and In­sur­ance.


Andreas Neuh­i­erl, Wash­ing­ton Uni­versity in St. Louis, En­gel­ber­t-­Dock­ner Fel­low

Struc­tural Deep Learn­ing in Con­di­tional As­set Pri­cing

(with Ji­an­qing Fan, Zheng Tracy Ke, and Yuan Liao)

We develop new struc­tural non­para­met­ric meth­ods for es­tim­at­ing con­di­tional as­set pri­cing mod­els us­ing deep neural net­works. Our method is guided by eco­nomic the­ory and em­ploys time-vary­ing con­di­tional in­form­a­tion on al­phas and betas car­ried by firm­-spe­cific char­ac­ter­ist­ics. Con­trary to many ap­plic­a­tions of neural net­works in eco­nom­ics, we open the “black box” of ma­chine learn­ing pre­dic­tions by in­cor­por­at­ing fin­ance the­ory into the learn­ing, and provide an eco­nomic in­ter­pret­a­tion of the suc­cess­ful pre­dic­tions ob­tained from neural net­works, by de­com­pos­ing the neural pre­dict­ors as risk-re­lated and mis­pri­cing com­pon­ents. Our es­tim­a­tion method starts with peri­od-by-period cross-sec­tional deep learn­ing, fol­lowed by local PCAs to cap­ture time-vary­ing fea­tures such as lat­ent factors of the model. We form­ally es­tab­lish the asymp­totic the­ory of the struc­tural deep­-learn­ing es­tim­at­ors, which ap­ply to both in-sample fit and out-of-sample pre­dic­tions. We also il­lus­trate the “double-­des­cent risk” phenom­ena as­so­ci­ated with over­-­para­met­rized pre­dic­tions, which jus­ti­fies the use of over­-fit­ting ma­chine learn­ing meth­ods.

Back to overview