Active allocation in private debt using natural language processing with alternative data sources

Abstract

Portfolio analysis is benefiting from the surge of alternative sources of data coupled with new modelling frameworks, introduced by machine-learning. I collected alternative data and applied new frameworks (machine-learning techniques and technology) to the domain of private debt. This is an interesting and complex asset class, which has a significant shortage of data from which to model. To counter this issue, I incorporated advanced macro-finance and asset-pricing models. Such will provide the correct context in which to model this asset class as part of a sophisticated multi-asset portfolio construction framework. To ensure that the credit-risk models are fully understood, I selected a modelling technique from a broad array of options in a mature environment of credit modelling largely performed in banking (whilst ensuring the technique is suitable for asset management). The modelling framework is geared to account for the dynamics of business cycles, this being an important results driver in unlisted credit and other asset classes alike. My thorough macro-finance research allows me to design non-trivial processes to incorporate alternative signals as part of an asset-pricing framework on which to generate information for use in portfolio construction via the data simulated. My economic scenario generator considers the relative changes to asset classes at various points in the business cycle as part of a long-term investment program suitable for a defined-benefit pension funds portfolio. The final portfolio models are put together using a reinforcement learning framework. This framework connects the macro- finance theory dealing with the business cycle dynamics, together with the credit-risk techniques, to portfolio modelling techniques which I believe compounds to a sophisticated modelling framework for strategic asset allocation in a data-sparse environment.

Portefeulje-analise trek voordeel uit die opwelling aan alternatiewe databronne, gekoppel met nuwe modelleringsraamwerke wat uit masjienleer spruit. Ek het alternatiewe data versamel en nuwe raamwerke (masjienleertegnieke en -tegnologie e) toegepas in die gebied van private skuld. Dit is 'n interessante en komplekse bateklas wat 'n beduidende gebrek aan modelleringsdata het. Om die kwessie die hoof te bied, het ek gevorderde makrofinansi ele en batebeprysingsmodelle ge nkorporeer. Daardeur word die korrekte konteks verskaf vir die modellering van hierdie bateklas as deel van die raamwerk vir die 'n geso stikeerde meerbate-portefeuljekonstruksie. Ten einde te verseker dat die kredietrisiko-modelle volledig begryp word, het ek 'n modelleringstegniek gekies uit 'n uitgebreide reeks opsies binne 'n volgroeide kredietmodelleringsomgewing wat hoofsaaklik in die bankwese verwend word (onderwyl daar verseker word dat die tegniek geskik is vir batebestuur). Die modelleringsraamwerk is gerat om die dinamiek van die konjunktuur (sakesiklus) in berekening te neem aangesien dit 'n beduidende dryfkrag is vir ongenoteerde krediet en eweneens vir ander bateklasse. My deeglike navorsing in makro nansies laat my toe om nie-triviale prosesse te ontwerp om alternatiewe seine as deel van 'n batebeprysingsraamwerk waarop inligting vir gebruik in die portefeuljekonstruksie gegenereer word, te inkorporeer. My generator vir ekonomiese scenarios neem die relatiewe veranderings in bateklasse op verskillende tydstippe in die konjunktuur in ag as deel van 'n langtermyn-beleggingsprogram wat geskik is vir 'n gede nieerde voordeel-pensioenfondsportefeulje. Die uiteindelike portefeuljemodelle word saamgestel deur 'n versterkingsleer-raamwerk te gebruik. Die raamwerk verbind die makro nansies-teorie vir die konjunktuurdinamiek asook die kredietrisiko-tegnieke met portefeuljemodelleringstegnieke wat (volgens my) tesame 'n geso stikeerde modelleringsraamwerk vir strategiese batetoewysing in 'n data-arm omgewing daarste.

Ukuhlaziywa kwephothifolio kuyinzuzo ekuqhubukeni kominye imithombo yedatha kanye nezihlaka eziyisibonelo eyethulwe yinqubo yokufunda yomshini. Ngiqoqe olunye uhlobo lwedatha mtase ngifaka uhlaka lusha ( inqubo yokufunda yomshini enobuchwepheshe ) esizindeni sesikweletu sangasese. Loki kuyinto ethokozisayo futhi nesigab sempahla esiyinkimbinkimbi, okunoku shoda kakhulu kwedatha ekuzomodelwa kuyo. Ukulwa naloludaba , ngihlanganise imali enkulu ethuthukiswe nokuthuthukisa imodeli yentengo yempahla. Lokhu kuzohlinzeka umongo olungile lokumodela isigaba sempahla nje ngenxenye yohlaka oluyinkimbinkimbi lokwakhiwa kwephothifoliyo yezimpahla eziningi. Ukuqinisekisa ukuthi kamamodeli engciphe yekhredithi aqondwa ngokugcwele, ngikhethe imodeli elinobuchwepheshe kusuka kuhlu olubanzi uma izinketho endaweni evuthiwe yokumodela izikweletu eyenziwa kakhulu emabhange (ngenkathi eqinisekisa ubuchwepheshe okufanele ukuphatha kwempahla). Loluhlaka lokumodela kwenzelwe ukulandelisa ngokuguquguquka kwemijikelezo yebhizinisi lokhu kuyimiphumela ebalulekile ekushayeleni kwekhredithi engafakwanga nezinye isigaba zempahla ngokufanayo. Ugcwaningo lwemali enkulu olujulile lungivumele ukuklama inqubo engeyona into encane ukuhlanganisa ezinye izimpawu njengenxenye yohlaka lwentengo yempahla lokhu kuzoletha imininingwane yokusebenzisa ukwakhiwa kwephothifoliyo ngedatha elingisiwe. Ijeneretha yami yesimo somnotho ibheka izinguquko ezihambisanayo ukuthola ikilasi lempahla ezindaweni ezahlukahlukene kwezamabhizinisi nje ngenxenye yohlelo lotshalozimali lwesikhathi eside ezifanele iphothifoliyo yezikwama zempesheni ezichaziwe. Amamodeli nephothifoliyo okugcina ahlanganiswe ngokusebenzisa uhlaka lokufunda oluqinisayo. Loluhlaka luhlanganisa imibono yezimali ezinkulu ekubhekene nemijikelezo okushintshashinrsha kwebhizinisi, kanye nequbo yobungozi besikweletu, kumasu nokumodela iphothifoliyo engikholwa engikholwa ihlangisa uhlaka lokumodela oluyindida ngokwabiwa kwempahla yeah endaweni eyingcosana yedatha