Identification and analysis of risk factors of peer to peer lending

Področje množičnega posojanja je relativno nova investicijska priložnost. Ravno zato trenutno ne obstaja veliko raziskav, ki se ukvarjajo s tveganji, ki jih take platforme prinesejo. Glavni namen diplomskega dela je ugotoviti, ali lahko potencialni investitorji ocenijo tveganje potencialne investicije s pomočjo modelov strojnega učenja in podatki, ki so na voljo javnosti. Namen dela je tudi ugotoviti vpliv pandemije koronavirusne bolezni COVID - 19 na profil posojil in stopnjo njihove uspešnosti. Izvirni nabor podatkov smo razširili in obogatili s podatki iz zunanjih virov, nad njimi izvedli metode čiščenja (iskanje osamelcev, pretvorbe v enotno valuto) ter preizkusili različne napovedne modele (logistična regresija, naključni gozdovi, nevronske mreže in algoritem AdaBoost). Najbolje z F1-oceno 0,783 se je najbolje obnesla nevronska mreža, ki je pravilno klasificirala 78,5 % posojil, takoj za njo je bil klasifikator naključni gozd. Med njunima zmogljivostma statistično pomembnih razlik ni bilo. Zaradi večje količine podatkov, ki jih o značilkah dobimo, smo si za napovedni model izbrali logistično regresijo. Med analizo smo ugotovili, da so faktorji, ki na uspeh posojil vplivajo najbolj njihova dolžina, višina, obrestna mera ter sam tip posojila, medtem ko makroekonomski indikatorji ter ocene posojil slabo ocenjujejo njihovo pričakovano uspešnost.The field of peer-to-peer lending is a relatively new investment opportunity. As such, there is not much research regarding the risk factors of loans on this kind of platform. This thesis aims to determine whether potential investors can estimate their expected risk using machine learning models and publicly available data to aid in the process. Additionally, we aimed to investigate the effect of the COVID-19 pandemic on loan profiles and success rates. The original dataset was expanded using additional data and cleaned (outlier detection, currency conversion). We tested several classificators (logistic regression, random forests, neural networks and the AdaBoost algorithm) on the expanded dataset. The neural network performed best with an F1 score of 0,783 and a classification accuracy of 78,5 %, followed closely by the random forest classificator, although a significant difference between the two models was not identified. Nevertheless, due to its\u27 ability to reveal more information about feature importance, the logistic regression model was chosen.Our analysis revealed that the most influential factors, in regards to loan performance, are their length, initial amount and loan rate percentage alongside the loan type itself. On the contrary, macroeconomic indicators and loan grades proved to be bad predictors of loan outcomes