machine learning - What type of neural network would work best for credit scoring? -

let me start saying took undergrad ai class @ school know enough dangerous.

here's problem i'm looking solve...accurate credit scoring key part success of business. rely on team of actuaries , statistical analysis suss out patterns in few dozen variables track each individual indicate may low or high credit risk. understand type of job neural nets great @ solving, is, finding high order relationships across many inputs human never spot , rendering decision or output on average more accurate trained human do. in short, want able input name, address, marital status, car drive, work, hair color, favorite food, etc in , credit score back.

my question type or architecture neural network best particular problem. i've done bit of research , seems i'm generating questions faster i'm finding answers @ point. best i've been able come kind of generative deep neural network multiple hidden layers each layer able abstract 1 level beyond previous one. im assuming it's going feed-forward because seems default. have historical data on previous customers including information used make initial score data on type of credit risk turned out be. seem lend unsupervised learning. i'm lost in number of layers, how layers different each other, size of each layer, connectedness of each of perceptrons , on. more dig more i'm getting research papers on head need smart person point me in right direction

does have ideas? again, don't need thorough explanation general area should focus on.

this supervised learning since have actual data can labelled. it's feedforward since you're not predicting time series assigning scores. further, should prepare data (assigning credit scores manually or rough heuristic) , start experimenting tools before invest time implementing state-of-the-art architectures. multi-layer-perceptron (mlp) 1 hidden layer sufficient starting point such problem. there on, can train network generalize credit assignment heuristic began with.

you should know "new" architectures read while researching dealing more difficult problems credit scoring (speech/image/character recognition/detection). there collection of papers on scenario of credit scoring / risk classification, i'd recommend reshifting focus architectures actual case studies (see e.g. this paper). pick recent paper mlps , apply parameters. start simple , improve system incrementally (as @roganjosh stated).