Quicken Loans 有兩個(gè)系統(tǒng)：
LOLA: captures the system where people can record all the behavior that the lender does before underwriter.
LOLA 系統(tǒng)非常強(qiáng)大专控，每一個(gè)申請過房屋貸款的都會記錄在案则奥，例如說banker 打了多少次電話，顧客回復(fù)了多少次狭魂，
回復(fù)的頻率是多少洽损，顧客查詢了多少次的proapproval document,在這個(gè)過程當(dāng)中有沒投訴售葡，顧客申請初次貸款的金額產(chǎn)品载慈，已經(jīng)leads的情況是多少毅桃。有多少顧客在underwriting 之前就drop 掉的，概率是多少翁狐。
當(dāng)進(jìn)入underwriting 的狀態(tài)（23） 类溢，就會進(jìn)入到AMP系統(tǒng)，這里主要是underwriting 谴蔑，title source 的記錄（check)到status clear (closing)
我們通過業(yè)務(wù)的梳理豌骏，基本上是做feature selection, 我們并沒有用一個(gè)model 來做prediction ，而是通過不同的申請貸款過程中的不同階段隐锭，挑選出非常關(guān)鍵的階段：例如從mortgage banker 到preapproval letter, preapproval letter 到 criteria checking 窃躲，從underwriting 到title source, title source 到closing的幾個(gè)階段來做variable selection. 然后分別建立decision tree模型。
conversion rate prediction:
Regression with time series errors,
variable:
page views,
unique visitors
ten year treasury yield
- Variable Selections: From business side: the channeling mix and interest rate. Originally the channeling mix has 15 channels including leadby, social media etc. As for some channeling mix data, the leadby data are missing. So I narrow down to the variables into 7, 10 year treasury yields, direct search, paid search, online ads, email, affiliate networks, relationship marketing. This part uses multiple regression technique to do the variable selection with lowest AIC and pass the F-test and t-test for independent variables.
- Modeling Procedure:
? Check that the forecast variable and all predictors are stationary.
? Fit the regression model with AR(2) errors for non-seasonal data or ARIMA(2,0,0)(1,0,0)m errors for seasonal data. Since this model does not have any seasonality, so I fit the model with AR(2) first.
? Calculate the ARIMA errors (Nt) from the fitted regression model and identify an appropriate ARMA model for them. Also check the AIC value.
? Re-fit the entire model using the new ARMA model for the errors.
? Check that the et series looks like white noise.
? Pass the Ljung-Box test to show that the errors are uncorrelated
? Evaluation of the model.
