About Bookrule

Bookrule is an analysis aid for the fraud and data analysts inside a lender. They load their own loan book, and its repayment history where they have it, and the tool returns the loans whose data points at fraud: borrowers who stopped paying on the first instalments, repayment patterns that do not fit, payments that do not reconcile with the balance or the status, incomes and ratios piled up inside an underwriting cut-off, the same borrower appearing more than once, and segments whose losses are out of line with the book. Each loan comes with a reason an analyst can paste into a case note. It is not an independent loan evaluation, it does not price anything, and it does not replace the fraud team.

The name is the book's own rule, the dashed line of the book's rate on every chart; a loan that falls through it is the one to look at.

Users

Fraud teams and data analysts at a lender, reviewing the book they already hold. The method is fixed: two runs, or two books, can be set side by side.

Method

Analyses use the standard field names below.

The upload is mapped to the standard schema: columns are matched by name, a language model on the local network proposes matches for the rest from column names and sample values, and the analyst confirms the mapping before any value is converted. Data checks run on the result. Each loan is then tested on its own row (payments against balance and status, instalment against amortisation, dates in order), against the rest of the book (duplicate identities, incomes and ratios bunched at a cut-off, payments stopping on the first instalments) and, where payment history is loaded, month by month. Each segment of each dimension is tested against the book for losses out of line with it. Flagged loans are ranked by a review score built from the evidence on them, and the demo shows the result on 886,837 Lending Club loans.

Early default, strength and the review score

Early default: charged off or late, with no payment received or a last payment within six months of origination.

Reasons carry a strength: strong (a loan to query on this alone: payment identities broken, status against the cash, no payment received, instalment off amortisation, a shared identity cluster), medium (early default within three months, a pile-up just inside a limit, round unverified income with an early default, dates out of order), weak (context: pricing residuals, loan against income, sentinel values).

Points: strong 30, medium 15, weak 5; half for an integrity reason; up to 15 more for an unusual combination of application values, capped at 100. The review score ranks loans for review. It is not a probability.

Loans to review holds loans with a fraud-pattern reason and a review score of 15 or more. Data to confirm holds loans whose only reasons are integrity reasons: the loan's own figures do not agree, a question for servicing or the data team.

Standard schema. Required fields must be mapped. Checks that need a missing optional field are skipped and listed on the data checks page.
FieldTypeMeaningRequired
loan_id identifier Unique loan identifier. Required
origination_date date Date the loan was issued or funded. Required
original_balance money Principal amount originated. Required
interest_rate rate Annual interest rate at origination, as a percentage. Required
term_months integer Contractual term in months. Required
status status Loan status at the snapshot date. Required
installment money Scheduled periodic payment.
grade category Lender risk grade or band at origination.
sub_grade category Finer risk grade, if the lender has one.
credit_score integer Bureau credit score at origination.
dti rate Debt-to-income ratio at origination, as a percentage.
annual_income money Borrower's stated annual income.
employment_length_years number Years in current employment.
home_ownership category Housing status: own, mortgage, rent, other.
income_verified category Whether income was verified at underwriting.
purpose category Stated purpose of the loan.
region category Borrower geography: state, region or country.
channel category Origination or listing channel.
current_balance money Outstanding principal at the snapshot date.
principal_paid money Cumulative principal received.
interest_paid money Cumulative interest received.
total_paid money Cumulative payments received, all components.
recoveries money Cash recovered after charge-off.
last_payment_date date Date of the most recent payment received.
charge_off_date date Date the loan was charged off or written off.
raw_status text The lender's own status label, kept verbatim.
funded_balance money Amount actually disbursed, where it differs from the amount originated.
application_date date Date the borrower applied.
first_payment_date date Date the first instalment was due.
broker_id category Broker, dealer or introducer that sourced the loan.
employer identity Borrower's employer as stated. Stored hashed.
postcode identity Borrower's postcode or zip. Stored hashed.
date_of_birth identity Borrower's date of birth. Stored hashed.
borrower_id identity Lender's identifier for the borrower, shared across their loans. Stored hashed.
modification_flag category Modification, forbearance, hardship, settlement or payment-plan marker.
last_payment_amount money Amount of the most recent payment received.
fees_paid money Cumulative fees received, late fees included.
bureau_inquiries_6m integer Credit bureau inquiries in the six months before origination.
open_accounts integer Open credit accounts at origination.
earliest_credit_line date Date the borrower's earliest credit account opened.
revolving_utilisation rate Revolving balance as a percentage of revolving limits at origination.
public_records integer Derogatory public records at origination.
delinquencies_2y integer Delinquencies in the two years before origination.
accounts_opened_24m integer Credit accounts opened in the 24 months before origination.
months_since_last_inquiry integer Months from the most recent bureau inquiry to origination.
latest_credit_score integer Bureau credit score at the latest pull, after origination.

Statuses

Source status labels are mapped to current, paid, late, charged off or other. The original label is kept in raw_status. Labels mapped to other on more than 1% of loans raise a data check.

Losses

Loss rates are to date: a loan is a loss if its status at the snapshot date is charged off (written off, where the file says so). Loans still current may charge off later. A segment is a finding when its loss rate by count differs from the book by three or more standard errors after allowing for the spread between segments (Spiegelhalter, 2005), on at least 1% of balance and 200 loans. Findings are ranked by excess losses: the segment's charged-off balance beyond what it would hold at the book rate, as a share of the book's charged-off balance.

Data checks

Blocking checks stop the analysis: duplicate loan ids above 0.1% of rows, a required field failing to convert on more than 5% of rows, more than 5% of loans with an unrecognised status, or origination dates after the snapshot. Warnings and notes do not.

Demo

Lending Club 2008–2015: 886,837 loans originated January 2008 to December 2015, status at March 2019, read as if it were one lender's own book. Unsecured US consumer instalment loans originated on the Lending Club platform, 36- and 60-month terms, from the public accepted-loans table. Open the demo

Caveats

  • 2007 is excluded: 603 loans over seven months, a partial year.
  • 2016 to 2018 are excluded (1,373,228 loans): under 36 months of seasoning at the snapshot, and mostly still current.
  • The snapshot date is the latest last payment date in the file (March 2019).
  • Credit score is the low end of the FICO range at origination.
  • Latest credit score is the low end of the FICO range at the platform's last pull (last_fico_range_low); a zero, meaning no score, is read as missing.
  • Accounts opened in 24 months (acc_open_past_24mths) and months since the last inquiry (mths_since_recent_inq) are blank on the early vintages, which the file does not carry them for.
  • Original balance is the requested loan amount (loan_amnt); the amount funded (funded_amnt) is mapped separately and used by the payment checks.
  • Postcode is the first three digits of the zip code, and employer is the borrower's job title as typed (emp_title), not an employer name; both are stored hashed.
  • The modification flag combines the debt settlement, hardship and payment plan markers.
  • Statuses 'Default' and 'Charged Off' map to charged off, 'In Grace Period' to current, and the two 'Late' statuses to late.