# LimitWise working file: data remediation log (Week 3)

CORVANA BANK: MODELING TEAM

LIMITWISE MODELING EXTRACT: DATA REMEDIATION LOG

*Prepared by the modeling team for the LimitWise program file, in response to the Week 2 review. One row per change the team has now made to the extract before the first model is trained, and one row per finding the team chose to document rather than change. Two terms are used the same way in every row. The working file is every column the team holds after remediation: the fields as loaded, the fields added beside them, and the outcome. The model inputs are the subset of those columns the model will be trained on. A field can be in the working file and not a model input. Two rules apply to every row. The raw field is never overwritten. A field held out of the model inputs stays in the working file for measurement. The team takes the label off the account, not the account out of the file. Rows leave only under a written rule, and one row of this log (R-8) is the only place that happens.*

**Extract:** 30,000 rows and 25 columns as loaded (23 fields, the account identifier, and the outcome). 29,995 rows and 40 columns in the working file after remediation (15 columns added, 5 rows removed under R-8). Those 40 columns are 27 model inputs, 12 fields held out of the inputs, and the outcome column. Outcome: default_payment_next_month, value of interest 1 ("default"), 22.1 percent of 29,995 (6,635 accounts, unchanged at one decimal from 22.1 percent of 30,000, and not the book rate).

| Row | Field or fields | What the Week 2 review found | Operation performed | Reason | Options considered and not taken | Before and after | Where it lands | Notebook |
|---|---|---|---|---|---|---|---|---|
| R-1 | The extract as drawn (all rows) | 22.1 percent default, 6,636 of 30,000, against a book rate the dictionary puts near 3 percent, about 7.4 times. The dictionary says the sample was drawn to include enough defaults to learn from and does not say how | Leave as is and document. Reweight or recalibrate for enrichment: recorded as a modeling-stage decision, not performed | The enrichment is a fact the file can show. The method is open: no sampling memo or extraction query is held. If the draw was random within each outcome, a defaulting account entered the extract about 9.2 times as often as a non-defaulting one, and that is the figure a reweighting would use. Without the method, reweighting fixes the mix but not a biased draw. No threshold, loss criterion, or calibration claim is made against 22.1 percent | Reweight non-defaulters now, about nine to one. Calibrate now on a book-like holdout. Set a threshold on the extract's rate | No change to the data. 22.1 percent, 3 percent, 7.4 times, 9.2 times, printed | Not a field. A figure (the enrichment) and an open item (the method) recorded for the modeling and evaluation stages | 11.1 |
| R-2 | PAY_0, PAY_2 to PAY_6 (six fields) | Values 0 and -2 appear in every repayment-status field, are not documented, and 0 is the most common value in every month. 25,939 records, 86.5 percent, hold one of them somewhere. Both default below the documented paid-duly code in PAY_0: 0 at 12.8 percent, -2 at 13.2, -1 at 16.8. The six fields are the six strongest predictors in the file | Recode undocumented values into a named level: six new fields, STATUS_0 and STATUS_2 to STATUS_6, beside the raw fields. -1 and 1 to 9 keep their documented meaning in words. 0 and -2 both become "no delinquency recorded, code undefined" | The meaning of 0 and -2 is open and stays open. The level name says what is known and nothing more: the documented delinquency codes are 1 to 9, and these accounts carry none. What the accounts did is visible in the documented amounts: in PAY_0, code 0 has a median prior statement of 48,261 and a median payment of 3,000, and code -2 a median prior statement of 1,099 with 34.5 percent having no prior balance at all. A reason for an account action must be writable in plain words, and "your status was 0" is not. The raw codes stay so the evaluation stage can test merged against separated under the Chief AI Officer's 5 percent tripwire | Define the codes from their behavior (asserts a meaning nobody confirmed). Drop the six fields from the inputs (the six strongest predictors would go). Leave the raw codes as the inputs (reasons could not be written) | Six raw fields with 11 values each, then six recoded fields beside them with 10 levels. "No delinquency recorded, code undefined" holds 17,496 records in STATUS_0 and 21,181 in STATUS_6. Rows 30,000 before and after | Raw in the working file, not model inputs. Recoded copies are the model inputs | 11.2 |
| R-3 | BILL_AMT1 to BILL_AMT6, PAY_AMT1 to PAY_AMT6 | The status codes cannot be defined, but the statement balances and the payments are documented, most recent first | Derive a feature from documented fields: five new fields, PAID_1 to PAID_5, each payment set against the prior month's statement balance under one rule: no balance (prior statement zero or negative), paid in full, paid part, paid nothing | One rule for every month, so the evaluation stage can test whether the model leans on a status code whose rule changed (R-4). Every level is a plain word a reason could use. The earliest month's payment (PAY_AMT6) is set against a statement the extract does not hold, so there is no sixth field, and the team says so rather than inventing one. Nothing here defines a code: the table that sets PAID_1 beside PAY_0 shows counts only | A sixth field with an assumed prior statement. Same-month pairing (payment against the balance of the same month, which is not what the documentation describes). Derive nothing and rely on the codes | Five fields added. PAID_1: paid part 17,782, paid in full 6,823, no balance 3,175, paid nothing 2,220. Of the 14,737 accounts with code 0 in PAY_0, 13,956 paid part of the prior statement. Rows unchanged | PAID_1 to PAID_5: model inputs, added. The twelve raw amount fields: model inputs, unchanged | 11.3 |
| R-4 | PAY_0, PAY_2 to PAY_6 | Value 1 appears 3,688 times in PAY_0 and then 28, 4, 2, 0, and 0 in the older months, while value 2 holds between 2,626 and 3,927 in every month. The fields are named PAY_0 and PAY_2 to PAY_6. There is no PAY_1 | Leave as is and document, as a coding inconsistency across months | The most recent month's status was computed by a different rule from the older months, and the team cannot say what that rule was or why the names skip 1. Six fields sharing a name and not a rule is a risk the evaluation stage tests with the PAID fields beside them (R-3), which carry one rule for all months | Use only PAY_0 and discard the five older months (five months of documented amounts would go with them). Recompute the older months (no rule is known to recompute by) | No change to the data. The counts by month are printed | Not a field. An open item recorded for the evaluation stage (whether one rule computed all six months) | 11.4 |
| R-5 | EDUCATION, MARRIAGE | EDUCATION holds three codes the dictionary does not document, 0 (14 records), 5 (280), and 6 (51), 345 in all. MARRIAGE holds one, 0 (54). Their default rates are a fraction of the file's, 0.0, 6.4, 15.7, and 9.3 percent, on groups too small to trust | Recode undocumented values into a named level: EDUCATION_MODEL and MARRIAGE_MODEL beside the raw fields, with 0, 5, and 6 in EDUCATION and 0 in MARRIAGE as one level, "unknown" | The dictionary states that all four codes appear in the data and are not documented in the codebook, and nothing else the team holds documents them. What any of the four records is open, and the recode does not close it. A named "unknown" is honest, and it stops the model from learning that one undocumented code is safe. The raw codes stay so the evaluation stage can test the merge against the codes kept separate under the Chief AI Officer's one-point tripwire | Treat the codes as missing and impute (nothing is missing, and imputing a demographic field is never done). Merge into "other" (asserts they belong with the documented "other" level). Drop the records | EDUCATION 7 codes, then EDUCATION_MODEL 5 levels, "unknown" 345 records at 7.5 percent. MARRIAGE 4 codes, then MARRIAGE_MODEL 4 levels, "unknown" 54 records at 9.3 percent. Rows unchanged | Raw EDUCATION and MARRIAGE: in the working file, not model inputs. EDUCATION_MODEL: model input. MARRIAGE_MODEL: in the working file, not a model input (R-10) | 11.5 |
| R-6 | default_payment_next_month | The dictionary defines the label as default on payment in the following statement cycle, one month, and says nothing operational | Define the label window: leave as recorded and document. The documented window is one statement cycle. The operational definition is open | The label sets the base rate and the meaning of every error. A definition the team supplied would be a guess dressed as documentation. Whether the label's definition matches the loss forecast's definition of default is open, and the mismatch, if any, is stated rather than resolved | Redefine the label to a standard such as days past due (invents a definition). Relabel from the payment fields (replaces the recorded outcome with a derived one) | No change. 6,636 default (22.1 percent), 23,364 no default (77.9 percent) as loaded | The outcome column, unchanged. Not a model input | 11.6 |
| R-7 | LIMIT_BAL, BILL_AMT1 to BILL_AMT6, PAY_AMT1 to PAY_AMT6 | The unit of the amounts is not stated. 1,930 accounts (6.4 percent) carry a negative balance in some month, smallest -339,603. The largest statement balance, 1,664,089, and the largest payment, 1,684,259, each exceed the largest credit limit, 1,000,000 | Leave as is and document. No cap, no trim, no sign correction | The unit is open, and the amounts are used as relative values. The dictionary says balances may be negative. 3,931 accounts (13.1 percent) carry a balance above their own limit in at least one month, so above-limit balances are common, not a handful of outliers. Capping real extremes would hide exactly the accounts the model most needs to see, and correcting a value nobody has confirmed is wrong would assert something | Cap or winsorize the amounts. Set negative balances to zero. Drop the accounts above the limit | No change to the data. The maxima, the counts, and the unit line are printed | All thirteen fields: model inputs, unchanged. The unit and the extremes: open items | 11.7 |
| R-8 | All rows (ID used by the rule) | 35 patterns repeat in every field except ID, 70 rows. 30 pairs show no activity in any month. 5 pairs carry six months of identical non-zero balances or payments | Drop rows under a stated rule: identical in every field except ID, with a non-zero balance or payment in at least one month. Keep the lower ID, remove the other | Two accounts matching in every one of 24 fields including six months of activity is one account extracted twice under two identifiers. Two unused cards with the same limit and application details can coincide, so the 30 inactive pairs stay. Removal happens before the split rule (R-12) so no copy of an account can sit on both sides. This is the only row removal in the log. Which of the two identifiers is correct is open. ID is the account identifier, never a model input, kept to trace rows | Keep all 70 rows (five accounts would be doubled across the split). Drop all 70 (removes 60 rows the team cannot show are duplicates). Drop the pattern rather than one row of it | 30,000 rows before, 29,995 after, 5 removed. The 30 inactive pairs, 60 rows, stay. Rate of "default" 22.1 percent after (6,635 of 29,995) | Every other row stays. ID: in the working file, not a model input | 11.8 |
| R-9 | BILL_AMT1, LIMIT_BAL, PAY_AMT1, BILL_AMT2 | The six statement-balance fields correlate at 0.80 to 0.95 (section 8a), so the top reason for an account can shift between near-identical fields at every retrain | Derive a feature from documented fields: UTIL_RECENT = BILL_AMT1 / LIMIT_BAL, and PAY_RATIO_RECENT = PAY_AMT1 / BILL_AMT2 where BILL_AMT2 is above zero. Both from the most recent month. Raw fields kept as inputs | Both are quantities a person would reason with, from documented fields, in the same unstated unit as their sources, so both are relative. Where there was no prior balance to pay, the ratio is left empty rather than invented, and PAID_1 (R-3) already records "no balance" for every such account. The evaluation stage tests reason stability across retrains under the 10 percent tripwire | Drop five of the six balance fields (loses month-to-month information). Fill the empty ratio with a value (asserts a payment behavior where there was nothing to pay). Derive nothing | Two fields added. UTIL_RECENT median 0.314, 2,113 accounts above 1. PAY_RATIO_RECENT median 0.078, 6,820 accounts at 1 or more, empty for 3,175 accounts (10.6 percent), the only empty values in the working file | UTIL_RECENT and PAY_RATIO_RECENT: model inputs, added. The raw fields: model inputs, unchanged | 11.9 |
| R-10 | SEX, AGE, MARRIAGE_MODEL (and the raw MARRIAGE, already out under R-5). EDUCATION_MODEL stays | Sex, age, and marital status are prohibited bases under the Equal Credit Opportunity Act and Regulation B. The four demographic fields are among the weakest predictors in the file, 0.034 to 0.074 (section 8b). The documentation never says whose demographics a supplementary-card account carries | Hold protected attributes out of the model inputs, keep them in the working file for measurement. EDUCATION_MODEL stays a model input, with monitoring, pending the evaluation-stage test | You cannot measure what you deleted. The Chief AI Officer's watch list measures false alarm and miss rates by sex, age band, marital status, and education, and tests whether the remaining inputs recover each field. The Chief AI Officer declined an untested drop in Week 2 and ruled that the decision is made on the test (with and without, worth more than half a point of area under the curve, AUC). This row records the recommendation and the test. The team takes the label off the account, not the account out of the file: every cardholder stays with all of their other fields, and the model is simply never told their sex, age, or marital status. One limitation sits on all four fields: whose demographics a multi-card account carries is open | Delete the four fields from the working file. Keep all four as inputs and test later. Drop EDUCATION now, ahead of the ruling | 30 model inputs, then 27. All four demographic fields and their recoded copies still in the working file. Rows 29,995 before and after | SEX, AGE, MARRIAGE, MARRIAGE_MODEL: in the working file, not model inputs. EDUCATION_MODEL: model input, monitored. The supplementary-card question: an open item on every measurement by these fields | 11.10 |
| R-11 | Credit bureau data (no field) | Nothing in the extract or its documentation refers to bureau data | Leave as is and document: no join has been made | No bureau field is present and the dictionary refers to none. Whether bureau data could join these accounts is open. Any later join is a new extract and a new datasheet | Join bureau data now (nothing to join). Assume it is absent from the book (unknown) | No change. The field list as loaded is printed | Not a field. An open item recorded in the datasheet | 11.11 |
| R-12 | Train/test split | The six months are columns, not rows, so there is no time index across rows. After R-8 each account appears once | Fix the train/test split: random by row, stratified on the outcome, 80 percent training and 20 percent held back as the test set, fixed seed, after the R-8 deduplication. Recorded now, executed at the modeling stage | The split decides whether the evaluation numbers are honest. With no time index and each account once, random stratified is the defensible rule, and the datasheet says so. The held-back test set is never used for training | Split by time (no time index across rows). Split by account (each account appears once after R-8, so a random split of rows already keeps every account on one side). Resample before the split (leaks) | Expected sizes: 23,996 training rows (about 5,308 defaults), 5,999 in the held-back test set (about 1,327 defaults). For the smallest measurement groups: age 65 and over, 110 records, about 22 in the held-back set. MARRIAGE_MODEL "unknown," 54 records, about 11. EDUCATION_MODEL "other," 123 records, about 25. Not executed this week | Not a field. A rule recorded for the modeling stage | 11.12 |

**Open items carried in the datasheet, not resolved by any row above:** how the 30,000 accounts were drawn and the book default rate for the period they cover, what repayment-status values 0 and -2 record in each of the six fields, whether one rule computed all six months and why there is no PAY_1, what EDUCATION 0, 5, and 6 and MARRIAGE 0 record, what "defaulted on payment" means operationally and whether it matches the loss forecast's definition, the unit of the amounts and how a balance or a payment can exceed the largest credit limit, whether the 30 inactive repeated patterns are one account or two and which identifier is correct for the five pairs treated as one, whose demographics a supplementary-card account carries, and whether credit bureau data can be joined. Section 12 of the notebook prints the same nine.

*End of log. Questions about any row go to the modeling team through the AI Governance Office.*