Chavín Microfinance — 90-day default risk

Chavín approves microloans in five regions. Risk wants every application in data/holdout.csv ranked, because the credit team can pull and review only 5% of them a week. They do not want a yes/no; they want the order.

Time: 90 minutes. Write everything in solution.py.

About 5% of applications default. A model that answers "no default" every time is 95% accurate, and finds nothing. Accuracy is not the metric here and never was.

Files

Columns

Known defects, stated as rules

1. income_monthly uses 999999 for "not declared". That is missing, not a large income.
2. Negative debt_to_income is invalid. Treat those as missing. Do not take the absolute value.
3. employment_sector gained codes after the training window closed. A handful of codes appear in holdout.csv and in no training row.
4. education is ordinal. Its ordering carries information that one-hot encoding discards.

Tasks

T1 — class_balance(df) -> dict (10 pts)

Exactly three keys: rows, defaults, default_rate (a share, not a percentage, rounded to four decimals).

T2 — unseen_levels(train, holdout) -> dict[str, list[str]] (15 pts)

Across the categorical columns employment_sector, education, channel and region, return {column: sorted list of levels present in holdout and absent from train} — only for the columns that have at least one.

T3 — clean(df) -> pd.DataFrame (15 pts)

Return the frame with: - 999999 in income_monthly replaced by missing; - negative debt_to_income replaced by missing; - a new integer column education_level: primaria=1, secundaria=2, tecnica=3, universitaria=4, postgrado=5.

No rows may be dropped. An undeclared income is a fact about the applicant.

T4 — design_matrix(train, holdout) -> (X_train, X_holdout) (20 pts)

Return two numeric matrices — arrays or DataFrames — with the same number of columns, one row per input row, and no missing values in X_holdout. Rule 3 is the whole point: whatever encodes the categories has to survive a level it has never seen, without widening the matrix and without losing the applicant.

T5 — fit_predict(train, holdout) -> np.ndarray (40 pts)

Both frames arrive raw. Return a risk score per holdout row, in file order — higher means likelier to default. Any monotone score works; probabilities are fine.

Graded on two numbers: - Average precision (area under precision–recall): must beat 2.5× the base rate, then >= 0.140 is competent and >= 0.155 is good. - Recall inside the 5% review budget: of all the applications that really defaulted, the share that lands in your top 5% must be >= 0.170.

Allowed: pandas, numpy, scikit-learn.

Visible cases — Chavín Microfinance — default risk

What CodeSignal shows next to Run tests: small inputs with the output they should produce. Run them with make drill-test. They use toy data, never the day's CSVs, so passing them is not a grade — make drill-grade is.

T1 — class_balance (How rare is the thing being predicted)

rows, defaults and a share rounded to 4 decimals

Input:

application_id  default_90d
            A1            0
            A2            0
            A3            1
            A4            0
            A5            0
            A6            0
            A7            1
            A8            0

Expected:

{'rows': 8, 'defaults': 2, 'default_rate': 0.25}

T2 — unseen_levels (Levels that only exist at scoring time)

only the sector column has levels the training file lacks

Input:

train:
application_id employment_sector education channel region
            A1            SEC-01  primaria  branch   Lima
            A2            SEC-02   tecnica     app  Norte
            A3            SEC-01 postgrado   agent    Sur

holdout:
application_id employment_sector education channel region
            H1            SEC-02   tecnica     app    Sur
            H2            SEC-90  primaria     app   Lima
            H3            SEC-91 postgrado  branch  Norte

Expected:

{'employment_sector': ['SEC-90', 'SEC-91']}  (sorted; no other keys)

nothing new means an empty dict

Input:

train:
application_id employment_sector education channel region
            A1            SEC-01  primaria  branch   Lima
            A2            SEC-02   tecnica     app  Norte
            A3            SEC-01 postgrado   agent    Sur

holdout:
application_id employment_sector education channel region
            H1            SEC-02   tecnica     app    Sur
            H2            SEC-01  primaria  branch   Lima

Expected:

{}

T3 — clean (Sentinels, impossible values, an ordering)

999999 becomes missing and no row is dropped

Input:

application_id  income_monthly  debt_to_income education  default_90d
            A1        999999.0             0.3  primaria            0
            A2          2500.0            -0.2 postgrado            1
            A3          3000.0             0.5   tecnica            0

Expected:

3 rows; income_monthly == [NaN, 2500.0, 3000.0]

a negative ratio becomes missing, not its absolute value

Input:

application_id  income_monthly  debt_to_income education  default_90d
            A1        999999.0             0.3  primaria            0
            A2          2500.0            -0.2 postgrado            1
            A3          3000.0             0.5   tecnica            0

Expected:

debt_to_income == [0.3, NaN, 0.5]

education_level follows the stated order

Input:

application_id  income_monthly  debt_to_income education  default_90d
            A1        999999.0             0.3  primaria            0
            A2          2500.0            -0.2 postgrado            1
            A3          3000.0             0.5   tecnica            0

Expected:

education_level == [1, 5, 3]

T4 — design_matrix (One design matrix for both files)

same width, every row kept, no NaN in the scoring matrix

Input:

train (6 rows):
application_id application_date employment_sector     education  income_monthly  credit_history_months  prior_defaults  requested_amount  debt_to_income channel region  default_90d
            A1       2025-09-01            SEC-01      primaria          1500.0                      3               0            3000.0             0.4  branch   Lima            0
            A2       2025-09-01            SEC-02    secundaria          2200.0                     40               0            5000.0             0.2     app  Norte            0
            A3       2025-09-01            SEC-01       tecnica        999999.0                     12               1            4000.0             0.9   agent    Sur            1
            A4       2025-09-01            SEC-03 universitaria          4100.0                     80               0            9000.0             0.3     app Centro            0
            A5       2025-09-01            SEC-02     postgrado          5200.0                    100               0           12000.0            -0.1  branch  Selva            0
            A6       2025-09-01            SEC-01       tecnica          2600.0                     24               2            2000.0             0.6   agent   Lima            1

holdout (3 rows, one sector SEC-90 the training file never saw):
application_id application_date employment_sector education  income_monthly  credit_history_months  prior_defaults  requested_amount  debt_to_income channel region
            H1       2026-07-01            SEC-02   tecnica        999999.0                     10               0            4000.0             0.5     app   Lima
            H2       2026-07-01            SEC-90  primaria          1800.0                      2               1            2500.0            -0.3   agent    Sur
            H3       2026-07-01            SEC-01 postgrado          6000.0                    120               0           15000.0             0.2  branch  Norte

Expected:

(X_train 6 x k, X_holdout 3 x k) for one k; numeric; X_holdout without NaN

T5 — fit_predict (Rank the applications)

one finite risk score per applicant, unseen sectors included

Input:

300-row toy training frame (every column, 999999 incomes, negative ratios); 20-row raw holdout whose first two sectors are SEC-90 and SEC-91

Expected:

array of 20 finite floats (higher = riskier)