SMB Underwriting Challenge - Participant Instructions

SUBMISSION LINKS: Links Sheet

STEP 0 (REQUIRED, DO THIS FIRST): Register your team

**Register here as early as possible: Form Link

You will not receive a submission link until you register. Don't leave this
to the last minute.

Provide your team name and, for each member (1-4 members per team):
full name (required), email (required). After you
register, we email your team a private link where you upload your submission.

The challenge

You are a small-business lender. Using a historical book of loan applications,
build a model that (A) decides whom to fund, (B) forecasts how those loans
default over time, and (C) answers causal "what-if" questions - then defend your
reasoning in a short writeup (D).

What you get

Everything is in this folder. Start with the dataset guide:
dataset/README.md.

The four deliverables

You submit exactly four files, with exactly these names. Scoring is
automated by joining on these names and IDs, so a wrong name or a missing ID
means your submission cannot be scored.

A - submission_A_decisions.csv (your lending policy)

Your approve/decline decision plus your calibrated probability of default (PD)
for every applicant.

• Rows: one per applicant in validation.csv + test.csv (13,306 total).

• Columns:

  column	meaning
  applicant_id	matches the dataset
  decision	1 = approve at the requested amount, 0 = decline
  predicted_pd	your PD point estimate in [0, 1] - required for everyone, including declines
  pd_lower_90	lower bound of your 90% interval on predicted_pd
  pd_upper_90	upper bound of your 90% interval

• Rule: pd_lower_90 <= predicted_pd <= pd_upper_90 on every row.

B - submission_B_trajectory.csv (the default-timing forecast)

This is not a single default number - it is the shape of how defaults
accumulate over the life of a loan, per origination cohort. It is what tells us
whether you modeled loan timing rather than just a yes/no classifier.

For each origination cohort week w (1-13) and loan age a weeks (1-13),
predict the cumulative fraction of your approved cohort-w loans that have
defaulted by day 7a.

• Rows: the full 13 x 13 = 169 grid. Use dataset/submission_B_template.csv:
  overwrite the three prediction columns; do not change the grid itself.

• Columns:

  column	meaning
  cohort_week	origination cohort 1-13 (from cohort_week_definitions.csv)
  loan_age_weeks	loan age 1-13 weeks after origination
  cumulative_default_rate	predicted cumulative default fraction by day 7a, in [0, 1]
  cdr_lower_90 / cdr_upper_90	your 90% interval bounds

• Rules: cdr_lower_90 <= cumulative_default_rate <= cdr_upper_90, and within
  each cohort the cumulative_default_rate must be non-decreasing as age
  increases (cumulative rates can only go up).

• Example row: cohort_week=3, loan_age_weeks=4, cumulative_default_rate=0.05
  means "5% of my approved week-3 loans defaulted within the first 28 days."

C - submission_C_counterfactuals.csv (the causal what-if)

dataset/intervention_queries.csv lists ~900 queries, each
(query_id, applicant_id, feature_name, intervention_value). For each query,
predict the applicant's PD if that one feature were set to that value by
intervention, holding everything else fixed - i.e. do(feature = value).

• Rows: one per query_id in intervention_queries.csv.

• Columns:

  column	meaning
  query_id	matches the queries file
  predicted_pd_cf	your post-intervention PD in [0, 1]
  pd_cf_lower_90 / pd_cf_upper_90	your 90% interval bounds

• Rule: pd_cf_lower_90 <= predicted_pd_cf <= pd_cf_upper_90.

D - submission_D_writeup.pdf (the technical writeup)

A short writeup defending your methodology. Human-reviewed; reviewers grade on
substance, not polish.

• Start from submission_D_writeup_template.md,
  fill in each section, then export to PDF named submission_D_writeup.pdf.
• Format (enforced): max 4 pages of body (excluding references), >=11pt
  font, >=0.75 inch margins. Content past page 4 is truncated for review.
• Required sections, in this order:
  1. Problem framing & assumptions violated
  2. Methodology
  3. Causal reasoning & counterfactual methodology
  4. Calibration & uncertainty quantification
  5. Limitations & what we'd do differently
• Section 3 (causal reasoning - distinguishing observational from interventional
  prediction, and how you'd defend your drivers to a regulator) carries the most
  weight.

How you're scored (high level)

The A/B/C files are scored automatically; the writeup is reviewed by humans:

• Portfolio profitability from your approve/decline decisions (A).
• Cohort-timing accuracy of your trajectory forecast (B).
• Calibration of your 90% intervals (A and B).
• Counterfactual accuracy of your what-if predictions (C).
• Technical writeup (D).

We deliberately do not publish the exact scoring weights or formulas.

Submission checklist

1. Register on the Google Form (Step 0) early - your private upload link is
   emailed to your team.

2. Unzip and explore the dataset (dataset/).

3. Build your four files with the exact names above. Use
   submission_B_template.csv for B and submission_D_writeup_template.md for D.

4. Put all four files flat in one folder (no subfolders).

5. Validate until it prints PASS:

   pip install -r requirements.txt
   python validate_submission.py path/to/your_submission_folder

6. Upload the four files to your team's private link from the email. Done.

Hard requirements (read this)

• Exact file names, exactly as listed above. A typo means your file is not
  found and not scored.
• Flat folder - the four files directly in one folder, no nesting.
• validate_submission.py must print PASS. Format, naming, ID-coverage,
  range, and monotonicity mismatches are caught here. Because scoring for some deliverables are objective and fully
  automated, a submission that does not pass the validator will be disqualified.