Credit Limit Adjustment

One account object carries everything: payment_history with on-time percentages and late payment counts, utilization with current and six-month figures, financial_behavior covering inquiries, tenure, and income verification, plus the current_credit_limit the final step adjusts.

Repayment quality is read from two fields at once under a first hit policy: an on_time_payment_percentage of 95 or better with zero late_payments_last_12_months rates 'excellent' and +30, 90 with at most 1 late payment rates 'good' and +20, 80 with up to 2 is score-neutral 'average', 70 with up to 3 is 'below_average' at -10, and the fallback is 'poor' at -30.

Giving payment history the widest score swing in the graph (a 60-point spread) parallels how consumer credit models weight it; in the FICO model payment history is the single largest component at about 35 percent. Requiring both a high on-time rate and few recent lates catches accounts whose lifetime average looks fine but which have deteriorated lately.

Usage patterns pair the snapshot with the trend: a current_percentage under 30 alongside a six-month average under 50 rates 'low' and +25, the 30 to 50 band with a 50 to 70 average is 'moderate' at +15, and successively heavier bands step through 'high' (0), 'very_high' (-15), and the 'excessive' catch-all at -25.

The 30 percent line is the utilization guideline most credit scoring commentary converges on, below which revolving usage is considered healthy. Testing average_last_6_months as well as the current reading separates a customer who briefly ran up a balance from one who is persistently maxed out, and only the latter should cost limit headroom.

Three signals combine into one rating: fewer than 1 credit inquiry in 12 months, 3 or more years_with_institution, and income_verified true earns 'excellent' at +25; slightly looser combinations grade 'good' (+15) and 'average' (+5), while fewer than 4 inquiries with under a year of tenure and unverified income falls to 'below_average' at -10, and the catch-all is 'poor' at -20.

Inquiry counts proxy for credit-seeking pressure, and long tenure gives the bank observed behavior it can trust more than bureau data alone. Verified income matters beyond risk appetite: US card issuers must consider a customer's ability to pay before raising a limit under the CARD Act, so an increase recommendation built on unverified income would be hard to action.

A single sum produces total_adjustment_score by adding evaluation.payment_history_score, evaluation.utilization_score, and evaluation.financial_behavior_score. The unweighted addition keeps each table's internal point scale as its implicit weight, and the result is the only input the recommendation ladder needs.

Eleven descending bands convert total_adjustment_score into recommendation.percent_change with a first hit policy: above 50 grants a 30 percent increase, then 25, 20, 15, 10, and 5 percent steps down to scores just above 0, while negative territory maps to a hold at 0 and cuts of -5, -10, -15, and finally -20 percent below -40. Every row also emits a plain-language justification, from 'Significant increase recommended' down to 'Major decrease recommended due to severe risk factors'.

Graduated moves in the 5 to 30 percent range match how issuers actually manage lines, since small periodic adjustments avoid sudden exposure jumps and customer shocks. The asymmetry is notable: the best score earns +30 percent but the worst only cuts -20, a common posture because aggressive decreases can trigger utilization spikes and attrition among salvageable accounts, and each row's justification text doubles as the explanation a customer or reviewer would see.

Dollar figures come last: recommendation.amount_change rounds account.current_credit_limit times percent_change over 100, and recommendation.new_credit_limit adds that change back to the current limit. Emitting concrete numbers means the account system can apply the decision without re-deriving the percentage math.