routing-and-cost-optimization

# Routing and Cost Optimization Solve payment routing optimization problems: find which card scheme or ACI minimizes/maximizes fees for a merchant over a time period. ## Question Types 1. **Card Scheme Routing**: "Which card scheme should merchant X steer traffic to in [month/year] to pay [minimum/maximum] fees?" 2. **ACI Optimization**: "For merchant X in [period], if we move fraudulent transactions to a different ACI, what is the preferred choice considering [lowest] fees?" **Answer format**: `{option}:{total_fee_rounded_to_2_decimals}` — e.g., `GlobalCard:142.37` or `B:102.59` ## Required Datasets | File | Purpose | |------|---------| | `payments.csv` | Transactions (merchant, card_scheme, aci, is_credit, eur_amount, issuing_country, acquirer_country, has_fraudulent_dispute, day_of_year, year) | | `fees.json` | Fee rules (1000 rules with matching criteria + fixed_amount + rate) | | `merchant_data.json` | Merchant profile (account_type, capture_delay, merchant_category_code, acquirer list) | Always read `manual.md` and `payments-readme.md` first for domain definitions. ## Fee Calculation Formula ``` fee = fixed_amount + rate * transaction_value / 10000 ``` When multiple fee rules match a transaction, **select the one producing the lowest fee**. ## CRITICAL: Acquirer Country for Intracountry **Always use the `acquirer_country` column from `payments.csv` directly** — one per transaction row. **NEVER** compute acquirer country by looking up the merchant's acquirer list in `acquirer_countries.csv`. The `acquirer_country` in `payments.csv` may differ significantly from the acquirer_countries.csv mapping. For example, a merchant's acquirer mapped to 'FR' in acquirer_countries.csv might have 'NL' in the payments.csv `acquirer_country` column for all its transactions. Using the wrong source causes incorrect intracountry flags and wrong fee totals. ```python # CORRECT intracountry = row['issuing_country'] == row['acquirer_country'] # from payments.csv # WRONG - do NOT do this acquirer_country = acq_countries[acq_countries['acquirer'] == merchant_acquirer]['country_code'].values[0] intracountry = row['issuing_country'] == acquirer_country # this may be wrong ``` ## Fee Rule Matching Logic Match ALL of these criteria. A rule applies when: | Field | Rule applies when | |-------|------------------| | `card_scheme` | Exactly matches transaction's card scheme | | `account_type` | `[]` (empty list) or `null` → all; else merchant's type must be in list | | `capture_delay` | `null` → all; else must match mapped merchant capture delay | | `merchant_category_code` | `[]` (empty list) or `null` → all; else merchant's MCC must be in list | | `is_credit` | `null` → all; else must match transaction's `is_credit` | | `aci` | `[]` (empty list) or `null` → all; else transaction/target ACI must be in list | | `intracountry` | `null` → all; `1.0`/`true` → domestic only; `0.0`/`false` → international only | | `monthly_fraud_level` | `null` → all; else merchant's period fraud % must fall in range | | `monthly_volume` | `null` → all; else merchant's period EUR volume must fall in range | **Empty list `[]` = applies to all** (same as null). This is the most common mistake to avoid. ## Capture Delay Mapping Merchant's `capture_delay` in `merchant_data.json` may be numeric (days) or a string bracket: - `"0"` or `"immediate"` → `immediate` - `"1"` or `"2"` (1-2 days) → `<3` - `"3"`, `"4"`, `"5"` (3-5 days) → `3-5` - `"6"`, `"7"`, or any number > 5 → `>5` - `"manual"` → `manual` ## Range Parsing **Monthly fraud level** (fraudulent_volume / total_volume × 100): ``` '<7.2%' → fraud_pct < 7.2 '7.2%-7.7%' → 7.2 <= fraud_pct <= 7.7 '7.7%-8.3%' → 7.7 <= fraud_pct <= 8.3 '>8.3%' → fraud_pct > 8.3 ``` **Monthly volume** (sum of eur_amount in EUR): ``` '<100k' → volume < 100_000 '100k-1m' → 100_000 <= volume <= 1_000_000 '1m-5m' → 1_000_000 < volume <= 5_000_000 '>5m' → volume > 5_000_000 ``` ## Step-by-Step Solution Process ### Step 1: Load data ```python import pandas as pd, json payments = pd.read_csv('.../payments.csv') fees = json.load(open('.../fees.json')) merchants = json.load(open('.../merchant_data.json')) ``` ### Step 2: Filter transactions for the period ```python # For a monthly question (e.g., September 2023): payments['date'] = pd.to_datetime(payments['year'].astype(str) + '-' + payments['day_of_year'].astype(str), format='%Y-%j') payments['month'] = payments['date'].dt.month subset = payments[(payments['merchant'] == merchant_name) & (payments['month'] == target_month)] # For an annual question (e.g., year 2023): subset = payments[(payments['merchant'] == merchant_name) & (payments['year'] == 2023)] ``` ### Step 3: Compute period metrics ```python total_volume = subset['eur_amount'].sum() fraud_volume = subset[subset['has_fraudulent_dispute'] == True]['eur_amount'].sum() fraud_pct = fraud_volume / total_volume * 100 ``` For **annual** questions, compute these metrics over the **entire year** (not per natural month). The fee rule fields `monthly_volume` and `monthly_fraud_level` are matched against the full-period aggregated volume and fraud rate for the queried time window. ### Step 4: Determine intracountry per transaction ```python # Use acquirer_country from payments.csv directly — do NOT look up from acquirer_countries.csv subset = subset.copy() subset['intracountry'] = subset['issuing_country'] == subset['acquirer_country'] ``` ### Step 5: Build the fee matching function ```python def matches_rule(rule, account_type, capture_delay, fraud_pct, volume, mcc, is_credit, aci, intracountry): # account_type at = rule.get('account_type') or [] if at and account_type not in at: return False # capture_delay cd = rule.get('capture_delay') if cd is not None and cd != capture_delay: return False # monthly_fraud_level mfl = rule.get('monthly_fraud_level') if mfl: if mfl == '<7.2%' and not fraud_pct < 7.2: return False elif mfl == '7.2%-7.7%' and not (7.2 <= fraud_pct <= 7.7): return False elif mfl == '7.7%-8.3%' and not (7.7 <= fraud_pct <= 8.3): return False elif mfl == '>8.3%' and not fraud_pct > 8.3: return False # monthly_volume mv = rule.get('monthly_volume') if mv: if mv == '<100k' and not volume < 100_000: return False elif mv == '100k-1m' and not (100_000 <= volume <= 1_000_000): return False elif mv == '1m-5m' and not (1_000_000 < volume <= 5_000_000): return False elif mv == '>5m' and not volume > 5_000_000: return False # merchant_category_code mccs = rule.get('merchant_category_code') or [] if mccs and mcc not in mccs: return False # is_credit ic = rule.get('is_credit') if ic is not None and ic != is_credit: return False # aci rule_aci = rule.get('aci') or [] if rule_aci and aci not in rule_aci: return False # intracountry intra = rule.get('intracountry') if intra is not None and bool(intra) != intracountry: return False return True def calc_fee(rule, amount): return rule['fixed_amount'] + rule['rate'] * amount / 10000 ``` ### Step 6: Calculate total fees per option **Card-scheme routing**: iterate over transactions already on each scheme. **ACI optimization**: iterate over only fraudulent transactions, testing each ACI (A–F). ```python # Card-scheme routing options = subset['card_scheme'].unique() results = {} for option in options: txns = subset[subset['card_scheme'] == option] total_fee = 0.0 no_rule_count = 0 for _, row in txns.iterrows(): matching = [r for r in fees if r['card_scheme'] == row['card_scheme'] and matches_rule(r, account_type, capture_delay, fraud_pct, volume, mcc, row['is_credit'], row['aci'], row['intracountry'])] if matching: total_fee += min(calc_fee(r, row['eur_amount']) for r in matching) else: no_rule_count += 1 # contributes 0 to total fee results[option] = {'total_fee': total_fee, 'no_rule': no_rule_count} # ACI optimization (fraudulent transactions only) fraud_txns = subset[subset['has_fraudulent_dispute'] == True] aci_options = ['A', 'B', 'C', 'D', 'E', 'F'] # G has no fee rules for aci_target in aci_options: total_fee = 0.0 no_rule_count = 0 for _, row in fraud_txns.iterrows(): matching = [r for r in fees if r['card_scheme'] == row['card_scheme'] and matches_rule(r, account_type, capture_delay, fraud_pct, volume, mcc, row['is_credit'], aci_target, row['intracountry'])] if matching: total_fee += min(calc_fee(r, row['eur_amount']) for r in matching) else: no_rule_count += 1 results[aci_target] = {'total_fee': total_fee, 'no_rule': no_rule_count} ``` ### Step 7: Select and format answer **For card-scheme questions**: Pick the scheme with min/max total fee across all schemes (transactions without matching rules contribute 0 — do NOT filter by coverage). **For ACI questions**: Only consider ACIs with **full coverage** (no transactions without a matching rule). Among fully-covered ACIs, pick the one with lowest total fee. ```python # Card scheme: no coverage filter best = min(results, key=lambda k: results[k]['total_fee']) # or max for maximum # ACI: filter to full coverage first full_coverage = {k: v for k, v in results.items() if v['no_rule'] == 0} best = min(full_coverage, key=lambda k: full_coverage[k]['total_fee']) fee = round(full_coverage[best]['total_fee'], 2) answer = f"{best}:{fee}" ``` ## Common Pitfalls - **Using acquirer_countries.csv for intracountry**: The `acquirer_country` in payments.csv is per-transaction and authoritative. Never override it with a lookup from acquirer_countries.csv — the two can differ for the same merchant. - **Treating `[]` as "no match"**: Empty list in `account_type`, `aci`, or `merchant_category_code` means "applies to all". - **Missing ACI coverage check**: For ACI optimization, ACIs with any uncovered transaction must be excluded even if their partial fee sum is lower. - **Applying ACI coverage filter to card-scheme routing**: For card-scheme routing, transactions with no matching rule just contribute 0 — do not exclude entire schemes. - **Not using minimum fee**: When multiple rules match, always use the rule producing the **lowest fee**. - **Capture delay numeric values**: `capture_delay: "7"` means 7 days → maps to `>5`. - **Annual period metrics**: For annual questions, compute `monthly_volume` and `monthly_fraud_level` categories using the **full year's aggregated totals**, not per-month breakdowns.