Fictional scenario. Synthetic data. Your notes stay on this browser; do not enter real customer information.
Week two is about numbers, and specifically about the number underneath the number. Every rate has a denominator, and most arguments on a risk desk are arguments about which one.
Day 08 — Compared with what.
"Declines are up 40 per cent" means nothing until you know against what. Yesterday? The same weekday last month? The same week last year, when the merchant ran the same promotion? A spike is only a spike relative to a baseline, and a baseline is a choice.
Today: take one metric your team watches and write down what it is compared against. If the answer is "the previous period", write down what the previous period was and whether it was typical. Then find the same metric for the same calendar week last year and compare that too.
Read: Compared with what, exactly.
Day 09 — Find the denominator.
A rate without a denominator is a rumour. "Twelve chargebacks" is a count; whether it matters depends on whether it is twelve of four hundred or twelve of forty thousand, and on whether the four hundred were counted in the same month, the previous month, or settled transactions only.
Today: for the metric you chose yesterday, write down its denominator exactly as the system computes it. Not what you assume. Find the query or the definition. If nobody can show you the definition, that is today's finding, and it goes in the second notebook.
Read: the decision study The missing denominator.
Day 10 — Precision is a ratio of your own work.
Of everything your rule flagged last month, how much was actually the thing it was looking for? That is precision, and it is the metric that tells you how much of your colleagues' time a rule spends. Recall is the other half, and it is harder to measure because the misses are, by definition, not in your queue.
Today: pick one rule. Count what it flagged in the last 30 days and how many of those were confirmed. Write the ratio down. Then write down how you would ever know what it missed.
Read: the Volume 01 lesson Measuring errors and look up precision.
Day 11 — Two ratios, two denominators.
The card schemes watch merchants through ratios, and the two main ones do not use the same denominator or the same numerator. One, as summarised by the acquirer guidance we cite, counts fraud reports and disputes against settled card-absent transactions in the month. The other, as summarised by processor documentation, sets this month's chargebacks against last month's transaction count. The same incident lands differently in each, and in a different month.
Today: find out which ratio your acquirer reports to you, for which merchants, and when. Write down the formula as they state it. If you have a merchant whose volume fell sharply last month, work out what that does to this month's ratio.
Read: Two ratios, two denominators, one portfolio.
Day 12 — A threshold is a price.
Every threshold is a decision about which error you prefer. Lower it and you catch more and interrupt more. Raise it and you miss more and interrupt fewer. The number is not a measure of risk. It is the price you have agreed to pay, and someone should have agreed to it.
Today: take one threshold in one rule and write a four-cell table: caught, missed, interrupted, left alone, with a rough cost in each cell. Synthetic numbers are fine. Then ask whether the threshold you have is the one that table would choose.
Read: The threshold is a price, not a score.
Day 13 — A rule is a claim.
A rule says: when these conditions hold, this is probably happening, and this is the right response. That is a claim with three parts, and each can be wrong separately. A rule nobody can state as a sentence is a rule nobody can review.
Today: take the rule from day ten and write it as one sentence. "When a card from this BIN range makes more than three gift-card attempts in an hour on the web channel, it is probably enumeration, and we step up." If you cannot fill in the middle clause, the rule is a reflex, not a claim.
Read: A rule is a claim. Write it like one.
Day 14 — Check the rule still fires.
Rules die quietly. An export field is renamed, a channel is migrated, a merchant changes its product code, and the rule goes on existing without ever matching anything again. Nobody notices, because a rule that fires nothing produces no queue.
Today: list your ten most important rules and the date each last fired. Any rule silent for longer than its usual gap is a question for today, not next week. Then add a line to the second notebook: which of your habits from week one would have caught this sooner?
Read: the case Silent rule, which is exactly this, at Northbank Pay, three weeks late.
Fictional learning material. Any numerical examples are synthetic. Use only authorised, appropriately minimised records for workplace exercises; do not upload customer information to this learning site.