Generate 10,000 realistic FX trades across 300 customers.

Real customer data is proprietary, so I generated a synthetic replica that mirrors it — 300 customers, 10,000 trades, 24-month window, weighted across 20 popular currency corridors (USD→EUR, GBP→USD, etc.) using approximate mid-market rates.

8% of customers are seeded as predictable (weekly, biweekly, or monthly cadence with ±1 day jitter). The other 92% trade at random intervals. The model doesn't know which is which — it has to find the predictable ones from the gap data alone.

For each customer, compute the gap between trades.

Sort every customer's trades by date, then use .diff() to get the days between each one and the next. That list of gaps per person is the raw signal the whole rest of the project sits on top of.

Then per-customer aggregates: mean gap, standard deviation, min, max, total revenue, tenure. One row per customer, ready for scoring.

One number per customer: how consistent is their rhythm?

Quick explainer — take a customer's gaps (say, 30, 29, 31, 30 days). Divide how much they wobble (standard deviation) by their average. Small wobble = low score = predictable. Statisticians call this ratio the Coefficient of Variation, or CV.

Compute CV for every customer with at least 3 trades — fewer than that and you can't tell a rhythm from a coincidence. The eligible pool: 278 customers. The CV distribution runs from a tight 0.01 (perfect clockwork) up to 1.55 (chaotic).

Three buckets. Two of them are worth emailing.

Apply the thresholds: CV ≤ 0.10 = Highly Predictable, 0.10–0.20 = Moderately Predictable, everything above = Irregular. That lands:

• 18 Highly Predictable (6.0%) — near-perfect cadence
• 9 Moderately Predictable (3.0%) — reliable but looser
• 251 Irregular (83.7%) — too noisy to trigger on
• 22 Insufficient Data (7.3%) — fewer than 3 trades

Total actionable: 27 customers, 9.0% of the base. And the seeded ground truth was 24 — the model recovered all of them and picked up 3 more irregulars who happened to fall into a tight rhythm by chance.

Do the score's winners actually look predictable?

A low CV should mean "their trades line up on a grid when you plot them." Bars are trades (height = send amount), red dashed line is the predicted next trade date. The pattern is visually clear without needing any stats background — useful when presenting to marketing and ops.

The #1 customer (CUST_0183, monthly cadence) has a CV of 0.02: 24 trades, mean gap 30.00 days. The next trade is predicted for Jan 4, 2026. Trigger the email on Jan 2.

Predictable customers earn 39.3% more revenue per head.

Being predictable is only useful if they also spend. Joining the revenue data onto the classification:

• Average total revenue: $2,339 (predictable) vs $1,679 (all 3+ trade customers) — +39.3%
• Average annual revenue: $1,231 vs $906
• Average trades per customer: 52 vs 36
• Interesting wrinkle: average send amount is actually lower ($2,620 vs $3,841) — they send smaller amounts, but much more often. Volume beats size.

Total annual revenue from predictable customers: $33,249, or 12.5% of the book. Retention on this slice is worth 3× its headcount share.

Cohort profile — over-indexed corridors point to acquisition.

Where do these predictable customers actually trade? Breaking their transactions down by corridor and comparing to the general population:

• AUD→GBP over-indexes 250% — predictable customers use this corridor 2.5× more than the base
• CHF→USD over-indexes 218%
• EUR→CHF over-indexes 211%
• GBP→JPY and USD→NZD both over-index around 160–167%

These are the hallmarks of expat living expenses and recurring remittance — not one-off trades. A Swiss worker paying UK bills. A retiree sending from Europe to Switzerland. That's the acquisition profile to target.

The output: a dated email trigger list.

This is what the marketing tool actually ingests. One row per predictable customer. For each: their predicted next trade date, their trigger date (2 days earlier), their cadence, their primary corridor, and their expected annual revenue. Sorted by trigger date so the campaign manager can see the week ahead.

Total annual revenue tied to this list: $33,249. Projected 3-year LTV for the moderate tier alone (the weekly shoppers) is $7,182 per customer — 3.7× the highly-predictable tier, because they trade so often. Counter-intuitive but real.