# Data & Methodology

This paper explains how the modeled results were produced and what they represent.
Sampo's rule is proprietary, and parts of it are patent pending; this paper describes the
data and the measurement, not the rule.

## The data is synthetic

Every number in these materials comes from a synthetic simulation. There is no
production bank data and no Fedwire transaction data, and the 125 "days" are
synthetic weekday labels, not an observation period. We built a synthetic book to
demonstrate the mechanism, run sensitivities, and design a bank-data pilot without
using any bank's confidential data. Real results will differ; a pilot on a bank's
own payments confirms the opportunity.

## How the synthetic book is built

The generator is referenced to published Fedwire structure along three axes:

- **Who pays whom.** A money-center core of five banks with a fixed directed graph
  of payment relationships, so the network structure stays stable across the
  history while the flows are redrawn each day.
- **How large payments are.** Large-value payment sizes are extreme right-skewed. We
  fit a lognormal to the admitted tail using published exceedance shares above
  $100m, $1bn and $2bn, then draw each in-scope payment from that distribution
  truncated to the configured $500m–$5.1bn range.
- **When they arrive.** A peak at the opening and a larger one in the late afternoon,
  matching the published intraday shape of a system with no coordination.

For each day, the generator draws new amounts, the times the banks send the payment intents,
and each payment's own time, from a documented, deterministic seed sequence. The result is a 125-day book
of 62,500 payments (500 a day across five banks), with values from ~$500m to
$5.1bn (median ~$972m, mean ~$1.08bn). The full parameters, sources and
regeneration procedure are documented in the internal data-generation methodology.

## How the saving is measured

For each day we compare the same payments without Sampo and with Sampo:

- **Without Sampo.** Every payment is submitted at its own time. We measure each
  bank's modeled in-scope peak net debit: its largest net debit at the end of any
  minute in the day, and zero if it never goes into net debit (technical name LNNCP).
- **With Sampo.** Sampo advises banks to submit some payments early, never later than
  their own time. Its no-worse-off guard aims to reduce the risk that a bank is worse
  off. It is not a guarantee.

A bank is worse off on a day when its realised peak net debit with Sampo is above its
peak without Sampo.

The headline **~49% reduction in modeled in-scope peak net debit** is in the cohort
monthly peak net debit: the sum of each bank's monthly peak, found separately, averaged
over the complete calendar months February and March of the synthetic history. It falls
by **49.3% (~$55.4bn)**. On an average-day basis the reduction is **59.0%**. Balances are sampled on the Fedwire convention: net all
settlements in an operating minute, then sample at minute end. An intraminute
convention, sampling after each settlement, is available for other rails.

## What the model leaves out

The metric is modeled in-scope peak net debit, not a bank's total intraday
liquidity requirement. The model omits opening account balances, below-threshold and
other out-of-scope flows, other rails, collateral and credit facilities, real RTGS
queue, rejection and finality behavior, cancellations and repairs, and
submission-to-settlement latency (the prototype uses one timestamp for both). It is suitable for demonstrating the
mechanism and designing a pilot, not for estimating a Fedwire tail, actual daylight
credit, bank P&L, regulatory capital or a required liquidity or HQLA buffer.

## The bank owns the economic translation

Any monetary value is a sensitivity the bank supplies, not a product claim:

`annual value = reduction in modeled in-scope peak × recognition factor × cost of liquidity`

The recognition factor (0–100%) is the share the bank's approved methodology
recognizes. The cost of liquidity is an annual rate that the bank supplies: what
holding one unit of liquidity for a year costs it. The annual value is computed for
each bank, at its own factor and rate, and summed over the banks. At 100% recognition
and a cost of liquidity of 1% a year, the monthly default implies an illustrative
~$554m annual value. This is not a funding-cost fact, a
Fedwire fee, or booked P&L.

## The current synthetic dataset

The current dashboard is a deterministic synthetic test of the mechanism.

| Property | Current dataset |
|---|---:|
| Banks | 5 pseudonymous banks |
| Synthetic history | 125 date labels |
| Payments | 62,500; exactly 500 each day |
| Payment values | ~$500m–$5.1bn; ~$972m median and ~$1.08bn mean |
| Mean selected value | ~$539bn per day; 11.74% of 2025 Fedwire daily value |
| Opening batch | 29.85% of payment intents sent at the opening |
| Comparison | Without Sampo: every payment submitted at its own time, which is unique for each payment |

The bank roster, complete directed pair graph and number of payments on
each pair stay fixed across the history. The generator redraws amounts, the times the
banks send payment intents, and own times for each day, using the documented,
deterministic seed sequence. Each payment's own time (`latest_ms`) equals its synthetic
settlement time (`historical_settlement_ms`). Sampo may advise a payment earlier than
its own time, never later.