## SOFR data republication notice The SOFR data is subject to the [Terms of Use posted at newyorkfed.org](https://www.newyorkfed.org/privacy/termsofuse). The New York Fed is not responsible for publication of the SOFR data by OpenProse, does not sanction or endorse any particular republication, and has no liability for your use. OpenProse is not affiliated with the New York Fed. The New York Fed does not sanction, endorse, or recommend any products or services offered by OpenProse. > Public demo input, adapted from the original developer brief at source revision `01ae3bdac6f61ecbabcd72e0116ebb2860e93157`. This is evidence of the developer’s model and choices, not a completed methodology document. The original numerical run is historical; this commission does not rerun it. References below to `out/` mean the public `evidence/` directory, and script names mean `model-source/`. Raw archives and copyrighted papers are not republished. The last section records the original brief’s missing institutional inputs; the current commission supplies fictional scenario facts in `institution.md`, and any remaining actual institutional facts remain unavailable. Source-specific assertions are attributable to this brief unless independently checked against the cited source. # Developer evidence: USD SOFR OIS discount curve This is the model-specific input to the composed run. It states what the model is, what the developer chose and why, and where every fact comes from. It cites no regulation (the base layer does that) and sets no document conventions (the house layer does that). Source keys resolve in the public [source register](https://openprose.ai/demos/sofr/v1/sources.md). Every number below is copied from the worked-example outputs named beside it. ## Model **Name.** USD SOFR OIS discount curve. **What it produces.** A discount factor `Z(0, t)` for every date `t` from the as-of date to 30 years, and beyond by a stated extrapolation: the present value of one US dollar paid on `t`. The curve is bootstrapped from the last published SOFR fixing and from USD SOFR overnight-indexed swap (OIS) par rates, and interpolated between the dates those instruments determine. **Purpose.** Two uses of one curve, because in a single-curve SOFR setting the curve that discounts is also the curve that projects SOFR: 1. Discounting of USD cash flows whose collateral or funding rate is SOFR. For cleared USD swaps at CME this is the planned convention: CME's August 2020 methodology paper sets the transition of discounting and price alignment for cleared USD swaps from EFFR to SOFR for the close of business on October 16, 2020, "pending regulatory approval" ([CME-SOFR-2020] §1, p. 3), with the entry "Market begins to quote swaps based on SOFR discounting" for October 19, 2020 in its timeline ([CME-SOFR-2020] §1.3, p. 5). 2. Projection of compounded SOFR for the floating legs of SOFR swaps: under a single curve the compounded overnight rate over a period is fixed by the discount factors at its ends (see the par relation below). **Not for.** Projection of any rate other than SOFR; discounting of cash flows collateralised at another rate or in another currency; anything requiring a tenor basis, a cross-currency basis, or a term-SOFR curve; anything that needs the curve to step on central-bank meeting dates or across the turn of the year. See exclusions. ## Scope and exclusions - One currency (USD), one curve, one as-of date in the worked example. - Out of scope, explicitly: multi-curve construction (separate projection and discount curves) and the tenor and cross-currency basis machinery described in [AB2013]; collateral-choice or CSA-specific discounting; SOFR futures and their convexity; term SOFR; turn-of-year and central-bank-meeting-date steps. ## Instruments and conventions | Item | Value | Source | |------|-------|--------| | Overnight anchor | Two deposits at the last SOFR fixing known at the as-of close, 3.85 % for 2026-09-17 (published 2026-09-18): overnight, 2026-09-18 to 2026-09-21, and tom-next, 2026-09-21 to spot, 2026-09-22 | [NYFed-SOFR], `out/sofr_fixing.json` | | OIS quotes | Close-adjusted median of the day's par prints per standard tenor (choice 2) | [DTCC-PPD], `derive_quotes.py`, `out/quotes.md` | | Tenors used | 1M, 2M, 3M, 4M, 6M, 1Y, 18M, 2Y, 3Y, 4Y, 5Y, 6Y, 7Y, 8Y, 9Y, 10Y, 12Y, 15Y, 20Y, 30Y (20 OIS plus the two deposits) | `out/quotes.csv` | | Tenors reported but not used | 9M (4 prints), 25Y (2 prints) | `out/quotes.csv` | | Spot lag | 2 business days (effective date 2026-09-22 for trade date 2026-09-18) | [DTCC-PPD] effective-date field | | Fixed leg | Annual payments (one payment for tenors of a year or less), Act/360 | [DTCC-PPD] fields "Fixed rate day count convention-leg 1" (code A004, defined as "Actual360" in [CFTC-TS] Appendix C, p. 55) and "Fixed rate payment frequency period-Leg 1" (YEAR, or EXPI for one payment at expiry); code counts in `out/quotes.md` | | Floating leg | SOFR compounded daily in arrears, same schedule as the fixed leg, Act/360 | [ARRC2021] p. 31 and p. 35 for the ISDA compound convention with N = 360; [DTCC-PPD] "Floating rate payment frequency period-Leg 2"; underlier name "USD-SOFR-OIS Compound" | | Accrual dates | Rolled back from maturity in 12-month steps, Modified Following on the SOFR calendar | QuantLib `OISRateHelper` defaults (rule Backward, convention ModifiedFollowing), [QuantLib] | | Payment dates | Accrual end plus 2 business days, Following | [ARRC2021] p. 17 for the two-day delay ("2 days after the end of the interest period"); business days and Following are QuantLib `OISRateHelper` settings, [QuantLib] | | Calendar | SOFR publication calendar as implemented in QuantLib `UnitedStates(SOFR)` | [QuantLib] | | Node dates | The last payment date of each instrument (so the 2Y node is 2028-09-26: maturity 2028-09-22 plus the payment lag). The instrument's price fixes the curve only to its last accrual end; the few days from there to the node are interpolated, which is why the methods' 1M node zero rates differ (3.9407 % and 3.9477 %) | `out/nodes.csv`, `out/results.md` | | Time axis and zero rates | Act/365 Fixed year fractions from the as-of date (a developer choice, see smaller choices), continuous compounding | [HW2008] §1 for continuously compounded rates | | Implementations | Method A: QuantLib 1.43 `PiecewiseLogLinearDiscount`. Method B: `hagan_west.py`, written from [HW2008] §6. Instrument dates for both: QuantLib `DepositRateHelper`, `OISRateHelper` | `out/run_info.json` | **Par relation.** For an OIS starting at spot with periods `k` accruing from `s_k` to `e_k` (Act/360 fraction `α_k`) and paying at `p_k`, the par rate is `R = Σ_k (Z(s_k)/Z(e_k) − 1) Z(p_k) / Σ_k α_k Z(p_k)`. Each floating payment is the compounded SOFR over its period; under one curve the daily compounding of the curve's own overnight forwards telescopes to `Z(s_k)/Z(e_k) − 1`, and the payment is discounted from its pay date. This is [HW2006] p. 92 eq. (1) generalised to a spot start and a payment lag; with no lag and a start at the as-of date it reduces to eq. (1). The formula is implemented in `hagan_west.ois_par_rate` and agrees with QuantLib's OIS valuation to 2.1e-13 bp on every input (`out/verify.md`). ## The developer's choices (each with a reason, alternatives, and a limitation) 1. **Interpolation: monotone convex on forward rates, as published** — the unameliorated method of [HW2008] §6 with the positivity collar of §6.1, implemented in `hagan_west.py` from equations (22)–(34), and solved for all instruments at once so that every one reprices. Log-linear interpolation on discount factors ([HW2008] §4.4, the "raw" method) is built from the same instruments as method A. *Reason.* The curve's second use is projection, and the case for B rests on the shape of the forwards it projects, with a smaller gain in pricing short broken dates. - Forward shape: from the as-of date to the 20Y node, the raw method's overnight forward is a staircase that steps at every node, by up to 59.71 bp (18 steps above 2 bp to the 30Y node); method B's forward is continuous, with a largest day-to-day move of 1.68 bp and none above 2 bp (`out/results.md`). The developer cites [HW2008] p. 72 as motivation for continuous forwards. Over any five business days, A's forward moves by up to 59.73 bp and B's by up to 5.34 bp (`out/results.md`), a measure that does not depend on a daily threshold. B's forward is not smooth everywhere: it has flat stretches where the paper's sectors (ii) and (iii) insert them (for example from 3Y to 5Y), and just after the 3Y node a short ramp takes it from 4.4705 % to 4.4317 % in three business days, close to a step (`out/results.md`). - Pricing trades outside the input set (`out/verify.md`, paired table): on the 41 spot-starting par prints under 1Y at maturities that are not inputs, B's par rate is closer on 30 and A's on 11, by a median 0.28 bp, and 34 against 23 are within 1 bp. The prints sit at 22 distinct expiration dates; taking one observation per date, B is closer at 16 and A at 6 (sign test p = 0.052), and on raw, unadjusted rates B is closer on 35 of the 41. The advantage is concentrated between 6M and 1Y (B closer on 24 of 28), the interval where the 9M tenor had too few prints to be an input: B's curvature fills a gap in the inputs better than A's flat segment. The 6M–1Y grouping was chosen after seeing the data, so it is exploratory. Suggestive rather than conclusive, and specific to where inputs are missing. In every other maturity bucket with more than one print, the median difference in absolute residual lies between −0.041 and +0.082 bp, and which method is closer more often varies by bucket (A in two, B in four): the methods cannot be told apart economically there. - [HW2008] p. 79 recommends this method, "in particular, the unameliorated version", and it is the version built. *Alternatives considered.* The raw method (kept as method A for a transparent comparison, citing [HW2008] p. 73). QuantLib's `ConvexMonotone` bootstrap: at the paper's setting (quadraticity 0, monotonicity 1) its iterative bootstrap, which solves one node at a time, fails on these instruments ("root not bracketed" at the 3M instrument), a limit of that solver rather than of the method, which `hagan_west.py` solves for all nodes at once; its default (0.3, 0.7) is a smoothed blend with a quadratic, not the paper's method; built as a variant, it differs from B by at most 8.0e-5 in discount factor on the monthly grid and has a largest daily forward jump of 20.56 bp (`out/results.md`). Cubic splines (rejected on [HW2008] Table 1: forwards not guaranteed positive, poor locality; not built). *Limitation admitted.* Method B is less local. A +1 bp bump to the 5Y quote moves overnight forwards outside 4Y–6Y by up to 4.15 bp under B against 0.03 bp under A, and 11.9 % of the absolute PV sensitivity of a 4.5Y cash flow sits outside the 4Y and 5Y inputs under B against 5.9 % under A (`out/results.md`). For a book hedged instrument by instrument, A's risk is cleaner. Where inputs are far apart the methods differ most: the largest discount-factor difference on the monthly grid is 2.27e-3 on 2050-04-18, between the 20Y and 30Y nodes, USD 227,458 of present value per USD 100 million paid on that date, or 2.9 bp of zero rate on that date (4.6715 % under A, 4.7006 % under B, `out/df_grid_monthly.csv`); this is a disagreement between the methods, not a gain from either, and it is the worst single date, not the effect on a diversified book. Beyond 30Y the gap widens: discount factors at 40Y are 0.16426043 (A) and 0.16853911 (B) (`out/results.md`). The long end is thin: the 20Y node rests on 6 prints with USD 0.04 billion of notional (`out/quotes.md`). The 20Y–30Y interval's discrete forward is low, and it shapes B's forward from about 18Y in two ways: through eq. (22) it pulls down B's node forward at 20Y, which bends the 15Y–20Y interval; and eq. (24) sets the forward's slope to zero at the 30Y node ([HW2008] p. 75), which places the end forward below the interval's discrete forward by half the 20Y node forward's excess over it. In overnight Act/360 terms (`out/results.md`), B's forward falls from 4.6127 % at the 20Y node to 3.9957 % at the 30Y node, 20.6 bp below B's own average over that interval (4.2013 %) and 67 bp below the 30Y par rate (4.6637 %, `out/quotes.md`); A holds 4.2248 % across the interval. That shape comes from the end rule on sparse inputs, not from prices. The positivity collar did not bind on the as-of date. 2. **Quotes: the close-adjusted median of the day's par prints.** Swap rates rose by about 7 bp during 2026-09-18 (`out/intraday_path.csv`: −6.8 bp at 00h, −2.8 bp at 08h, 0 at 16h New York time, for tenors of 2Y and longer), so an all-day median mixes prints struck at different market levels. Each print is moved to the close along a common intraday path: the fixed rate is modelled as its tenor's close level plus a tenor-specific multiple (the loading) of the path; the path is fitted by median regression on tenors of 2Y and longer (loading 1 by definition), and each shorter tenor's loading is fitted where it has at least 8 prints in at least 5 distinct hours (1Y 0.49, 18M 0.86, 6M 0.14, 3M 0.08) and interpolated otherwise; loadings are held to the range 0 to 1, so 1M's fitted −0.04 is used as 0 (`out/intraday_loadings.csv`). The close is the last traded hour, prints executed 16:00–16:59 New York time, which are not adjusted. The quote is the median of the adjusted prints. *Reason.* The only free, per-date, citable source of USD SOFR OIS levels is the CFTC-mandated public dissemination [DTCC-PPD]; the adjustment removes the day's drift, which was larger than the differences between neighbouring tenors. For tenors of 1Y and longer, the interquartile range of prints within a tenor falls from 1.7–5.8 bp raw to 0.1–0.8 bp adjusted (`out/quotes.md`); unadjusted, the 4Y median sat 5.9 bp below the 3Y median, an inversion made by the timing of the prints. The fit is a median regression, the regression counterpart of the median used for each quote. *Alternatives.* The all-day median (reported beside each quote in `out/quotes.md`); the median of prints in the last hours only; a vendor or clearing-house closing curve (not used: the worked example rests only on data vendored in this repository). The narrowing of the spread is partly in-sample by construction, since the path is fitted to the same prints. *Evidence that the adjustment is not fitting the quotes to themselves.* Refitting the path without each path tenor's own prints moves that tenor's quote by at most 0.165 bp (`out/quotes.md`, leave-one-tenor-out table). *Limitation.* The loading model assumes one intraday factor; tenors between fitted loadings are interpolated; a forward-starting print takes the loading of its time to expiration, so a one-year swap starting in one year is treated like a two-year swap. The prints priced outside the input set (choice 1) are adjusted with the same path, so they are outside the bootstrap's inputs but not outside the adjustment's. The adjusted prints are still executions, not a fixing. In a deployment this step is replaced by the institution's market data (`[BANK-SUPPLIED: Data lineage to internal market data feeds]`). 3. **Print selection.** New trades (action NEWT, event TRAD) executed on the as-of date in New York time, effective at spot, not flagged as packages, with an Act/360 fixed leg, annual or single-payment schedules on both legs, no upfront payment, and a maturity within four calendar days of a standard tenor. A tenor enters the bootstrap with at least 5 prints and a close-adjusted interquartile range of at most 2 bp. *Reason.* Each filter keeps prints whose fixed rate is a par rate for the instrument the curve prices: a leg reported inside a package may carry a rate that reflects the package as a whole, so the developer does not rely on it as a stand-alone par rate; a print with an upfront payment trades at a standard coupon (on the as-of date 26 % of the 1,661 such prints carry a rate on a round 5 bp step, against 3 % of the 2,390 others, `out/quotes.md`); monthly-paying prints have a different schedule. The trade date is taken in New York time because the report's day is cut in UTC ("all transactions published for that report day based on UTC", [DTCC-PPD-Guide] p. 7) and the as-of close is New York's. Five prints with a sub-basis-point spread fix a median to well under 1 bp. *Alternatives.* Include packages; convert rather than drop other schedules; a count floor of 10 (which on the as-of date would drop 1M, 2M, 18M, 8Y, 9Y, 12Y, 15Y and 20Y). *Limitation.* 9M and 25Y are dropped; the 20Y-to-30Y gap is the widest on the curve. Code A004 is Act/360 by the CFTC specification ([CFTC-TS] Appendix C, p. 55). Counts (`out/quotes.md`): the report has 23,563 rows, of which 4,258 are new USD SOFR OIS trades; 663 spot-starting prints pass the filters and 529 match a standard tenor; across all effective dates, 76 monthly-paying prints and 309 prints with an upfront payment are excluded. The filters do not make every print a par print: 20 of the 529 sit 5.7 to 154.3 bp from their tenor's quote after adjustment (off-market or mis-reported). They are kept and disclosed; removing them would move no quote by more than 0.124 bp. 4. **Anchor: overnight and tom-next deposits at the last known fixing.** *Reason.* Pins the curve from the as-of date to spot with a published rate, so no interpolation is needed before the first swap starts. *Alternatives.* Start the curve at spot; use the fixing dated 2026-09-18 (published on the next business day, Monday 2026-09-21, so not known at the as-of close; same value, 3.85 %, in the snapshot). *Limitation.* Assumes the last fixing holds from the as-of date to spot, four calendar days and two business days. On this date it did: the fixings later published for 2026-09-18 and 2026-09-21 are both 3.85 % (`out/verify.md`); an anchor 1 bp higher would move discount factors by at most 1.1e-6 (A) and 1.3e-6 (B) on the monthly grid (`out/results.md`). 5. **OIS only; no SOFR futures.** *Reason.* OIS par rates alone fix the curve to 30Y, and every input then comes from the one vendored report; futures would need a convexity adjustment, which "will depend on some time and volatility parameters" ([HW2006] p. 92), and settlement data this example does not vendor. *Alternatives.* Futures at the short end, as in CME's SOFR curve construction ("SR3/SR1 Futures (with convexity adjustments) at the short-end", [CME-SOFR-2020] §2.1, p. 7). *Limitation.* Front-end detail is limited to the OIS tenors with enough prints. 6. **Single curve.** *Reason.* Scope; [AB2013] describes the multi-curve setting. *Alternatives.* A multi-curve framework. *Limitation.* The curve projects SOFR only. 7. **Extrapolation past the 30Y node: the instantaneous forward held at its value at the last node.** *Reason.* [HW2006] p. 109: "Beyond tn, we should use flat extrapolation: f(t) = f(tn) for all t > tn." The forward is then continuous at the node (jump 0.00 bp for both methods). *What is measured.* The held level is 3.9957 % under B and 4.2248 % under A (A holds its last interval's forward); discount factors at 40Y are 0.16853911 (B) and 0.16426043 (A) (`out/results.md`). *Alternatives.* A long-rate assumption; an institutional extrapolation policy. *Limitation.* B's held level comes from the paper's end condition (eq. (24)), which extends the slope of the 20Y–30Y forwards; beyond 30Y the curve rests on no quote. `[BANK-SUPPLIED: Extrapolation policy beyond the last node]`. ## Smaller choices (reason · alternatives · limitation) Construction - **Positivity collar on B** ([HW2008] §6.1 step (3)). Reason: guarantees a positive forward whenever the discrete forwards are positive. Alternatives: no collar (built as a variant; identical here). Limitation: where it binds, the forward is flattened near a node and the curve departs from the uncollared interpolant; it did not bind on the as-of date. - **B solved as one simultaneous system** (SciPy hybrid Powell). Reason: B's interpolation on an interval depends on neighbouring intervals, so a node-by-node solve can fail (as QuantLib's does). Alternatives: iterate node-by-node solves to a fixed point ([HW2006] p. 93). Limitation: convergence is not proven in general; the solve fails loudly if any par-rate residual exceeds 1e-11, and the repricing table shows the residuals. - **Method A as the comparison method.** Reason: [HW2008] p. 73 recommends comparing a sophisticated method with the raw method to find mistakes. Alternatives: QuantLib's ConvexMonotone default (also reported); a cubic spline (not built). Limitation: A is a comparison, not a best competitor; beating A is necessary, not sufficient. - **One set of instrument dates for both methods, from QuantLib.** Reason: any difference between the curves is then due to interpolation alone. Alternatives: separate schedule code for B. Limitation: a schedule error in QuantLib would affect both curves alike; the separately written pricer agrees with QuantLib's to 2.1e-13 bp, which checks the valuation, not the dates. - **Node at each instrument's last payment date.** Reason: QuantLib's default pillar for OIS (the last relevant date), kept so that both methods share one node set. Alternatives: node at the last accrual end, the latest date whose discount factor the par rate pins (for a one-period swap the pay-date factor cancels out of the par relation). Limitation: the few days between last accrual end and payment are interpolated (the 1M node difference in the conventions table). - **Act/365 Fixed time axis, continuous compounding.** Reason: continuous compounding is how [HW2008] §1 writes the curve; Act/365 Fixed is a common neutral time measure and does not affect discount factors. Alternatives: Act/360 time. Limitation: zero rates and the forward integral are stated on this axis; converting to another convention changes quoted zero rates, not prices. Close adjustment - **Median regression.** Reason: the regression counterpart of the median used for each quote, so the path and the quotes treat off-market prints alike. Alternatives: least squares; least squares after trimming. Limitation: not compared with least squares here. - **Path fitted on tenors of 2Y and longer, loading 1.** Reason: their fitted loadings, when each is fitted separately, lie between 0.97 and 1.08 (`out/intraday_loadings.csv` records the fits used), so one common path describes them; shorter tenors moved less. Alternatives: a separate loading for every tenor. Limitation: small differences in how long tenors moved are ignored. - **Loading fitted only with at least 8 prints in at least 5 hours.** Reason: fewer prints or hours cannot separate a tenor's level from its intraday move. Alternatives: fit every tenor. Limitation: 2M's loading is interpolated. - **Loadings held to 0–1.** Reason: a negative loading or one above the long tenors' has no economic reading for this day and would come from noise. Alternatives: unconstrained. Limitation: 1M's fitted −0.04 is used as 0. - **The close is the last traded hour, 16:00–16:59 New York.** Reason: the latest hour with prints at the path tenors. Alternatives: a fixed clock time such as 15:00. Limitation: the curve reflects that hour, not a published closing fixing. - **Forward-starting prints take the loading of their time to expiration.** Reason: a simple rule that needs no forward-start model. Alternatives: loading by the start and length of the swap. Limitation: a one-year swap starting in one year is treated like a two-year swap. Print selection - **Maturity within four calendar days of a standard tenor.** Reason: absorbs business-day adjustment of maturity dates over a weekend and a holiday. Alternatives: exact-date matching; recomputing the adjusted date. Limitation: a print with an unusual stub near a tenor could be matched to it. - **Trade date in New York time.** Reason: as choice 3. Alternatives: the report's UTC day. Limitation: a print in the New York evening after the close is still counted in the as-of day. - **Close-adjusted interquartile range of at most 2 bp for a tenor to enter.** Reason: with sub-basis-point adjusted spreads at every liquid tenor, a wider spread signals a quote that the median cannot fix. Alternatives: no dispersion test. Limitation: the dispersion screen excluded no tenor on the as-of date; 9M (4 prints) and 25Y (2 prints) failed the five-print count floor instead (`out/quotes.md`, 9M and 25Y rows; `derive_quotes.py`, tenor eligibility condition). - **Off-market prints kept in the medians.** Reason: removing prints needs a rule that could itself be tuned; a median already limits their effect, and removing them would move no quote by more than 0.124 bp. Alternatives: remove prints beyond 5 bp. Limitation: at a thin tenor one off-market print shifts the median's position more. - **Only prints coded A004 (Act/360) are used.** Reason: the curve's instruments accrue Act/360. Alternatives: convert prints with other day-count codes. Limitation: the 47 of 4,258 prints with other codes are excluded rather than converted. Tests - **Where B's bump response spills over.** Bumping the 5Y quote changes the 4Y–5Y discrete forward, which moves B's node forward at 4Y (eq. (22)) and therefore the shape of B's 3Y–4Y interval: that is why B's largest change outside 4Y–6Y (4.15 bp) sits just after the 3Y node. - **Locality: +1 bp on the 5Y quote, window 4Y–6Y; ladder for a 4.5Y cash flow.** Reason: 5Y is mid-curve and has 76 prints; it is not the most-traded path tenor (10Y has 99; `worked-example/out/quotes.md`, rows 5Y and 10Y). A 1 bp perturbation follows [HW2008] p. 78; the 4.5Y cash flow follows the hedge example of [HW2006] p. 126. Alternatives: bump every input and report the stability norm of [HW2006] pp. 124–125. Limitation: one bump and one cash flow, not a norm over all inputs. - **Pricing outside the input set at close-adjusted rates, median absolute residual at most 1 bp.** Reason: 1 bp is above the close-adjusted interquartile range of the prints at every used tenor (0.1–0.8 bp, `out/quotes.md`), so a curve within it prices unseen prints about as closely as the prints at one tenor agree with one another. Alternatives: a tolerance per maturity bucket. Limitation: the tolerance is a judgment, not derived. - **Late-print row.** Reason: a result that does not use the intraday model. Alternatives: the last hour only (fewer prints); the last three hours (larger adjustments). Limitation: 49 prints only. - **Paired sign test by maturity bucket.** Reason: compares the two curves on the same prints by counting which is closer, and needs no model of the residuals' size. Alternatives: a paired t-test; a Wilcoxon test. Limitation: counts which method is closer, not by how much (the median difference is reported beside it). - **Money measure: largest discount-factor difference on a monthly grid to 30Y, as the present value of USD 100 million.** Reason: states what the choice is worth in the unit a desk uses. Alternatives: a portfolio's PV difference (no portfolio is public). Limitation: a single cash flow on the worst date overstates a diversified book. - **Leave-one-tenor-out check.** Reason: tests whether a tenor's own prints drive its adjustment. Alternatives: cross-validation over hours. Limitation: covers the path tenors, not the fitted loadings. - **Check tolerances (`out/verify.md`).** Reason: two implementations of the same arithmetic should agree to rounding (1e-9 and below); the round trip to 1e-6 bp; the paper examples to half the printed precision; condition (i) to 1e-7 in rate, because the midpoint rule with 2,000 steps per interval approximates each interval's average (the largest error found is 6.3e-10); continuity to 1e-6 in rate, because the forward is evaluated 1e-9 years either side of each node, where a continuous piecewise quadratic moves far less; method B through QuantLib to 1e-4 bp, because QuantLib values B from daily discount factors with log-linear interpolation between business days, an approximation (the error found is 1.3e-11 bp). Alternatives: tighter or looser. Limitation: set by the developer. - **Paper example p. 102 uses the capitalisation factors.** Reason: the printed 4Y yield disagrees with its own factor and forward; the factor is internally consistent. Alternatives: use the printed yield (the check would then fail on the paper's misprint). Limitation: relies on the page image reading. - **Convention sensitivities: payment lag 0, quarterly payments.** Reason: the two conventions the prints do not fully fix. Alternatives: roll convention, calendar. Limitation: not measured for the others. ## Assumptions - The DTCC report codes the day count as the CFTC specification defines it (A004 is "Actual360", [CFTC-TS] Appendix C, p. 55). A004 is the code on 4,211 of the day's 4,258 USD SOFR OIS prints (`out/quotes.md`). The day count matters: were the fixed legs Act/365 Fixed, discount factors would move by up to 5.0e-3 on the monthly grid (`out/results.md`). - A payment lag of two business days on both legs ([ARRC2021] p. 17). Sensitivity: a lag of zero changes discount factors by at most 1.3e-5 (A) and 4.8e-6 (B) on the monthly grid (`out/results.md`). - Annual payments on both legs. Sensitivity: building the same par rates as quarterly-paying swaps changes discount factors by up to 6.6e-3 on the monthly grid, so the schedule is read from the prints and non-annual prints are excluded. - The SOFR publication calendar is the business-day calendar for all dates. Reason: SOFR is published for the days the Treasury repo market is open, as determined by SIFMA ([ARRC2021] p. 5), and QuantLib's `UnitedStates(SOFR)` implements that calendar. Alternative: the New York Fed's holiday list applied directly. Limitation: date-by-date equivalence of the QuantLib calendar with SIFMA's is not verified. - The intraday move of each tenor is its loading times one common path. ## Worked example and how it is checked As-of 2026-09-18, spot 2026-09-22. The original calculation used a checksum-checked raw snapshot and pinned versions. Historical stage receipts are supplied under `evidence/environment/`; the frozen outputs below are inputs to this documentation commission. The public bundle includes the calculation source for inspection, but omits the original raw archive and third-party PDFs. Therefore it supports documentation and numerical inspection, not a fully offline raw-data rerun. No new calculation is claimed. | Check (`out/verify.md`) | Result | |------|--------| | Repricing of every input | max abs error 3.9e-9 bp (A), 2.8e-10 bp (B): solver tolerance | | The paper's two example curves ([HW2006] pp. 102, 107) | B recovers every discrete forward and stays positive on the curve on which the paper reports that every cubic spline goes negative (minimum forward 0.87 %). The p. 102 table prints a 4Y yield of 4.40 %; its own capitalisation factor, 1.22140276 = e^0.2, and discrete forward, 4.33 %, both imply 5.00 %, and the check uses the consistent columns | | The paper's conditions (i)–(iv) on the curve of record | discrete forwards recovered to 6.3e-10; minimum forward 3.85 %; continuous at every node (6.1e-11); no interval breaks local monotonicity | | Round trip | a flat 4 % curve priced into deposit and swap quotes and bootstrapped back is recovered to 9.1e-11 bp (A) and 6.0e-12 bp (B) | | Method A, two implementations | QuantLib against the log-linear bootstrap in `hagan_west.py`: discount factors agree to 1.9e-13 on every business day to 30Y | | The pricer, two implementations | `hagan_west.ois_par_rate` against QuantLib's OIS helper: 2.1e-13 bp | | Method B through QuantLib | QuantLib's swap engine prices every input off method B's curve to 1.3e-11 bp | | Method B's interpolation, two implementations | against QuantLib's `ConvexMonotoneInterpolation(0, 1)` on the same nodes: 3.5e-13 bp. QuantLib's positivity option floors only the end forwards ([QuantLib-CM], `ConvexMonotoneImpl::update()`), not the interior collar of [HW2008] §6.1; the collar does not bind here | | Out-of-sample repricing | 983 par prints starting on or after spot, close-adjusted; the 460 the bootstrap never saw (other maturities, forward starts) have a median absolute residual of approximately 0.40 bp under B (corrected rounding in this public adaptation; the original historical report remains unchanged). The adjustment uses the intraday model of choice 2; on the 49 held-out prints from the last two traded hours at their raw rates, where the adjustment is at most 0.2 bp, the median residual is −0.47 bp (B) and −0.76 bp (A): both curves sit slightly above the late prints | ## Institutional facts this brief cannot supply Every one of these is a slot in the document, in the house marker syntax: inventory id and tier; data lineage to internal feeds; prior validation findings; approval conditions; usage inventory; monitoring plan and thresholds; roles; internal implementation and change record; the benchmark curve from the institution's market data and the tolerance for comparing with it; the production instrument set and its conventions; the quote snapshot time and stale-quote rules; the extrapolation policy beyond the last node; the risk system's bump methodology; the document's version history; the sign-off block. The repricing tolerance in the evidence policy is owned by the institution through its house layer.