limitations.md

Model and implementation boundaries

• The original CLI/portfolio replay uses single-curve projection and discounting. The v2.1 extension adds OIS discounting and separately calibrated projection curves. The basis extension calibrates same-currency compounded overnight/overnight basis swaps, with no multi-currency support or XVA. Real Fed fitted Treasury curves provide a historical valuation proxy; the swap portfolios are hypothetical.
• Deterministic deposits, fixed/floating and compounded overnight basis swaps. No options, caps, floors, convexity adjustment, stochastic models or credit risk.
• The v2.2 SOFR APIs value future, seasoned, matured/unpaid and settled coupons with availability-aware realized fixings, accrued value, lookback, observation shift, lockout and spot/payment lags. Legacy vanilla pricing APIs still require future coupons. Neither API supports floating spread/gearing, amortization, ex-coupon treatment or principal exchanges. The basis API adds a simple spread on its base leg. A generic clearing ledger models acceptance/rejection, explicit payment settlement and termination closeout; a cash-collateral account models calls, transfers, interest and return. Neither connects to a CCP or implements novation, initial-margin calculation, settlement variation, noncash collateral, default management or external settlement reconciliation. Full trade amendment workflows remain unsupported. The v2.3 basis workflow generates final obligations from contractual coupons and checks complete normal-maturity settlement. SQLite journals add local transactional persistence/concurrent writers; source authenticity, distributed coordination and a network service remain outside scope.
• Gregorian dates restricted to 1900..2199; month-based forward schedules with short final stubs. No long/front stubs or IMM rolls. C++ calendars take explicit holidays; the SOFR pipeline exports a versioned QuantLib US Government Bond calendar with finite coverage. Operators must review calendar updates.
• The default bootstrap uses log-linear discount interpolation, with discontinuous instantaneous forwards. No extrapolation or smoothing. The separate curve ambiguity experiment supports quote intervals and explicit forward bounds on a complete cash-flow grid. It reports conditional attainable PV ranges, not probability intervals or market-calibrated uncertainty. Targets are unseasoned vanilla swaps; an optional projection curve is held fixed. Joint discount/projection ambiguity and seasoned SOFR/basis ambiguity are unsupported.
• Bootstrap solves within log-discount bounds [-50,50]. Pathological or inconsistent inputs may fail; failure is surfaced. No claim is made that arbitrary mixed quotes satisfy a complete arbitrage analysis.
• Quote risk uses full rebootstrap per scenario, shared across a portfolio. No adjoints, automatic differentiation or parallel execution.
• rates_cli handles one trade with configurable conventions; rates_batch supports portfolios using default conventions and strict unquoted CSV. Export files are written after computation; disk failure can leave partial files and is reported with nonzero exit status.
• Independent Python and QuantLib tests cover selected alternate day counts, holidays, business-day rules and irregular schedules, not every combination. Expanded SOFR checks cross five schedule families with lookback/shift/lockout, distinct payment calendars and lifecycle states; two-curve market-quote risk is checked by independent recalibration. This does not extend the supported schedule model to front/long stubs or establish exhaustive convention coverage.
• QuantLib rate helpers independently calibrate six compatible cases. Irregular forward-stub calibration contracts are excluded from helper comparisons where QuantLib's schedule conventions differ; irregular trade schedules are checked separately. No QuantLib performance comparison is claimed.
• Benchmarks describe one machine and workload. They are not service latency guarantees. Measured coverage identifies remaining unexecuted paths; it does not establish mathematical completeness.
• Historical replay supports synthetic snapshots and real fitted zero curves with an in-memory timestamp/offset index and immutable local files. The first snapshot defines the quote-label universe; there is no independent feed completeness manifest. File size and modification time detect ordinary mutation, not adversarial tampering.
• Replay values at the selected snapshot's valuation date. It does not roll curves to a later query date or supply missing past fixings. The example is intraday in the original example; trades with already-started coupons fail explicitly.
• Fed GSW historical data are fitted Treasury estimates from the downloaded, possibly revised vintage. Original publication times are unavailable. The observation-date labels support retrospective analysis, not point-in-time strategy backtesting. Source zero yields are not executable swap quotes.
• The SOFR data adapter retrieves public NY Fed observations over HTTPS with bounded retries, validates completeness and archives raw responses with hashes. Exact historical publication timestamps are unavailable: a downloaded fixing becomes usable only at its first local capture time. Changed values become new immutable local revisions. This cannot reconstruct an uncaptured historic information set. Missing required past fixings fail, including before their publication; they are never replaced with forecasts.
• OIS quote import checks source, units, timestamps, freshness, calendar and conventions, required tenors, bid/ask bounds and configured rate/spread limits. It does not supply a licensed tradable quote feed, source authentication, market-wide arbitrage certification, operational monitoring or a service SLA. Stored hashes detect corruption, not maliciously regenerated snapshots. The committed real-fixing demonstration uses synthetic OIS quotes.
• Optional firm-quote policy additionally checks declared execution size, expiry, venue and source/usage-rights references. These are caller-supplied assertions, not independently verified tradability or permission to redistribute data.
• The synthetic hedging study minimizes unweighted quote-DV01 norm with box limits and rounded notionals. It excludes funding, margin, basis and liquidity risks. Reducing the chosen norm does not guarantee lower loss under every scenario.

This is a reproducible quantitative engineering project, not a complete institutional derivatives platform or a claim of production readiness.

The separate ETF simulator is single-instrument long/cash, with fixed modeled spreads/fees and next-original-open fills. It has no liquidity, market-impact, corporate-action, settlement or interest model. Its first public TLT example covers an action-free interval and is explicitly retrospective. Availability checks protect the signal from future inputs under the supplied timestamps; they cannot certify the historical provenance of those timestamps.

Next improvements: front/long stubs, IBOR/arithmetic-average/cross-currency basis, spread/gearing support beyond the basis leg, provider/CCP/payment connectors, initial-margin models, validated event snapshots and distributed operations.