Skip to content

PROGRESS.md

Latest commit

 

History

History
67 lines (59 loc) · 3.93 KB

PROGRESS.md

File metadata and controls

67 lines (59 loc) · 3.93 KB
Raw

PROGRESS

Project 1 — FX implied vol surface engine. All milestones complete.

Milestone Status Evidence
M1 pricing core done fxvol/pricing.py; parity/known-value/Greek-FD tests (tests/test_pricing.py)
M2 delta machinery done fxvol/delta.py; 4-convention round trips, PA branch test, DNS tests (tests/test_delta.py)
M3 smile calibration done fxvol/calibration.py; market-strangle reprice + RR exactness (tests/test_calibration.py)
M4 smile + surface done fxvol/smile/*, fxvol/surface.py; headline round-trip at every pillar (tests/test_surface.py)
M5 arbitrage done fxvol/arbitrage.py; clean set passes, frown/calendar violations detected (tests/test_arbitrage.py)
M6 viz + report + demo done fxvol/viz.py, fxvol/io_report.py, examples/run_demo.py, examples/demo.ipynb

Test suite: 209 passed, 2 skipped (PA-call strikes below the turning point — not on the market branch by construction). Coverage ≈ 94% lines (python -m pytest --cov=fxvol). ruff check and mypy fxvol clean.

Round-trip acceptance on sample EURUSD data: worst error 1.2e-11 vol points across ATM/RR25/BF25/RR10/BF10 at all five pillars (target < 0.01). Same test passes for a synthetic premium-adjusted USDJPY market including a 2Y forward-delta pillar.

Convention decisions (with citations)

  1. Premium adjustment is a per-pair registry flag, set by the premium currency being the base ccy (Reiswich–Wystup 2010 §2.2; Clark 2011 §3.3). EURUSD/GBPUSD/AUDUSD/NZDUSD unadjusted; USDJPY/USDCHF/USDCAD/USDBRL/ EURJPY/EURGBP adjusted. Unknown pairs raise — no silent guessing.
  2. Spot vs forward delta cutoff: spot delta for T ≤ 1Y inclusive, forward beyond (Clark 2011 §3.3); configurable per pair (spot_delta_cutoff_years).
  3. ATM = delta-neutral straddle by default (RW 2010 §3.2): K = F·e^{+σ²T/2} unadjusted, K = F·e^{−σ²T/2} premium-adjusted (sign flip tested). ATMF/ATM-spot selectable.
  4. PA call delta inversion picks the higher-strike branch after locating the turning point σ√T·N(d2) = n(d2) (RW 2010 §2.2.2). Unattainable deltas raise ValueError.
  5. Butterfly = market strangle (RW 2012 §3): calibration solves the 2-constraint system (RR difference + strangle-price reprice) with Brent, seeded by the naive vols atm + bf ± rr/2. The smile-strangle BF diagnostic (WingCalibration.smile_bf) shows the wedge vs the quoted BF (e.g. 1Y EURUSD sample: 1.015% smile vs 1.050% quoted).
  6. Smile interpolation: PCHIP in (log-moneyness, total variance). Wing extrapolation: asymptotically flat-in-vol with a C1 exponential slope decay (default extrap_decay = half node span). Rationale: hard flat wings are C0 and their kink produces a genuine negative-density spike at the 10Δ strikes — discovered by our own arbitrage checker during development and kept as a selectable "flat" mode for demonstration.
  7. Term structure: linear in total variance at constant forward log-moneyness (Gatheral 2006); flat-vol extrapolation outside pillars. Zero curves linear in T, forwards from covered interest parity.
  8. Day count act/365F on calendar-day tenor rolls (EOM-clamped month adds); no holiday calendar yet — extension point, documented in market_data.add_tenor.
  9. File units: percent in CSV/JSON, decimals in code; conversion happens only at the IO boundary.

Known limitations / next steps

  • No business-day calendar / cut times (ON pillar = 1 calendar day).
  • mypy fxvol and ruff check both pass clean (config in pyproject.toml). Python here is 3.10, so 3.10-compatible syntax is used (spec asked 3.11+; nothing 3.11-specific is required).
  • Arbitrage repair mode intentionally not implemented (diagnostics only, by design).
  • Project 2 attach point: TenorCalibration nodes + pricing.greeks vega/vanna/volga. Project 3 attach point: smile/sabr.py stub.