UK Gilt / SONIA OIS Yield-Curve Construction, Forward Extraction & Term-Premium Decomposition

A research-grade fixed-income project that reads what the market prices about future monetary policy, inflation and risk premia out of UK rates. The spine of the project is one decomposition:

nominal yield = expected average short rate + term premium

and its inflation analogue (breakeven = expected inflation + inflation risk premium − liquidity premium). Everything here — curve construction, bootstrapping, forwards, the policy path, and the planned term-premium model — exists to populate that decomposition and then ask where the market's implied story disagrees with macro fundamentals.

This is SONIA-native (post-LIBOR): the GBP risk-free discount curve is the SONIA OIS curve, and Bank Rate expectations are read from meeting-dated OIS.

---

What is implemented (this slice)

A complete, tested, real-data pipeline for Objectives 1–4:

1. Curve construction — ingest the Bank of England's published SONIA OIS curve.
2. Bootstrapping — strip discount factors from par OIS quotes, from scratch.
3. Forward extraction — zero rates, instantaneous and period forwards.
4. Implied Bank Rate path — forwards sliced at MPC meeting dates (≈ WIRP).

The numerical core is built from scratch in numpy/scipy — no black-box pricing library on the critical path — so the mathematics is fully inspectable. rateslib / QuantLib are wired in only as optional cross-validation (pip install -e ".[validate]").

Headline result (BoE curve, as of 19 Jun 2026)

The front end prices a gradual rise in forward SONIA of ~+45bp to end-2027. Note this is the raw forward path, which conflates expected policy with term premium — disentangling the two is the next phase (ACM), and is exactly why the term-premium decomposition matters.

Validation (on real BoE data, reported by the runner)

Check	Result	What it proves
Bootstrap round-trip (par → DF → bootstrap)	**max	ΔDF
Our forwards vs BoE's published instantaneous forwards	RMSE = 0.16 bp	the compounding convention matches the BoE's own

The continuous-compounding convention (D(t)=exp(−R(t)·t)) was pinned empirically, not assumed: it reproduces BoE's published forward curve to 0.3bp vs 8.5bp for annual compounding.

---

Architecture

curves/
├── data/
│   ├── raw/            # BoE workbooks (re-fetchable; git-ignored)
│   └── processed/      # implied_policy_path.csv
├── src/giltcurve/
│   ├── conventions.py          # day-count, spot<->DF (continuous)
│   ├── ingest/boe.py           # download + parse BoE SONIA OIS curve
│   ├── curves/
│   │   ├── discount.py         # DiscountCurve (log-linear in log DF)
│   │   ├── ois_bootstrap.py    # par OIS -> discount factors (Step 2)
│   │   └── forwards.py         # zero/forward reporting tables (Step 3)
│   ├── policy/
│   │   ├── mpc.py              # MPC meeting-date calendar
│   │   └── meeting_dated.py    # implied Bank Rate path (Step 4)
│   └── viz/plots.py            # desk-style charts
├── scripts/run_policy_path.py  # end-to-end pipeline + validations
├── tests/                      # 32 tests (TDD); 1 real-data integration check
├── reports/figures/            # generated charts
└── notebooks/                  # (narrative analysis — roadmap)

Quickstart

python3 -m pip install -e ".[dev]"   # or: pip install numpy scipy pandas matplotlib openpyxl pytest
python3 -m pytest                    # 32 passing (real-data check auto-skips if offline)
python3 scripts/run_policy_path.py   # downloads BoE curve, prints path + validations, saves charts

Data

• Source: Bank of England published UK yield curves — the same curve the MPC and UK desks watch, with published methodology (Anderson–Sleath VRP spline). Free and fully reproducible; no terminal required. https://www.bankofengland.co.uk/statistics/yield-curves
• Desk path: a bloomberg.py adaptor returning the same (asof, maturities, rates) tuple — or raw par-swap quotes straight into bootstrap_ois — drops in unchanged.

Roadmap (Objectives 5–8)

Phase	Module	Note
Yield-curve PCA	premium/pca.py	level/slope/curvature; foundation for ACM
Term premium (ACM)	premium/acm.py	the centerpiece: split yields into expected-rate + term premium
Breakeven decomposition	inflation/breakevens.py	nominal vs index-linked gilts; RPI wedge + 2030 reform
Cross-checks & divergence	notebooks/	implied path vs Consensus; flag pricing inconsistent with fundamentals

Caveats

• MPC dates for 2026–2027 in policy/mpc.py are projected from the BoE's ~6-weekly cadence and should be reconciled against the official calendar.
• SONIA → Bank Rate basis (bank_rate_minus_sonia_bp) is a small, explicit parameter; it cancels from changes in the path, so the cumulative-bp figure is the robust headline.