Modernise OG-UK: new PolicyEngine API, 8-sector calibration, OBR-grounded TPI charts

[nwoodruff-co]

Summary

This PR modernises OG-UK end-to-end: new PolicyEngine API integration, UK macro calibration, 8-sector industry model, and OBR-grounded transition path charts.

Core modernisation

• New PolicyEngine API: Clean functional interface (solve_steady_state, run_transition_path, map_to_real_world) replacing the old imperative runner — see oguk/api.py
• UK macro calibration: Fiscal params calibrated to OBR Nov 2025 EFO; real-world mapping anchored to ONS GDP/expenditure data — see oguk/macro_params.py
• Tax function estimation: Fixed two critical bugs; switched from DEP to GS functional form; added age-specific estimation option
• Demographics: Use ogcore.demographics for UK population data, drop legacy demographics.py and setup.py
• CI/docs fixes: Pin jupyter-book<2, fix lint, stub UN API token, modernise pyproject.toml

8-sector industry calibration (oguk/industry_params.py)

• M=8 production industries mapped to SIC 2007 sections (Manufacturing last — capital good producer per OG-Core convention)
• Capital shares (gamma): ONS GDP income approach, shrunk 40% toward aggregate mean (0.35) for solver stability
• IO matrix: ONS Input-Output Supply and Use Tables 2022 (Table 2), aggregated from ~100 product groups into 8 sectors, row-normalised
• TFP (Z): Solow residuals from ONS capital stocks (CAPSTK) and workforce jobs (JOBS02), GVA-weighted normalisation
• CES elasticities (epsilon): Calibrated from Chirinko (2008) and Knoblach et al. (2020), but currently set to Cobb-Douglas (1.0) because OG-Core TPI diverges with heterogeneous epsilon. Raw values preserved in _EPSILON for when this is fixed upstream
• c_min: Non-zero energy floor for subsistence consumption
• Energy cost shares: For future energy price shock modelling

OBR-grounded TPI charts (scripts/tpi_charts_plotly.py)

• Macro charts: 6-panel (GDP, Consumption, Investment, Government, Tax Revenue, Debt) from 2000-2030. Baseline = ONS outturn + OBR forecast projected forward. Reform = baseline x (1 + model % change). Single continuous baseline line, dashed reform from 2027
• Sector charts: Per-industry % change summary (Output, Capital, Labour) + 8x3 panel with historical ONS outturn (GVA, capital stocks, workforce jobs) from 2000, projected forward using OBR GDP growth
• No stitching artefacts: Both macro and sector charts use the same approach — model % changes applied to OBR-consistent baselines

TPI export (scripts/run_tpi_and_export.py)

• Runs baseline + reform SS+TPI for 1pp basic rate increase from 2027
• Exports to tpi_results.xlsx: 3 macro sheets + 12 sector sheets (baseline levels, reform levels, % changes for Y_m, K_m, L_m, p_m)

Key model results (1pp basic rate increase)

• GDP impact: -0.3% by 2029 (consistent with OBR/HMRC ready reckoners)
• Labour impact: ~60-80k FTE reduction (consistent with OBR NICs/threshold estimates)
• Energy sector least affected (most capital-intensive); Public and Other goes positive (more tax revenue -> government spending)

Test plan

• run_tpi_and_export.py produces tpi_results.xlsx with macro + sector sheets
• tpi_charts_plotly.py produces tpi_charts.html with smooth baseline/reform lines
• CI passes (lint, format, tests, docs build)
• Manufacturing at index 7 (last sector) per OG-Core convention
• All IO matrix rows sum to 1.0
• Sector TFP normalised so GVA-weighted mean = 1.0

[jdebacker]

One thing I'd like to see changed in this PR before merge: make manufacturing the 8th industry in the multi-sector calibration.

[vahid-ahmadi]

Just wanted to flag these points on the multi-sector calibration:

1. With io_matrix = np.eye(8), there are no inter-industry flows. An energy price shock only hits households directly; manufacturing doesn't get more expensive because it uses energy as input. The supply-chain channel is absent. ONS publishes Supply and Use Tables (1997–2023) with the inter-industry flows needed for a real 8x8 coefficients matrix.

2. Cobb-Douglas everywhere. OG-Core already supports sector-specific CES (p.epsilon[m]), but the PR sets epsilon=1.0 for all 8 sectors. This means capital-labor substitution is identical across energy and business services, which isn't realistic. Rough literature values: Energy ~0.5, Manufacturing ~0.8, Construction ~0.7, Services ~1.1–1.3, Real Estate ~0.4 (Chirinko 2008, Knoblach et al. 2020).

3. Uniform TFP. With Z=1 everywhere, all cross-sector output differences come from capital shares and factor quantities alone. A sector with high productivity (finance) looks the same as one with low productivity (agriculture). Factor prices don't reflect true sectoral productivity differences. Sector-level Z can be backed out as a Solow residual from ONS data: Z_m = GVA_m / (K_m^γ * L_m^(1-γ)) using ONS capital stocks by industry and workforce jobs by SIC.

[jdebacker]

@vahid-ahmadi Thanks for flagging -- very good points to keep in mind.

On your first point, about supply chains, that's right - other than the capital good, no outputs are used by inputs in other industries. This is a limitation from OG-Core. I think it should be updated when we build out trade in goods other than the capital good.

[vahid-ahmadi]

Addressing sector heterogeneity concerns

1. IO matrix — Fully addressed ✅

Replaced np.eye(8) with a full 8×8 IO matrix sourced from ONS Input-Output Supply and Use Tables 2022 (Table 2: "Use of products by industry at basic prices"). Raw ~100 product-by-industry flows were aggregated into the 8-sector classification and row-normalised. The matrix is diagonal-dominant (each good primarily from its own industry) with off-diagonal entries capturing cross-industry supply chains (e.g. energy inputs to manufacturing, business services to finance).

2. CES elasticities (epsilon) — Partially addressed ⚠️

Raw literature values from Chirinko (2008) and Knoblach et al. (2020) are calibrated and preserved in _EPSILON:

• Energy: 0.50, Manufacturing: 0.80, Construction: 0.70, Trade & Transport: 1.00, Info & Finance: 1.20, Real Estate: 0.40, Business Services: 1.30, Public & Other: 0.90

However, epsilon is currently set to 1.0 (Cobb-Douglas) for all sectors in get_industry_params(). The reason: OG-Core's TPI solver produces NaN at iteration 1 with any heterogeneous epsilon — even when shrunk 80% toward 1.0 (range 0.88–1.06). The SS solver handles heterogeneous epsilon fine; only TPI breaks.

Next step: Investigate the root cause inside OG-Core's TPI.run_TPI() — specifically trace where NaN first appears in the inner loop when p.epsilon varies across sectors. This is likely a numerical issue in how TPI computes CES production function derivatives or initial guesses along the transition path. Once identified and fixed upstream, we can enable the calibrated epsilon values by changing one line in get_industry_params().

3. TFP (Z) — Fully addressed ✅

Sector-level TFP is computed as Solow residuals exactly as suggested:

Z_m = GVA_m / (K_m^γ_m × L_m^(1 − γ_m))

Using ONS capital stocks by industry (CAPSTK, 2022 net capital stock) and workforce jobs by SIC (JOBS02, 2022 Q4), normalised so the GVA-weighted mean equals 1.0. Raw (unshrunk) gamma values are used for the Solow residual calculation to ensure data consistency.

[vahid-ahmadi]

Update: Per-Industry TPI Impacts + Historical Outturn Charts

This latest push adds sector-level transition path results and real-world anchored charts for the 8-sector model.

What's new

Per-industry TPI export (run_tpi_and_export.py):

• Extracts Y_m, K_m, L_m, p_m (output, capital, labour, prices by sector) from OG-Core's TPI solve
• Exports 12 new sheets in tpi_results.xlsx: baseline levels, reform levels, and % changes for each variable across all 8 sectors

Sector charts with historical outturn (tpi_charts_plotly.py):

• ONS outturn (GVA by SIC section, capital stocks, workforce jobs) from 2000 onwards
• Projected forward using OBR nominal GDP growth — same approach as the macro charts (no stitching artefacts)
• Reform line = baseline × (1 + model % change from TPI)
• Sector % change summary charts (Output, Capital, Labour side by side)

Key findings from the 1pp basic rate increase

• Energy least affected (most capital-intensive, less sensitive to labour income tax)
• Public & Other goes positive: more tax revenue → more government spending
• Private sectors cluster together because epsilon is Cobb-Douglas (1.0) for all — heterogeneous CES would differentiate them more but breaks OG-Core TPI
• Total labour impact ~60-80k FTE, consistent with OBR's own estimates for similar-magnitude tax changes (e.g. NICs cuts ~50k FTE per pp, threshold freeze ~130k FTE). Most is intensive margin (hours) not job losses.

Limitations

• Epsilon homogeneity: OG-Core TPI cannot converge with heterogeneous CES elasticities. This limits cross-sector differentiation in reform impacts.
• Labour supply magnitude: The Frisch elasticity in OG-Core may be calibrated higher than UK-specific estimates. Direction is correct; magnitudes need validation against UK labour supply literature.

[rickecon]

@nikhilwoodruff and @vahid-ahmadi (cc @jdebacker). This PR looks great. I just reviewed and approved it. The only recommendation I have is that we update the version in pyproject.toml and /oguk/__init__.py to 0.3.1, with a corresponding update to the CHANGELOG.md. This is a significant calibration update with some updated Python scripts.

[nikhilwoodruff]

Thanks- merging!