Model Data Inventory — Shared Material (PC / RCA model + optimization tool)

Compiled: 2026-06-25 · Author of source material: Alon & Ishana (NZIPL modeling team) Purpose: inventory of two shared drops on OneDrive (000 Data Infrastructure/), what each contains, and — critically — whether the material is reproducible and whether EVs are covered.

1. CICE Updated/ — the model inputs & per-tech RCA analyses

Folder holding the Predicted-Competitiveness (PC) / RCA model material.

Inside ML_vars.zip

ML_vars/
├── Bentley_data_cleaning.R          # WDI download + cleaning (context variables)
├── wdi.xlsx · wdi_impute.csv        # World Bank Development Indicators (imputed) — model covariates
├── IRENA_INSPIRE_Patents_April_2024(INSPIRE_data).csv   # patent counts (feature)
├── Analysis/RCA <Tech> Analysis.ipynb   # 14 per-tech notebooks (the actual model runs)
└── RCA Construction/<Tech>/         # ⚠️ ALL EMPTY directory stubs (no files shipped)

Context / covariate data present (reproducibility inputs):

• wdi_impute.csv + Bentley_data_cleaning.R — the WDI macro covariates (GDP, population, FDI, electricity use, governance, trade openness, manufacturing share, …) and how they’re built.
• IRENA_INSPIRE_Patents…csv — patent feature.
• baci.zip — raw trade (the RCA base).

Per-tech model notebooks (Analysis/) — 14 technologies, each a fully-executed notebook (RandomForest classifier on an RCA>1 target + shap.TreeExplainer for feature importance): Batteries · Biofuel · Electrolyzer · Geothermal · Heatpump · Magnets · Nuclear · Solar · Transmission · Wind · EV · plus DRI, DAC, Mass Timber (not in CVCE).

Reproducibility status — ⚠️ PARTIAL

The notebooks read their main inputs from hardcoded local paths on the author’s machine that are not bundled, e.g. (from RCA EV Analysis.ipynb):

ev      = pd.read_csv("…/RCA Construction/EV/ev_trade.csv")        # ❌ missing (folder empty)
rca_hs17= pd.read_csv("…/RCA Construction/EV/hs17_rca_cyh.csv")    # ❌ missing
hs_master=pd.read_csv("…/concordance/HS_Descriptions.csv")        # ❌ no concordance/ in zip
wdi     = pd.read_csv("…/ML_vars/wdi_impute.csv")                  # ✅ present

So: present = model code (notebooks), WDI covariates, patents, raw BACI. Missing = the intermediate “RCA Construction” per-tech panels (ev_trade.csv, hs17_rca_cyh.csv, …) and the HS concordance file. Those RCA Construction/<Tech>/ folders ship as empty stubs. → You cannot push-button rerun the model as-is; the intermediate RCA panels between raw BACI and the RF must be regenerated or requested.

EVs specifically — answering “is it empty?”

• Not empty as code: there is a complete RCA EV Analysis.ipynb (22 cells, fully executed, dated 2026-05-15) — RandomForest + SHAP, identical pattern to the other techs.
• Empty as data: the RCA Construction/EV/ output folder is an empty directory in the zip. The notebook writes its results there on the author’s machine (out_dir = …/RCA Construction/EV/):
  • ev_rca_by_category_year.csv — RCA by category × year (→ atlas radar axis)
  • X_rank_save[['country_code','year','predicted_comp']] — PC scores by country-year (→ pc_scores)
  • df_2024_ev[['country_code','rank','rank_label','predicted_comp']] — 2024 ranks
  • a per-HS6 SHAP table (shap_df, shap_by_cat) (→ pc_features SHAP weights)

  None of these output CSVs are in the zip. They exist only inside the notebook’s run.

Implication for Task 3 (wire EVs into the atlas radar + PC scatter): the three CSVs the CVCE pipeline needs (PC scores, RCA-by-category, SHAP-by-HS6) are computed but not shipped. Cleanest path is to ask Ishana/Alon for the contents of RCA Construction/EV/ (3 small CSVs) — that is exactly the pc_scores.parquet / pc_rca.parquet / pc_features.csv material for EVs. Failing that, the values are partially recoverable by re-executing the EV notebook, but only after the missing ev_trade.csv / hs17_rca_cyh.csv / HS_Descriptions.csv inputs are obtained.

2. Alon's optimization tool.zip — CICE-V1 Interactive POC

A self-contained, email-able proof-of-concept dashboard: a country-competitiveness sensitivity tool. “If country X invested an extra ΔExport (as % of GDP) in technology T, how would its probability of being competitive — P(RCA>1) — move, and which HS categories drive it?”

web/
├── interactive_poc.html        # the dashboard (template + manifest + HS6 meta inlined)
├── interactive_template.html   # hand-written template (__INLINE_DATA__ substituted at build)
├── pipeline_flowchart.html + FLOWCHART_GUIDE.md   # methodology flowchart + its style guide
├── README.md                   # methodology + build instructions
└── data/
    ├── manifest.json           # {tech:{iso3:[years]}} availability + country names + ALPHAS grid
    ├── _hs_meta.js             # HS6 code → description cache
    └── <Tech>.json.gz          # one gzipped bundle per tech (country × year payloads)

Scope: 10 technologies (Batteries, Biofuel, Electrolyzer, Geothermal, Heatpump, Magnets, Nuclear, Solar, Transmission, Wind) × ~153 countries × 21 years (2003–2023). No EVs / DRI / DAC / Timber.

What the dashboard does (methodology):

• Left panel — P(RCA>1) vs. investment expressed as ΔExport (% of GDP, with a $ axis). The blue curve is the full log-linear-form (LLF) model prediction; drag the marker to read P, ΔP, ΔExport (%GDP), ΔExport ($). An orange dotted “validity range” marks the largest investment α at which the local linearization stays within 50% of the full prediction. A static label reports P_baseline, marginal dP/dExport ($), and the validity threshold.
• Right panel — RCA Impact (%) by HS2 category, recomputed live as you drag. Bar height = Σ_{i∈C} gᵢ² / Σ_all gⱼ² × 100 (bars sum to 100%); stacked segments are individual HS6 products by their L2² share. Categories: Chemicals, Industrial Materials, Electronics, Machinery, Metals, Textiles, Other.

Per-country-year payload (in each <Tech>.json.gz): pb (baseline P), gdp, exp (export level), val, and pf (the per-HS6 gradient / prediction vector) — the inputs that drive the draggable sensitivity curve.

Reproducibility: the bundles are build artifacts (gitignored upstream), regenerated by scripts/build_interactive_data.py --techs ALL --countries ALL --years ALL. That build script is not in this zip (only its outputs + the assembled HTML are). Runs over HTTP only (uses fetch); Plotly.js + pako load from CDN.

Bottom line