China-HQ green FDI readiness / Method appendix
Back to the story

EV capability measurement

Purpose. Documents how EV joined the scored readiness panel: the dedicated capability model that unblocked it, the shortened pre-FDI window and its consequences, the 101-code EV basket and its stage-type assignments, the third-market bloc definition, and the HQ-audit sensitivities specific to EV. EV cells score under the same floors and the same lagged-capability rule as solar and battery; every deviation from the solar/battery setup is listed here.

Why EV joins now

The v1 station scored only Solar and Battery. EV was excluded for one reason: the shared predicted-competitiveness model behind the capability axis covers ten green technologies and EV is not one of them (methodology decision 1, 2026-05-17). The 111 EV projects in the FDI panel were retained for ad hoc analysis but never scored.

A dedicated EV predicted-competitiveness model now exists. It was built on the same design as the shared model: a machine-learned score in [0, 1], fit on revealed-comparative-advantage patterns across the HS6 product space, representing the modeled probability that a country will be a competitive EV exporter given its current export portfolio. It is a modeled probability, not the truth. With that model in hand, EV enters the quadrant analysis under the identical classification rule: PC at or above 0.10 plus disclosed China-HQ capex at or above $100m separates the quadrants, with the fifth label for active cells below the capex floor.

The EV capability source

The EV model outputs are frozen copies in the station at data/raw/ev/ (see the README there for file-level provenance):

The top of the EV capability ranking on the pre-window mean reads as expected for the 2010s auto-export order: Belgium 0.97, Germany 0.94, Korea 0.91, Slovenia 0.90, Slovakia 0.88, the UK 0.86, the US 0.81, France 0.79. China itself scores low pre-window (0.01 to 0.05 in 2018 to 2019) and rises to 0.73 by 2023, consistent with how late China's EV export capability became visible in trade patterns.

Window deviation

The EV model series starts in 2018, so the EV pre-FDI capability window is the 2018 to 2019 mean, not the 2015 to 2019 mean used for solar and battery. The broad-RCA pre-window for EV is 2017 to 2019. Three consequences:

  1. The lagged-capability guardrail still holds. Both EV pre-windows end in 2019 and strictly precede the 2020 to 2024 FDI window, so same-period FDI cannot feed back into the capability score. The lag is shorter (one year of separation instead of five), which is a weaker break on reverse causality than the solar/battery setup. Read EV Prepared Magnet labels with that in mind.
  1. EV momentum is structurally unavailable. The station's momentum term is an OLS slope requiring at least three annual observations; the EV pre-window has two. Every EV cell carries a blank pc_momentum_pre, and the ranking composite reweights to PC level and broad RCA only. Quadrant labels are unaffected (they never used momentum).
  1. Machine-readable provenance. The capability table and the worktable carry per-row pc_window and rca_window columns (2015-2019 / 2018-2019 and 2015-2019 / 2017-2019 respectively), so no downstream consumer has to guess which lag a cell carries.

Floor calibration

The PC = 0.10 floor was chosen for solar and battery as a top-decile incumbent-capability threshold. It calibrates the same way for EV: 16 of 158 EV economies (10.1 percent) sit at or above 0.10 on the 2018 to 2019 mean, against 11 of 155 (7.1 percent) for each of solar and battery. Prepared Magnet keeps its intended meaning of confident incumbency in all three sectors.

The scored EV panel lands at: 6 Prepared Magnet (Thailand, Hungary, Slovakia, Korea, Spain, US), 12 FDI-First Experiment (led by Mexico $3.8b, Brazil $1.7b, and Indonesia $1.3b), 8 Untapped Candidate (Germany, Belgium, Slovenia, the UK, Austria, the Netherlands, Japan, Georgia), 11 Active below/unknown (including Serbia, Czechia, Oman, Qatar, and Ethiopia), 119 Bypassed. With 37 colored cells, EV is the most FDI-active panel of the three technologies.

The basket and its stage types

The source basket lists 101 HS6 codes. Two curation layers sit between that list and the pipeline, both recorded in data/manual/ev_stage_types.csv (one row per resolved HS17 code, with the source code and a note for every fix):

HS-vintage repair. Nine source codes are not valid HS2017 and one is a typo; unrepaired they would silently vanish from every BACI scan. Splits: 845610 into 845611 + 845612; 722610 into 722611 + 722619; 854280 into 854231/32/33/39. Merges: 870831 + 870839 into 870830; 870860 into 870850. Recodes: 854800 (an HS22 code) to HS17 854890. Typo: 61610 to 261610 (silver ores; the source file's own description confirms the lost leading digit). Dropped: 741410 (heading deleted from the HS, no successor carries the trade). The result is 103 valid HS17 codes. scripts/build_ev_basket.py validates every code against the BACI HS17 product list and fails hard on any miss, so vintage rot cannot recur silently.

Stage-type assignment. The station's supply-chain reads (manufacturing and factory-stage cuts, scatter dot colors) need each code typed as Final Product, Product Component, Process Equipment, Processed Material, or Raw Material. Two tiers:

  1. Inherit. 43 of the resolved codes already appear in the green dictionary under other technologies with a stage type; EV inherits it (for example 848620 Process Equipment, 281820 Processed Material). One code needed adjudication: 850511 permanent magnets is Final Product for the Magnets tech but a motor input for EV, typed Product Component here.
  1. HS4-family rule. The remaining codes are typed by heading family: 870380 is the single Final Product; 8708/8501/8504/8511/8512 vehicle and drivetrain parts are Product Component; the 845x/846x machine-tool block plus welding, induction, coil-winding, and instrument codes are Process Equipment; chapter-28 chemicals, electrical steel, nickel forms, and films are Processed Material; ores and scrap are Raw Material.

Final counts: 1 Final Product, 39 Product Component, 22 Process Equipment, 24 Processed Material, 17 Raw Material. The four cuts follow the solar/battery definitions exactly (manufacturing excludes Raw Material; factory-stage keeps Final + Component + Equipment).

Multi-use caveat. Eight codes are flagged multi-use in the table: the four integrated-circuit codes (854231/32/33/39), the two semiconductor-equipment codes (848610/848620), generic measuring instruments (903089), and generic iron castings (732599). These carry trade far beyond EVs. They stay in the basket because the model was trained on it, but per-code reads on them inherit the station's existing multi-use-commodity caveat, and the third-market scatter tooltips label them.

Third-market competition: bloc and concordance

The EV third-market scatter uses the incumbent bloc Germany + Japan + Korea + US, the 2010s auto-export leaders, defined tech-specifically in the same spirit as the battery bloc. Third markets exclude China plus the bloc as destinations.

The pre-window (2012 to 2014) is read from HS2012 data, which requires antecedents for three codes: 870380 did not exist in HS2012; its antecedent is 870390 ("other vehicles, including electric"), the standard pre-2017 EV proxy and a broad parent, so the pre-window baseline mixes in non-BEV trade and the cell is flagged broad_parent in the figure. 845611 and 845612 both carry the broad parent 845610. All three flags surface in the scatter tooltips and the QA note.

Headline read: 69 of 86 plotted EV products sit in the China-gains, incumbents-lose quadrant, the broadest shift of the three technologies. In battery-electric vehicles China rose from about 1 percent (on the broad-parent baseline) to 26 percent of third-market exports between 2012-2014 and 2022-2024, while the four-country bloc slipped from 53 to 45 percent. Generic vehicle parts (870899) moved from 4 to 10 percent China share with the bloc down 7 points.

HQ-audit interaction

The HQ audit's control-linked and ambiguous records are concentrated in EV, so the sensitivity that was previously outside scope now binds. Within the 2020 to 2024 window, the panel-rule versus documented-exclusions deltas that matter:

Under the strict screen the EV Prepared Magnet set is therefore Thailand, Hungary, and Spain. The Windrose (Belgium, France) and Leapmotor (Spain, Malaysia) ambiguous records are all announced 2025 or later, outside the scoring window, and affect no scored cell. As elsewhere in the station, panel-as-ingested remains the reproducible data layer and documented-exclusions is the preferred screen for narrative claims; both magnitudes stay side by side in the audit table.

Scope limits

Markdown source on GitHub